Characterizing gradients of functional connectome underpinning rumination and their alteration in depression

Abstract

As a prominent psychopathological process of major depressive disorder (MDD), rumination’s brain underpinnings remain unclear. Emerging studies have highlighted that the brain is organized along several macroscale gradients, which could serve as a powerful framework for better understanding how the functional connectome underlies rumination. In this study, we leveraged two datasets (Rum-Beijing and Rum-MDD) to characterize the gradient structure during an active ruminative state. Rum-Beijing consisted of 40 healthy controls (HC) who underwent 3 repetitive scans, while Rum-MDD consisted of 45 patients with major depressive disorder (MDD) and 46 HCs. We used a modified rumination state task (RST) to induce participants into a continuous, active rumination state and investigated the gradient profiles. Several global features of the gradient (range, explanation ratio, variance) and the regional differences of each gradient were also compared. In the Rum-MDD dataset, we further examined the interaction effect between the group (MDD vs. HC) and condition (rumination vs. distraction) regarding the gradient’s global and local metrics. Two gradients were identified: the primary-transmodal gradient and the visual-sensorimotor gradient. We found that the rumination state exhibited reduced gradient values in the default mode network (DMN) as compared to the distraction state. We identified replicable altered gradient values in the left dorsolateral prefrontal cortex (DLPFC), superior frontal gyrus (SFG), and posterior cingulate cortex (PCC). Relative to the distraction state, individuals’ gradient range exhibited a replicable and significant reduction along the primary–transmodal gradient, accompanied by a significant increase along the visual–sensorimotor gradient during rumination. Moreover, the MDD group showed a significantly higher primary–transmodal gradient range during the rumination state than HCs. In conclusion, the present study showed that rumination may correspond to a specific underlying functional gradient profile, which was altered in patients with MDD. These results shed new light on the neural mechanisms underlying rumination, highlighting a global functional coupling characteristic across the whole brain during active rumination.

Introduction

Rumination is repetitive thinking about the meanings, possible causes, and consequences of distress [1]. Rumination is associated with higher susceptibility to depressive episodes and predicts the onset of depression [2, 3]. It also impairs problem-solving behavior and predicts delayed treatment response and higher relapse rates in major depressive disorder (MDD) [3, 4]. Existing evidence has shown that rumination is a treatable psychological process and could be alleviated by several interventions, such as mindfulness meditation, rumination-focused cognitive behavioral therapy, repetitive transcranial magnetic stimulation (rTMS), real-time neurofeedback, and transcranial direct current stimulation (tDCS) [5–9]. Recent studies have highlighted an association between the clinical efficacy of neuromodulation techniques and the extent to which they target the underlying functional network mechanisms of a given symptom [10, 11]. Accordingly, a better understanding of the rumination’s underlying functional network mechanism may lead to the development of next-generation rumination-targeted treatment for MDD [12–14].

Existing research has highlighted the pivotal roles of activities and interactions among brain regions across several large-scale brain networks, especially the default mode network (DMN) and frontoparietal network (FPN), in the neural underpinnings of rumination [15–19]. Task-based functional magnetic resonance imaging (fMRI) studies have found increased activity in the typical DMN regions, including the posterior cingulate cortex (PCC), medial temporal regions, and medial prefrontal cortex (MPFC), as well as in the FPN regions, including the inferior parietal lobe [20–24]. Furthermore, functional coupling among nodes of these two networks, typically characterized using resting-state fMRI (R-fMRI), was also associated with participants’ self-reported rumination tendency [25–27].

Functional connectivity (FC) may be insufficient to fully characterize the brain’s dynamic changes and complex functional organization. The cortical gradient is an analytical framework that captures the continuous, spatially organized pattern of variation in a specific characteristic across the cerebral cortex [28–30]. Since rumination is a recurrent and automatic form of emotional regulation and a particularly intense and concentrated form of inner speech [31], it may bear specific alterations regarding the global properties of the functional brain network. The cortical gradients place discrete large-scale networks and associated regional units on a continuous spectrum, extending from unimodal systems that support perception and action to association areas involved in more abstract cognitive functions. Such an approach reveals that the organization of the brain is not merely a collection of isolated regions but rather a dynamic and continuous system. This holistic and continuous perspective provides a more comprehensive understanding of the brain’s structure and function and can effectively capture reliable features of the interpretable functional biology of the human brain [32–34]. Using connectivity data from both humans and macaques, Margulies and colleagues [28] identified a primary-transmodal gradient (Gradient 1) that ranges from a major sensorimotor functional area on one end to a cross-modal area (i.e., DMN) on the other end, and a visual-sensorimotor gradient (Gradient 2) differentiates from the sensorimotor and auditory cortex to the visual cortex.

It is worth noting that some studies have explored gradient abnormality in MDD. Patients with MDD showed a compressed primary-transmodal gradient. The MDD group showed lower gradient scores, mostly in the DMN, than healthy controls. The regional changes also include parallel alterations in subcortical regions such as the caudate, amygdala, and thalamus. In addition, gene transcription profiles accounted for 53.9% of the gradient pattern changes. The patients’ baseline gradient maps significantly predicted symptom improvement after treatment [35, 36]. Similarly, existing research found reduced cortical gradient dispersion in the primary intrinsic brain networks of patients with treatment-resistant depression. Decreased dispersion within nodes of the default mode, control, and salience networks correlates with baseline levels of trait anxiety, depression, and mindfulness. Meanwhile, baseline dispersion in the salience network predicts trait anxiety scores at 24 weeks post-intervention [37].

These insightful studies have shed light on rumination’s neural underpinnings, albeit caveats remain. The task-related alteration in the brain’s macroscale functional organization during an active rumination state remains unknown and may have critical implications for its neural underpinnings. R-fMRI is suitable for investigating the functional network features and cortical gradients, but its association with self-reported rumination traits only provides indirect evidence of rumination’s network mechanism. To address these issues, we developed a rumination state task that induced participants into an active, continuous rumination state and investigated its network underpinnings. We have previously characterized reproducible FC and dynamic stability mechanisms underlying rumination [38, 39], as well as abnormalities in degree centrality in patients with MDD [40]. The induced, continuous rumination state provides an appropriate condition for characterizing cortical gradients, as this metric is derived from a functional connectivity (FC) matrix that quantifies the synchronization of activity among spatially distributed brain regions over an extended time window (approximately 8 min). In addition, the recent “replication crisis” across various scientific disciplines has sparked widespread concerns about the reliability of published findings, especially those related to neuroimaging results [41, 42]. Therefore, it is critical to establish the reliability of novel findings by assessing their stability across repeated measurements and across different datasets [43].

Here, we first aimed to examine the gradient profile of the functional connectome during an active rumination state in a group of healthy adults. Of note, we ensured the replicability of our findings by implementing a repeated-measured design using different scanners at different sites. We further explored the abnormality in this rumination gradient profile in MDD patients using an independent dataset. We hypothesized that both healthy and MDD individuals would show a broadly similar yet subtle variate macroscaled gradient structure during rumination compared to the typical resting state. A narrower primary-to-transmodal cortical gradient, as indicated by global gradient metrics and lower gradient scores in the DMN, would be observed during rumination. We further hypothesized that, compared with healthy individuals, the MDD group would exhibit abnormal macroscale organization across both gradients.

Methods

Participants

This study used two independent datasets [39, 40]. The first dataset (Rum-Beijing) consists of 41 healthy adults recruited from the local community around the Institute of Psychology, Chinese Academy of Sciences, Beijing, China. This dataset has been openly shared through the R-fMRI Maps project (http://rfmri.org/RuminationfMRIData). For more details about this dataset, please refer to our previous study [39]. We excluded one left-handed participant, resulting in a final analytical sample of 40 participants. The second dataset (Rum-MDD) included patients with MDD and healthy controls (HC) from Guangji Hospital in Suzhou, Jiangsu, China. HCs were recruited from the local community through advertisements. All individuals diagnosed with MDD underwent a 17-item assessment using the Hamilton Depression Rating Scale (HAMD). Patients with MDD were required to meet DSM-5 criteria as determined by a structured clinical interview conducted by a trained evaluator and to have a HAMD score of ≥17. Most MDD patients were receiving antidepressant treatment, including selective serotonin reuptake inhibitors and serotonin–norepinephrine reuptake inhibitors. Current medications mainly consisted of escitalopram oxalate, venlafaxine, fluvoxamine, duloxetine, and sertraline, and a subset of patients had previously been treated with paroxetine, sertraline, venlafaxine, fluvoxamine, or escitalopram oxalate (see Table S1 for details). For more detailed exclusion criteria for both MDD and HC, please refer to Jia et al. [36]. The analytical sample of the Rum-MDD dataset comprised 45 MDD patients (33 females, mean age = 26.09) and 46 healthy controls (35 females, mean age = 29.22, Table 1). The local Institutional Review Board approved the study protocol for both datasets, and all participants signed informed consent forms. This study represents a reanalysis of existing datasets, for which our previous analyses have successfully detected robust and statistically significant effects. Therefore, no a priori sample size calculation was performed.

Table 1.

Demographic information of participants.

	N	Sex (female/male)	Age	RRS	RRS-B	RRS-R	HAMD
Rum-Beijing
HCs	40	22/18	22.71 (4.01)	46 (11.32)	11.24 (3.28)	10.85 (2.33)	NA
Rum-MDD
MDD patients	45	33/12	26.09 (8.32)	61.08 (9.80)	14.58 (2.95)	12.16 (2.86)	23.29 (5.41)
HCs	46	35/11	29.22 (10.25)	36.10 (8.54)	9.10 (2.50)	8.35 (2.63)	NA

RRS ruminative response scale, RRS-R the reflection subscale of RRS, RRS-B the brooding subscale of RRS, HAMD Hamilton depression rating scale, HC healthy control, MDD major depressive disorder.

Experimental design

All participants in these two datasets completed RST, which has been previously validated in both healthy individuals [38, 39] and patients with MDD [40]. Before the MRI scan, the researcher interviewed the participants to brief the purpose of the study and the definition of the psychological state during the scan. The rumination state was defined as “passive and repetitive thinking about negative events and their possible consequences,” and the distraction state was defined as “imagining images unrelated to oneself.” The distraction state was intended to prevent the participant from spontaneously immersing in the rumination state and to serve as the control condition [39].

The fMRI scan consisted of four states. A resting state came first, during which the participants were asked to look at the fixation on the screen and to remain alert. Then, during the sad memory state, the participants were prompted to recall negative autobiographical events triggered by individualized keywords generated by the participants themselves before the scanning. During the rumination state, participants were asked to engage in repetitive, self-focused negative thinking in response to the prompts, such as, “Think: Why can’t I handle things better in the events I just remembered?” or “Think: Why do I experience these events that others do not?” In contrast, during the distraction condition, they were instructed to imagine and focus on concrete external scenarios, such as “Think about: A train stopped at a station” or “Think about: The layout of a typical classroom.” The prompts used in the rumination and distraction conditions were adapted from the validated rumination induction task (RIT) [44, 45] and the Ruminative Response Scale (RRS) [46], ensuring that the rumination condition specifically elicited negative, self-referential, and repetitive thinking patterns. Finally, during the distraction state, participants were prompted to imagine some objective scenarios. Except for the resting state, all mental states (sad memory, rumination, and distraction) included four sequentially displayed stimuli (keywords or prompts). An 8-minute continuous mental state was induced by transitioning from one stimulus to the next every 2 min, with no inter-stimulus intervals (ISIs). While the order of rumination and distraction states was balanced across subjects, the resting state and recall of negative autobiographical events consistently came first and second, respectively.

Participants also reported their momentary emotional states at the end of each session. Detailed findings have been described elsewhere; therefore, only a concise overview is presented here. For full results, see our previous reports [38, 40]. Briefly, these data demonstrated significantly increased negative affect following the rumination condition, whereas emotional states returned to baseline after the distraction condition, supporting the effectiveness of the state induction.

MRI data acquisition and preprocessing

All participants from the Rum-Beijing dataset repetitively underwent MRI scans on 3 different scanners at two sites: two 3 Tesla GE MR750 scanners at the Institute of Psychology, Chinese Academy of Sciences (IPCAS), and Peking University Magnetic Resonance Imaging Research Center (PKUGE), as well as a 3 Tesla SIEMENS PRISMA scanner (PKUSIEMENS) at Peking University. MRI images from the Rum-MDD dataset were acquired in a 3 Tesla SIEMENS SKYRA scanner at Guangji Hospital. Before functional image acquisition, all participants underwent a 3D T1 scan, and functional images for both resting state and mental states (sad memory, rumination, and distraction) were obtained. Data preprocessing was implemented with the toolbox for Data Processing & Analysis for Brain Imaging on Surface (DPABISurf) [47]. For details, please refer to the Supplementary Information.

Functional connectome and gradient mapping

Following data preprocessing, cortical data were mapped onto the cortical parcellation proposed by Schaefer and colleagues [48], consisting of 400 regions. The 54-parcellation Tian’s subcortical atlas [49] was used to extract functional signals from the subcortical regions. Subsequently, the time series of regions of interest (ROIs) were extracted to calculate Pearson correlation coefficients, which underwent Fisher’s r-to-z transformation, yielding a 454×454 FC matrix.

We used the BrainSpace toolbox [30] run in a MATLAB R2020a (The MathWorks Inc., Natick, MA, US) environment to calculate gradient matrices for resting state, sad memory, rumination, and distraction states within each group. The input matrix consisted of the FC matrix, capturing both local and long-range connections. To impose sparsity, the top 10% of connectivity values were retained, while the remaining entries were set to zero, but the retained values were not binarized [28, 30]. A cosine similarity matrix was then calculated to quantify the similarity of connectivity profiles between each pair of ROIs, i.e., how similar the connectivity pattern of one ROI is to that of another. Subsequently, a normalized angle matrix was adjusted to eliminate negative values within the similarity matrix, following the methods outlined by Vos de Wael et al. [30] and Paquola et al. [50] We used diffusion map embedding [51, 52], a nonlinear dimensionality reduction technique, to recover a low-dimensional embedding from high-dimensional connectivity data. Diffusion maps translate connectivity similarities into distances and represent the global connectivity structure as a distribution of ROIs in the embedding space. ROIs that are strongly connected, either by many connections or few very strong connections, are close in this space, whereas weakly connected or unconnected ROIs are far apart. The parameter α was set to 0.5 to control the influence of sampling density on the embedding. This procedure ensures that the gradients capture the global structure of the functional connectome while being robust to noise. The eigenvectors of the diffusion operator define the gradients, and the corresponding eigenvalues indicate the variance explained by each gradient, i.e., the relative contribution of that gradient to the overall functional connectivity structure. To facilitate comparison across states and groups, Procrustes rotation [53] was applied to realign all the connectome gradients, and the gradients from Rum-Beijing and Rum-MDD were aligned to a common reference, ensuring that corresponding gradients share the same orientation and scale across subjects.

Furthermore, several quantitative metrics of the global feature of each gradient were computed. The explanation ratio represents the proportion of variance explained by each gradient, calculated as the normalized eigenvalue (i.e., the eigenvalue of a given gradient divided by the total sum of eigenvalues). This metric reflects the relative contribution of a specific gradient to the overall organization of the functional connectome. The gradient range was defined as the difference between the maximum and minimum gradient scores across all brain parcels along a given gradient, reflecting the spatial extent of differentiation along that functional axis. The gradient variance was calculated as the variance of gradient scores across cortical regions. This measure quantifies the degree of dispersion of regional embedding values and indicates how strongly cortical areas are differentiated along a given gradient dimension. Together, these metrics characterize both the global contribution of each gradient (explanation ratio) and the spatial distribution properties of cortical functional organization (range and variance) [54–56].

Statistical analysis

We first computed the average first and second gradients across all the subjects in the rumination/distractions condition. To test replicability, we calculated these average gradient distribution maps across all repeated scans in the Rum-Beijing dataset and across both groups in the Rum-MDD dataset.

We further examined the network-level gradient features by aggregating gradients across all subjects, grouping the 400 parcellations according to Yeo’s 7-network atlas, then calculating the mean gradient values within each network. All subcortical region gradient values were averaged separately to form a subcortical network. We also examined the gradient values at the node level. A number of global metrics, including explanation ratio, range, and variance of both gradients, were examined.

We applied paired t-tests by fitting a general linear model (GLM) to test the condition effect (rumination vs. distraction) in both Rum-Beijing and Rum-MDD datasets. T-tests were performed on network-wise and regional gradient values at the parcellation level, with the gradient values as the dependent variable. For the Rum-Beijing dataset, t-tests were conducted separately for each scanner to verify that the observed effects were consistent across sites. An effect was considered replicable if it was statistically significant in at least one measure or group (MDD or HC) and showed at least moderate effects in the same direction (increase or decrease) across all other measures and groups. Previous studies in the gradient literature have demonstrated that t-tests can yield robust and reliable results.

The head motion was included as a covariate. In the Rum-MDD dataset, we further examined the interaction effect between the condition (rumination vs. distraction) and the group (MDD vs. HC) by fitting a linear mixed-effects model (LMM). Because the Rum-Beijing dataset included healthy participants only and did not involve a group factor, no LMM was performed for that dataset, and no interaction results were generated or reported.

$$\mathrm{y}~\mathrm{Group}*\mathrm{Condition}+\mathrm{H}\mathrm{ead}\mathrm{m}\mathrm{otion}+\left(1\mathrm{|Subject}\right)$$ 1

An independent sample t-test was performed to examine the group effect in the rumination and distraction conditions, with head motion controlled as the additional covariate. All analyses were conducted in R (version 4.4.2). All statistical tests were two-tailed. The False Discovery Rate (FDR) correction was used for multiple comparisons, with separate corrections applied to all regional-level and network-level comparisons.

Spatial correlation and spin permutation analysis

To assess cortical gradient similarity between rumination or distraction and resting states, parcel-wise spatial correlation analyses were conducted using the Schaefer-400 atlas on the fsaverage5 spherical surface [57]. For each participant in the MDD group, gradient maps were extracted for the primary-transmodal gradient and the visual-sensorimotor gradient. Parcel centroids were computed by averaging the 3D coordinates of all vertices within each parcel, excluding vertices in the medial wall or with unassigned labels (label = 0), following established procedures for surface-based parcellations [48]. Left and right hemisphere centroids were combined to obtain a set of 400 cortical parcels.

Pearson correlation coefficients were calculated between the resting-state and rumination/distraction gradient maps, separately for the primary-transmodal gradient and the visual-sensorimotor gradient. To account for spatial autocorrelation, spin permutation tests were performed using the BrainSpace toolbox, with 10,000 random rotations applied to the cortical centroids. Observed correlations were compared against the null distributions to obtain spin-test p-values.

FDR correction was applied separately for the primary-transmodal gradient and the visual-sensorimotor gradient to control for multiple comparisons. Additionally, a more conservative approach was tested by combining the p-values of the first two gradients and applying FDR across all parcels. All analyses were conducted in Python 3.13.5 using numpy, pandas, scipy, and BrainSpace (0.1.16) [30].

Results

Connectome gradient mapping during rumination

The first two gradients accounting for the greatest variance were selected for the present study. Generally speaking, we found that the overall gradient architecture revealed by previous studies [28, 35] remains during both rumination and distraction states in MDD patients and HCs. As illustrated in Fig. 1, the first gradient was organized along an axis from the primary visual/sensorimotor area to the association area (named “primary-transmodal gradient” henceforth). Meanwhile, the second gradient is anchored at one end by the visual areas and by the sensorimotor area at the other end (named “visual-sensorimotor gradient” henceforth, Figs. 1 & 2). These gradient architectures were highly reproducible across all different conditions, scanners, and datasets.

Fig. 1. Distribution of cortical gradients during rumination and distraction states.

Distribution of the primary-transmodal gradient and the visual-sensorimotor gradient of the rumination and distraction states in 3 sites in the Rum-Beijing and the Rum-MDD datasets.

Fig. 2. Scatter plots of the gradient score for the first two connectome gradients in rumination states in 3 sites in the Rum-Beijing and the Rum-MDD datasets.

Each dot was colored according to 7 cortical networks and a subcortical network. Abbreviations: SMN, somatomotor network; VN, visual network; DAN, dorsal attention network; VAN, ventral attention network; LN, limbic network; FPN, fronto-parietal network; DMN, default mode network; SC, subcortical network.

Spatial correlation between resting and task gradients

In both the Rum-Beijing and Rum-MDD datasets, parcel-wise spatial correlations between resting-state and rumination/distraction gradient structures were consistently significant for both primary-transmodal gradient and visual-sensorimotor gradient (spin permutation ps < 0.05, FDR-corrected, see Figure S3). These findings indicate a robust similarity of cortical gradient patterns between resting and task states across both cohorts.

Network-wise gradient alterations during rumination

We first displayed the ranges of network-wise gradients during the rumination and distraction states, using both the Rum-Beijing and Rum-MDD datasets (Fig. 3A). As expected, all networks exhibited distributed gradients along the primary-transmodal gradient, with the DMN at one end and the SMN at the other. On the other hand, the scores on the visual-sensorimotor gradient clearly discriminated between the VN and SMN, with these two networks on opposite sides of this gradient.

Fig. 3. Network-level distribution and alterations of cortical gradients during rumination and distraction states.

A) The range of gradient values across the 7 networks and a subcortical network in all 3 sites in the Rum-Beijing dataset and in the Rum-MDD dataset in the primary-to-transmodal gradient, as well as the visual-sensorimotor gradient. B) Bar plot showing network-wise comparison of gradient value differences between rumination and distraction states. Note the replicability of the reduced DMN’s primary-transmodal gradient values in the rumination condition as compared to the distraction condition. Abbreviations: SMN, somatomotor network; VN, visual network; DAN, dorsal attention network; VAN, ventral attention network; LN, limbic network; FPN, fronto-parietal network; DMN, default mode network. *: q_FDR < 0.05.

Next, we examined network-level alterations in gradient values between the rumination and distraction conditions. Specifically, for the primary-transmodal gradient, we found reduced mean DMN gradient values during rumination as compared to distraction. This effect could be reproduced in both datasets (IPCAS: t(39) = −1.509, q_FDR = 0.223, Cohen’s f² = 0.074; PKUGE: t(39) = −2.721, q_FDR = 0.026, Cohen’s f² = 0.252; PKUSIEMENS: t(39) = −1.446, q_FDR = 0.249, Cohen’s f² = 0.065; Rum-MDD HC: t(44) = −2.904, q_FDR = 0.021, Cohen’s f² = 0.221; Rum-MDD MDD: t(43) = −2.503, q_FDR = 0.031, Cohen’s f² = 0.214; Fig. 3B).

Regional gradient alterations during rumination

We first examined the interaction effect between group (MDD vs. HC) and condition (rumination vs. distraction) in the Rum-MDD dataset, but found no significant interaction or group effect.

We further performed paired t-tests at the parcellation level to delineate the condition effect. For surface regions, we observed reproducible spatial patterns along the primary-transmodal gradient. Across healthy participants in both the Rum-Beijing and Rum-MDD datasets, compared to the distraction state, the rumination state exhibited decreased gradient values in the posterior cingulate cortex (PCC) (all q_FDRs < 0.01) and increased gradient values in the left dorsolateral prefrontal cortex (DLPFC) and superior frontal gyrus (SFG). Patients with MDD showed similar alterations in the primary-transmodal gradient. Along the visual–sensorimotor gradient, reproducibly greater gradient values in the left SFG and right superior occipital lobe were observed during rumination as compared to distraction (Fig. 4). Overall, rumination-related gradient alterations were similar in patients with MDD and HCs. However, reproducible effects were found in only a small proportion of brain regions.

Fig. 4. Regions with significant differences between the rumination state and the distraction state in primary-transmodal gradients and visual-sensorimotor gradients in all 3 sites in the Rum-Beijing dataset and in the Rum-MDD dataset.

Abbreviations: rum., rumination; dis., distraction.

For sub-cortical regions, significant condition effects were observed for the primary-transmodal gradient. In both the PKUGE and PKUSIEMENS datasets, rumination was associated with significantly higher gradient values in the putamen (PKUGE: t(39) = 2.972, q_FDR = 0.005, Cohen’s f² = 0.227; PKUSIEMENS: t(39) = 2.866, q_FDR < 0.001, Cohen’s f² = 0.370) and caudate (PKUGE: t(39) = 3.209, q_FDR = 0.003, Cohen’s f² = 0.264; PKUSIEMENS: t(39) = 2.866, q_FDR < 0.001, Cohen’s f² = 0.370; Figure S5) area compared to the distraction condition. Importantly, the effect in the caudate was also replicated in the MDD group of the Rum-MDD dataset (t(43) = 3.049, q_FDR = 0.004, Cohen’s f² = 0.216, Figure S5), where rumination was associated with higher gradient values compared to the distraction condition. The consistent directionality of effects across independent scanners and cohorts suggests a robust involvement of dorsal striatal structures in rumination-related gradient reorganization. Meanwhile, no reproducible condition effects were identified for the visual–sensorimotor gradient; thus, the findings for the visual–sensorimotor gradient were not further reported.

Gradient alterations regarding the global metrics during rumination

For the primary-transmodal gradient, the Rum-Beijing dataset consistently showed a lower gradient range during rumination compared with distraction (IPCAS: t(39) = −2.291, q_FDR = 0.027, Cohen’s f² = 0.179; PKUGE: t(39) = −2.202, q_FDR = 0.031, Cohen’s f² = 0.061; PKUSIEMENS: t(39) = −2.060, q_FDR = 0.046, Cohen’s f² = 0.145; Rum-MDD HC: t(44) = −3.034, q_FDR = 0.003, Cohen’s f² = 0.103; Rum-MDD MDD: t(43) = −1.524, q_FDR = 0.134, Cohen’s f² = 0.092), although this effect did not reach statistical significance at all sites. Similarly, for the visual–sensorimotor gradient, both datasets demonstrated a consistent increase in gradient range during rumination relative to distraction (IPCAS: t(39) = 3.163, q_FDR = 0.006, Cohen’s f² = 0.359; PKUGE: t(39) = 4.043, q_FDR < 0.001, Cohen’s f² = 0.389; PKUSIEMENS: t(39) = 5.851, q_FDR < 0.001, Cohen’s f² = 0.747; Rum-MDD HC: t(44) = 3.372, q_FDR = 0.003, Cohen’s f² = 0.345; Rum-MDD MDD: t(43) = 4.664, q_FDR < 0.001, Cohen’s f² = 0.676; Fig. 5). Notably, individuals with MDD exhibited a significantly elevated primary–transmodal gradient range relative to healthy controls (t(88) = 2.755, q_FDR = 0.014, Cohen’s f² = 0.224, Table 2).

Fig. 5. The differences in gradient range of the primary-transmodal and the visual-sensorimotor gradient between the rumination state and the distraction state of the Rum-Beijing dataset and the Rum-MDD dataset.

Abbreviations: rum., rumination; dis., distraction; HC, healthy control; MDD, major depressive disorder. *: p < 0.05; **: p < 0.01; ***: p < 0.001.

Table 2.

Differences in gradient range between rumination and distraction in the Rum-Beijing and the Rum-MDD datasets.

		IPCAS			PKUGE			PKUSIEMENS			HC			MDD
		t	p	Cohen’s f2	t	p	Cohen’s f2	t	p	Cohen’s f2	t	p	Cohen’s f2	t	p	Cohen’s f2
Gradient Range	primary-transmodal gradient	−2.291	0.027	0.179	−2.202	0.031	0.061	−2.060	0.046	0.145	−3.034	0.003	0.103	−1.524	0.134	0.092
	visual-sensorimotor gradient	3.163	0.006	0.359	4.043	<0.001	0.389	5.851	<0.001	0.747	3.372	0.003	0.345	4.664	<0.001	0.676

HC healthy control, MDD major depressive disorder.

In addition, for the primary–transmodal gradient, both the explanation ratio and gradient variance showed a consistent decreasing trend during the rumination state in healthy participants, whereas this pattern was not replicated in individuals with MDD. For the visual–sensorimotor gradient, the explanation ratio and gradient variance similarly exhibited a consistent increase during rumination across all groups. No significant interaction effect was observed.

Discussion

This is the first study to delineate alterations in functional gradients during an active ruminative state in both healthy adults and patients with MDD. In two independent datasets, we consistently observed that the brain’s two prominent macro-scale organizational dimensions, the primary-to-transmodal gradient and the visual-sensorimotor gradient, remained in the rumination state. We further found that large-scale brain networks were distributed along a gradient in a specific order similar to that observed in the resting state. Importantly, we found a reduced DMN score in the primary-to-transmodal gradient during rumination in both HCs and patients with MDD. We observed that during the rumination state, individuals exhibited a consistent and significant reduction in gradient range along the primary–transmodal gradient, accompanied by a significant enhancement along the visual–sensorimotor gradient. Moreover, the MDD group showed a significantly higher primary–transmodal gradient range than healthy controls.

Recent evidence has shown that the human brain’s activity exhibits macroscale spatial organization, characterized by gradients [29]. Such a cortical hierarchy might serve as an organizational axis for controlled psychological processes, including depressive rumination [58]. A question remained, though, whether or to what extent gradients would be altered in an active rumination state. Here, using two independent datasets, we found a replicable, significant decrease in DMN gradient values during the rumination state along the primary-transmodal gradient compared to the distraction state. In previous studies, patients with MDD have been observed to exhibit lower DMN gradient values than healthy controls [35]. From the perspective of cortical gradient architecture, a decrease in the DMN gradient value indicates reduced functional differentiation within the brain’s macroscale topology, suggesting a shift toward lower-order, more sensorimotor-like functional characteristics [28, 36]. This finding adds to a growing literature highlighting the prominent role of the DMN in the neural mechanisms underpinning rumination [24, 38, 39, 59–63]. Such reductions in gradient scores may indicate a more dynamic involvement of these regions in self-referential and past-oriented cognitive processes, which are core features of ruminative thinking [21, 25].

Regionally, in the Rum-Beijing dataset, the rumination state was characterized by a lower gradient score in the PCC within the DMN. This change observed in participants during the rumination state suggested decreased dissimilarity in the embedded FC pattern between the primary and transmodal systems [35, 54]. Previous studies have proposed that the DMN has three subsystems: a midline core system, a dorsal medial prefrontal cortex (DMPFC) subsystem, and a medial temporal lobe (MTL) subsystem [39]. The results of the current study mainly involved the midline core system and the DMPFC system. According to previous investigations, the midline core system is theorized to integrate the functions of other subsystems and is implicated in introspective processes regarding one’s own mental states. And the DMPFC subsystem is associated with processes such as mentalizing and metacognition [64]. The reduction in the gradient within the PCC region indicates increased similarity in its functional connectivity patterns with other brain regions, highlighting the pivotal role of the PCC in rumination. For instance, higher connectivity between the amygdala and PCC was correlated with increased rumination, while alterations in connectivity between the left PCC and right inferior temporal gyri were linked to changes in depressive symptoms and rumination [65, 66]. This is consistent with the central role of the midline core system in integrating other subsystems’ functions, as well as the crucial role of self-referential processing in rumination [17, 67]. Compared with the distraction state, previous research has found a reproducible decrease in FC between the core and the DMPFC subsystem during rumination [39]. It can be inferred that the functional differences between these two subsystems are amplified in the ruminating state.

We also observed replicable increases in the primary-transmodal gradient in the left DLPFC, one of the key nodes of the FPN. Such enhancement may suggest reduced interaction of FPN with primary regions, thereby displaying a more unique functional connectivity pattern and implying decreased involvement of executive control functions during the rumination state. Importantly, DLPFC, along with FPN, plays a critical role in inhibitory control, and one potential mechanism underlying rumination is a deficiency in this capacity. Therefore, the reduced engagement of the FPN in executive control during the rumination state may indicate weakened inhibitory control, which could in turn contribute to persistent self-focus and negative ruminative thought patterns [16, 68, 69]. Likewise, the higher gradient values observed in the DLPFC during the rumination state may suggest reduced capacity to regulate and manage spontaneous thought processes, potentially leading individuals to become trapped in repetitive negative thinking [70]. Moreover, our findings support previous findings that an imbalance between these two networks appears to be critical to MDD’s cognitive impairment [71]. Our results also added to existing literature, which showed a reduction in FCs between the SFG region and both the DAN and FPN during rumination in MDD patients [40], as well as an observed correlation between repetitive negative thinking and decreased resting-state functional connectivity (rsFC) between the DMN and FPN [72].

During the rumination state, we observed a consistent and significant decrease in gradient range along the primary–transmodal gradient, reflecting a compression of the principal functional hierarchy [33, 55, 73]. This suggests reduced differentiation between unimodal sensory regions and high-order transmodal regions, indicative of a shift of neural resources from externally oriented processing toward introspective, self-referential activity, consistent with the role of the default-mode network in rumination [18, 74, 75]. Simultaneously, the visual–sensorimotor gradient exhibited a significant increase in gradient range, indicating enhanced functional segregation between VIS and SMN systems [54, 55]. Previous studies have shown that ruminative thinking is closely associated with depressive cognition manifested in the form of inner speech [45, 76]. Along this gradient, these two networks became more spatially separated and functionally decoupled, implicating disengaged VIS and SMN during rumination. Such a configuration may support the maintenance of repetitive, internally focused thought by reducing integration between perceptual and motor systems [28, 77–79]. Notably, a significant between-group difference was observed, with individuals with MDD exhibiting a higher primary–transmodal gradient range than healthy controls. Importantly, we observed an increased range of the primary–transmodal gradient in patients with MDD relative to healthy controls during rumination. This pattern suggests that transmodal regions exhibit more idiosyncratic functional connectivity profiles and are relatively decoupled from unimodal systems. Such decoupling may contribute to the abstract, decontextualized nature of ruminative thought in MDD [1]. Thinking that lacks specific contextual grounding has been proposed to be particularly detrimental to mental health [80]. In contrast, situating thoughts within a concrete temporal and spatial context may depend on tighter coupling between unimodal and transmodal networks. These gradient alterations may underlie the persistent rumination and heightened negative affect observed in MDD, highlighting a disorder-specific reorganization of cortical hierarchy that contrasts with the gradient compression typically seen in healthy individuals during internally focused states [25, 28, 29, 81, 82]. Taken together, the combination of compressed primary–transmodal gradient and expanded, decoupled visual–sensorimotor gradient suggests that rumination is associated with a large-scale reorganization of functional gradients, shifting the brain from an externally integrated, perceptually grounded mode of processing to a more internally oriented, introspective state during rumination. These findings provide a macroscale, network-level explanation for the cognitive characteristics of rumination, and highlight the importance of considering gradient-based measures to understand dynamic functional reconfiguration during both adaptive and maladaptive internally focused states [36, 39].

Interestingly, no significant difference in gradient values was found between MDD and HC during the distraction state in the Rum-MDD dataset. Distraction involves diverting attention, attempting to manage attention, and reducing unpleasant feelings [83]. It has been reported that in a rumination state, negative emotions were sustained by narrowing attention span and leading to difficulty separation from information when it is no longer relevant. And this effect can be most effectively controlled by the distraction state over a short period of time [69]. Research has shown that a distraction state in the form of mental imagery resulted in greater affective improvement and higher high-frequency heart rate variability, without a significant difference in the skin conductance response [83, 84]. This evidence suggests that distraction may be utilized as a vital measure to alleviate the adverse reactions caused by the rumination state and may be applied to the clinical treatment of depression.

One strength of our study is the replicability of findings across multiple measures and independent sites. Nevertheless, variability across datasets was observed, which may reflect interindividual and sample-related heterogeneity. Notably, the alignment step in gradient computation can substantially influence cross-site reproducibility, highlighting the necessity of maintaining consistent computational procedures when conducting multi-cohort analyses on gradient. Therefore, we performed a Procrustes alignment of the two datasets to ensure that the gradients were directly comparable across cohorts. The test–retest reliability of the gradient structure is an important topic that remains to be investigated. Some initial evidence suggested moderate reliability of gradient metrics [85]. Future studies that comprehensively evaluate the reliability and reproducibility of gradients are warranted.

The present study comes with certain limitations. The Schaefer’s 400 template, applied here for whole-brain gradient calculation, could be too coarse. Future studies could employ vertex-level FC maps to calculate gradients. Furthermore, the absence of a correlation between gradient scores and behavioral data in this study highlights an avenue for improvement. Subsequent investigations may benefit from employing a broader range of meta-correlations to explore the behavioral relevance of gradient scores. Moreover, the potential effects of medication should be taken into account. Lastly, given the pivotal role of the DMN in the gradient distribution of the functional connectome and its close relation to rumination, future studies could further delineate the gradient structure within the DMN.

In sum, we found that both healthy subjects and MDD patients maintained the primary-transmodal gradient and visual-sensorimotor gradients in both the rumination and distraction states. At the network level, convergent alterations in network gradients emerged during rumination, characterized by replicable decreases in DMN gradient values along the primary-transmodal gradient. Furthermore, the present findings showed that rumination is associated with a compressed primary–transmodal gradient, reflecting reduced hierarchical differentiation, and an expanded, decoupled visual–sensorimotor gradient, reflecting increased functional segregation between VIS and SMN. Our results collectively suggested a large-scale reorganization of functional gradients from an externally integrated, perceptually grounded mode toward a more internally oriented, introspective state. Critically, we observed that individuals with MDD exhibited a significantly greater primary–transmodal gradient range than HCs. These findings suggested dysfunctional large-scale network reorganization in patients with MDD, which shed new light on the neural mechanisms of depressive rumination.

Author contributions

Concept and design: Chao-Gan Yan, Xiao Chen. Data acquisition: Xiao Chen, Feng-Nan Jia, Yan-Song Liu. Analysis and interpretation of data: Zheng-Jia-Yi Hu, Xiao Chen, Chao-Gan Yan. Drafting of the manuscript: Zheng-Jia-Yi Hu, Xiao Chen. Critical revision of the manuscript for important intellectual content: Zheng-Jia-Yi Hu, Feng-Nan Jia, Yan-Song Liu, Xiao Chen, Chao-Gan Yan. Statistical analysis: Zheng-Jia-Yi Hu, Xiao Chen. Administrative, technical, or material support: Xiao Chen, Chao-Gan Yan. Supervision: Xiao Chen, Chao-Gan Yan.

Funding

This work was funded by the National Natural Science Foundation of China (No. 32300933, No. 82122035, No. 81671774, and No. 81630031), the Beijing Nova Program of Science and Technology (No. 20230484465), the Beijing Natural Science Foundation (No. J230040), the Sci-Tech Innovation 2030 - Major Project of Brain Science and Brain-inspired Intelligence Technology (No. 2021ZD0200600), the Scientific Foundation of Institute of Psychology, Chinese Academy of Sciences (No. E2CX4425YZ, No. E3CX1315 and No. Y9CX422005).

Data availability

The Rum-Beijing dataset is publicly available at https://rfmri.org/RuminationfMRIData. The Rum-MDD dataset is publicly available through the Open Science Framework at https://doi.org/10.17605/OSF.IO/WTS7F.

Code availability

Analysis codes are available online (https://github.com/sameeniii26/Gradient_Rumination/tree/main).

Competing interests

The authors declare no competing interest.

Ethics approval and consent to participate

All methods were performed in accordance with the relevant guidelines and regulations and the principles of the Declaration of Helsinki. The experimental procedures were approved by the Institutional Review Board of the Institute of Psychology, Chinese Academy of Sciences (Approval No. H17019). Written informed consent was obtained from all participants prior to their participation in the study.

Supplementary information

The online version contains supplementary material available at https://doi.org/10.1038/s41398-026-04294-x.