Abnormal ventromedial-to-dorsolateral hierarchical topography of striatal circuits in cocaine use disorder and its modulations by brain stimulation

Abstract

Cocaine use disorder (CUD) has been linked to cortico-striatal dysfunctions, particularly within the prefrontal-striatal circuitry. However, previous studies have typically focused on discrete parcellations of the striatum, overlooking its continuous variations of neural organization. Moreover, while repetitive transcranial magnetic stimulation (rTMS) has shown benefits in CUD treatment, the neural effects of rTMS on striatal dysfunction in CUD remain poorly understood. Using connectome gradient-mapping techniques on three resting-state functional magnetic resonance imaging datasets, we derived the ventromedial-to-dorsolateral striatal functional topography. We identified specific alterations in this topography in the discovery cohort (41 CUD patients and 44 controls), validated findings in an independent cohort (53 CUD patients and 45 controls), and examined whether rTMS targeting the left dorsolateral prefrontal cortex (dlPFC) could normalize abnormalities in the rTMS-treatment cohort (44 patients). Across all datasets, we found a positive correlation between gradient variation and drug dependence severity in CUD. Compared to controls, CUD in both the discovery and replication cohorts exhibited elevated gradient values in the ventral striatum, while decreased values in the dorsal striatum were observed only in the discovery cohort. Furthermore, in the rTMS-treatment cohort, 5-Hz rTMS targeting the left dlPFC significantly normalized the aberrant gradient values in the ventral striatum, and these changes also related to cocaine craving changes. Overall, our study provides novel evidence of specific alterations in the ventromedial-to-dorsolateral functional topography of the striatum in CUD patients and highlights the impact of rTMS on striatal circuits through prefrontal modulation.

Introduction

Cocaine use disorder (CUD) affects approximately 0.35% of the global adult population [1] and is associated with cognitive impairment [2, 3], abnormal emotional processing [4, 5], and an increased risk of comorbid psychiatric disorders [6, 7]. These clinical manifestations underscore the urgency of elucidating the neurobiological mechanisms underlying CUD and advancing novel therapeutic strategies, such as neuromodulation.

Converging evidence from animal and human studies highlights the pivotal role of striatal dysfunction in CUD [8–11], particularly in multiple domains critically implicated in its pathogenesis, including associative learning, behavioral control, incentive salience, and reward processing [12–14]. The striatum’s functional heterogeneity and multifaceted contributions to CUD are rooted in its organization into distinct subregions, each integrated into specific fronto-striatal loops that interconnect with the entire cortex [12, 15, 16]. The dorsal striatum (DS), which receives projections from the premotor and motor cortices, is critical for regulatory control, habit formation, and the development of compulsive behaviors [12, 15–18]. In contrast, the ventral striatum (VS) receives inputs from the ventromedial prefrontal cortex (vmPFC), orbitofrontal cortex (OFC), and dorsal anterior cingulate cortex (dACC) and plays key role in reward processing, incentive-based learning, and impulsive behavior [15, 16, 19, 20]. Additionally, the dorsal prefrontal cortex (dPFC) – a region implicated in higher cognitive processes or executive functions - projects to the central striatal structures such as the head of the caudate nucleus and the rostral putamen [15], with potential extensions to the VS and lateral putamen [21]. This ventral-to-dorsal axis provides a crucial framework for understanding cognitive deficits in CUD, aligning with the dual-systems model that posits compulsive drug use arise from heightened drug-seeking motivation coupled with impaired inhibitory control mechanisms [9, 22].

Despite progress, previous methodologies have struggled to directly characterize the striatum’s ventral-to-dorsal functional topography in humans, limiting deeper mechanistic insights. Conventional resting-state functional magnetic resonance imaging (rsfMRI) studies typically employ hard parcellations to delineate striatal regions of interest (ROIs) [9, 10, 23, 24], which assume functional homogeneity across anatomically defined areas and disregard overlapping functional representations [25]. For instance, recent work revealed distinct connectivity alterations in VS subregions among individuals with cocaine dependence - changes obscured when analyzing the VS as a single seed region [23]. Emerging evidence highlights the striatum’s organization along continuous spatial gradients (ventral-to-dorsal, medial-to-lateral, and anterior-to-posterior) [25–27], aligning with tract-tracing findings in non-human primates [15, 28]. Such findings emphasize  the need for novel methodologies that capture the striatum’s continuous functional architecture in humans to advance research on CUD pathophysiology.

Gradient-based approaches offer a powerful framework for mapping continuous neural organziations [29–33]. By decomposing high-dimensional functional connectivity (FC) data into low-dimensional representations, these methods reveal hierarchical spatial patterns, positioning voxels with similar connectivity profiles along shared axes [30]. Applied to the striatum, this technique has delineated ventromedial-to-dorsolateral functional gradients [25], with demonstrated reproducibility and predictive validity for psychiatric phenotypes [27, 34].

Moreover, repetitive transcranial magnetic stimulation (rTMS) targeting the dorsolateral prefrontal cortex (dlPFC) shows promise in reducing drug use and craving [35–38], potentially via modulation of striatal circuits [39, 40]. However, mechanistic insight into the neurobiological effects of TMS remains limited. Investigating whether and how rTMS affects the striatum’s ventral-to-dorsal functional topography, particularly the central and partially ventral areas, could provide further insights.

In this study, we integrated gradient-based analysis with rsfMRI to: (1) characterize ventromedial-to-dorsolateral striatal gradient abnormalities in CUD, (2) localize specific deviant regions along this axis, and (3) evaluate whether left dlPFC-targeted rTMS modulates these aberrant gradients. As shown in Fig. 1, we calculated ventromedial-to-dorsolateral striatal gradient maps for 41 patients with CUD and compared these maps with those from 44 controls in the discovery cohort, subsequently correlating them with clinical variables. We then validated findings in an independent replication cohort (n = 98). Finally, using rTMS-treatment cohort (n = 44) we assessed effect of rTMS over the left dlPFC on the ventromedial-to-dorsolateral striatal gradient in CUD.

Fig. 1. Gradient analysis, main questions and corresponding analytic workflow.

A Connectivity gradient analysis identified the ventromedial-to-dorsolateral axis of functional organization within the striatum. B Research goals and analytic workflow of the present study. In the discovery cohort (n = 85) and replication cohort (n = 98), we aim to determine the altered metrics and areas of striatal gradient and their relationships with clinical measures in patients with CUD. In the rTMS-treatment cohort (n = 44), we examine whether rTMS targeting on the left dlPFC could modulate the striatal gradient in CUD.

Materials and methods

Participants in the discovery, replication, and rTMS-treatment cohorts

In the discovery cohort, rsfMRI data were obtained from the Mexican SUDMEX CONN dataset [41]. After excluding participants with incomplete tobacco use history, inactive cocaine consumption, or neurological and psychiatric disorders (see Supplementary Methods for full inclusion/exclusion criteria), the final sample comprised 41 CUD patients (3 females; mean age ± SD: 32.34 ± 6.67 years) and 44 controls (7 females; mean age ± SD: 30.73 ± 7.73 years). All CUD participants met DSM-IV criteria for substance dependence (symptom count ≥3) and did not have any other psychiatric disorders. Demographic and clinical characteristics, including severity metrics (e.g., years of cocaine use), are summarized in Table 1.

Table 1.

The demographic information in the discovery cohort.

Characteristic	CUD (n = 41)	Controls (n = 44)	Statistics
Sex			χ2 = 1.509 p = 0.219
Male	38	37
Female	3	7
Age	32.34 ± 6.67	30.73 ± 7.73	t = 1.027 p = 0.307
Years of education	10.97 ± 2.96	13.16 ± 3.40	t = −3.102 p = 0.003
Years of cocaine consumption	10.73 ± 7.12	/	/
Onset age of consumption	21.20 ± 5.74	/	/
Dependence score	6.25 ± 1.07	/
Years of tobacco use	14.95 ± 8.13	5.95 ± 8.02	t = 5.134 p < 0.001
Mean framewise displacement	0.20 ± 0.04	0.18 ± 0.04	t = 1.994 p = 0.049

Data are expressed as mean ± SD (SD, standard deviation). All subjects in the CUD group reported consuming cocaine at least twice a week in the month preceding the scanning session.

CUD cocaine use disorder.

To validate our findings, the replication cohort included rsfMRI data from 53 non-treatment-seeking cocaine users (11 females; mean age ± SD: 39.91 ± 7.96 years) and 45 controls (13 females; mean age ± SD: 37.93 ± 7.96 years), collected at the National Institute on Drug Abuse Intramural Research Program [9, 42]. Cocaine users met ≥1 DSM-IV-TR substance dependence criteria (criteria 3–7), reflecting moderate-to-severe dependence. None had any major illness or history of neurologic or psychiatric disorders other than current abuse/dependence on cocaine, nicotine, or both. Further demographic and clinical details are provided in Supplementary Table S1 and Supplementary Methods.

The rTMS-treatment cohort included 44 CUD patients who met the inclusion criteria such as high cocaine consumption for at least one year from the SUDMEX-TMS clinical trial [37, 43], randomized into active (n = 24; 4 females; mean age ± SD: 35.79 ± 7.05 years) or sham (n = 20; 2 females; mean age ± SD: 33.30 ± 8.66 years) rTMS groups. Notably, the same individuals were included in the study by Angeles-Valdez et al. [37], and we primarily focused on the 2-week treatment phase. During this phase, patients received 10 weekdays of active/sham 5-Hz rTMS targeting the left dlPFC (peak MNI coordinates: x = −30, y = 44, z = 38) over two weeks. Stimulation was delivered using a MagPro R301+Option device with a B65-A/P coil (MagVenture, Denmark). Each session comprised two blocks of 50 trains (5-Hz, 50 pulses/train, 10-second inter-train interval), separated by a 15-minute interval (5,000 pulses/day). Baseline and post-treatment fMRI assessments were conducted before and after the 2-week intervention. Additional details on treatment protocols, CUD severity, and psychiatric comorbidities are available in Supplementary Table S2 and Supplementary Methods.

The discovery and rTMS-treatment cohorts were approved by the Ethics Committee of the Instituto Nacional de Psiquiatría “Ramón de la Fuente Muñiz” [41, 43]. The replication cohort was obtained from subjects who consented to the protocols approved by the Institutional Review Board of the National Institute on Drug Abuse.

Clinical assessment

The studies systematically gathered standardized clinical evaluations across all three participant groups [9, 41, 43]. Here, we primarily focused on drug dependence symptoms across all cohorts and on cocaine craving and impulsivity in the rTMS-treatment cohort. For the discovery and rTMS-treatment cohorts, studies used the official Spanish version 5.0 of the Mini-International Neuropsychiatric Interview Plus (M.I.N.I.-Plus), a tool designed to assess the 17 most common Axis I psychiatric disorders according to the DSM-IV and ICD-10 [44], and we selected scores from the DSM-IV section on substance use disorders from the M.I.N.I.-Plus for detailed analysis of dependence symptoms. In the replication cohort, dependence symptoms were evaluated using the DSM-IV-TR. Moreover, the rTMS-treatment cohort used the cocaine craving visual analog scale (VAS), which consists of a 10 cm continuous line (with 2 decimal places), to assess participants’ subjective current craving, and used the Spanish translation of the Barratt impulsiveness scale version 11 (BIS-11) to examine impulsivity [45].

Image acquisition and preprocessing

For discovery and rTMS-treatment cohorts, rsfMRI data were acquired using a Philips Ingenia 3 T MR system with the following parameters: repetition time (TR) /echo time (TE) = 2000/30 ms, flip angle = 75°, matrix = 80 × 80, field of view = 240 mm², voxel size = 3 × 3 × 3 mm³, number of slices = 36. For replication cohort, rsfMRI was scanned using a Siemens Allegra 3 T MR scanner with the following parameters: TR/TE = 2000/27 ms, flip angle = 77°, voxel size = 3.44 × 3.44 × 4 mm³, number of slices = 39. We used fMRIPrep 21.0.0 (RRID: SCR_016216) [46] and SPM 12 (v7771; https://www.fil.ion.ucl.ac.uk/som/software/spm12/) for preprocessing across all cohorts, including slice timing correction, motion correction, alignment with structural images, spatial normalization to MNI152NLin2009cAsym space, smoothing, nuisance regression, motion scrubbing, and bandpass filtering. Additional image acquisition, preprocessing, and quality control details are provided in the Supplementary Methods.

Connectivity gradient analysis and group comparisons

In line with previous research [25–27], striatal connectomes were generated from time series, resulting in 3835 × 3835 FC matrices at the voxel level. Masks for the striatum were derived by thresholding the respective regions from the Harvard-Oxford atlas at a 25% probability threshold. The top 10% of row-wise connections were retained, and the cosine similarity of connectivity patterns between nodes was calculated. To capture the gradient components explaining the variance in the connectivity pattern of the functional connectome, we applied the diffusion map embedding approach [47], a non-linear dimensionality reduction technique implemented in the Brainspace toolbox (http://github.com/MICA-MNI/BrainSpace). Consistent with previous studies [30, 33, 48], we set the manifold learning parameter α to 0.5 during the diffusion process to preserve the global relationships among data points in the embedded space. To ensure the comparability and reliability of the results across different cohorts, striatal template gradients were generated from group-average rsfMRI connectomes of participants in an independent dataset [49]. Individual gradient maps were aligned with reference gradients using Procrustes rotation. Given that the ventromedial-to-dorsolateral axis of functional organization within the striatum is a key feature in the pathology of addiction, we primarily focused on alterations in this gradient mode related to CUD. Additionally, we calculated three global metrics: gradient explanation ratio, range, and variation. Details of the connectivity gradient analysis are provided in the Supplementary Methods.

To better characterize the global altered distribution of ventromedial-to-dorsolateral hierarchical gradient, global histogram analyses were performed. Based on findings, we tested for a linear correlation between X and (Y-absolute(X)) per corresponding voxel (X = CUD group gradient values, i.e., orange histogram of Fig. 2B, Y = controls group gradient values, i.e., blue histogram of Fig. 2B, for similar approach see Dong et al. [33]). Two-sample t tests were conducted on the explanation ratios, gradient ranges, and gradient variations, as well as on individual gradient maps, with age, sex, and years of tobacco use regressed out to determine specific gradient differences between the CUD and controls groups. For voxel-wise gradient maps, results were reported after correcting for multiple comparisons using permutation-based threshold-free cluster enhancement (TFCE) in PALM (Permutation Analysis of Linear Models) within FSL [50], with a threshold of TFCE p < 0. 05 (10,000 permutations) and family-wise error (FWE) corrected at peak-level. A polar plot was generated to illustrate the proportion of overlaps between the significant clusters identified in our analysis and the well-established striatal parcellation atlas from Pauli et al. [51], thereby identifying the specific locations of the alterations.

Fig. 2. Ventromedial-to-dorsolateral striatal functional gradient in patients with CUD and controls in the discovery cohort.

A The second-order striatal gradient who expands along a gradual axis from the ventromedial to the dorsolateral striatal areas. B Alterations in the ventromedial-to-dorsolateral striatal gradient were observed in patients with CUD in the discovery cohort. Specifically, the global histogram showed the prominent expansion of the lowest and highest portion of the striatal gradient in the CUD group compared to the controls group, and the violin plot displayed the case-control differences in the global matrices of the striatal gradient. Voxel-wise statistical comparison between the CUD and controls groups was displayed in the lowest panel (TFCE-FWE corrected p < 0.05). Higher/lower gradient values in CUD are presented as warm/cold colors. Arrows in the raincloud plot indicate the direction of the significant differences between the patients with CUD and controls. CUD cocaine use disorder, VS ventral stratum, Ca anterior caudate, Pa anterior putamen, Cp posterior caudate, Pp posterior putamen, **P < 0.01, ***P < 0.001.

To further validate the reliability of our findings, we repeated the group comparisons by considering additional potential influencing factors, such as years of education and mean framewise displacement (FD), as detailed in the Supplementary Methods. Additionally, we included an independent sample set to confirm the reproducibility of the main analyses. In this replication cohort, the ventromedial-to-dorsolateral striatal gradient differences between the CUD and controls groups were examined using the same pipeline as in the discovery cohort.

Associations between gradient variables and clinical characteristics

To investigate the relationship between CUD-related gradient alterations and clinical characteristics, we calculated partial correlations between gradient variables showing significant between-group differences and the following measures: (1) symptoms of drug dependence and (2) the average weekly dose of cocaine consumed over the past year. Notably, in the report below, a p value < 0.05 was considered statistically significant, while a p value in the range of 0.05 ≤ p < 0.1 was regarded as indicative of a trend to significance.

Gradient alterations-based FC analysis

To determine the altered ventromedial-to-dorsolateral striatal gradient underlying FC effects, we employed seed-based FC analyses (Supplementary Methods).

Effects of rTMS over the left dlPFC on the ventromedial-to-dorsolateral striatal axis

We aim to evaluate the effect of 10 sessions of 5-Hz rTMS targeting the left dlPFC on the aberrant ventromedial-to-dorsolateral hierarchical topography of striatal circuits in CUD. To this end, ventromedial-to-dorsolateral gradient maps were generated for the rTMS-treatment cohort. Gradient values were then extracted from two predefined ROIs (i.e., VS, peak MNI center coordinates: x = −12, y = 12, z = −6, and DS, MNI center coordinates: x = −14, y = 2, z = 22, 4 mm radius) for each participant across different sessions. The effects of rTMS on the gradient values within ROIs were analyzed using a 2 × 2 mixed-effects model to examine the interaction between group (active vs. sham) and session (baseline vs. post-treatment). Significant interactions were further investigated using post hoc tests. Specifically, independent t-tests were conducted within each ROI to compare the active and sham groups at both baseline and post-treatment. Of note, age, sex, and education years were included as covariates in repeated statistical analyses to control for potential confounding factors. Additionally, building on the findings of Angeles-Valdez et al. [37], which demonstrated that active rTMS treatment resulted in greater reductions in craving and impulsivity than placebo, we employed Spearman correlation analysis to examine the relationship between the changes in ventral striatal gradients (i.e., post-treatment gradient values of VS minus baseline gradient values of VS) and changes in these significant clinical variables (i.e., post-treatment craving VAS minus baseline craving VAS, and post-treatment impulsivity minus baseline impulsivity) in active and sham groups, respectively. Finally, to determine whether the alterations in the abnormal gradient areas were mediated by changes in FC following rTMS treatment, we performed seed-based FC analysis using seeds corresponding to the average target cones within the left dlPFC, as identified in a previous study [37] (Supplementary Methods).

Meta-analytic decoding of altered ventral striatal regions

We conducted a functional decoding analysis based on Neurosynth [52] to further explore the functional characterization of these CUD-related alterations. The spatial associations between CUD-related gradient alterations and the meta-analytic map of each cognitive term in the database was estimated.

Results

Abnormal ventromedial-to-dorsolateral striatal gradients in CUD and its relationship with clinical performances

Consistent with previously identified striatal gradient patterns in healthy individuals [25], we found that the second-order gradient of striatal FC delineates a ventromedial (red voxels) to dorsolateral (blue voxels) gradient in both CUD and controls groups (Fig. 2A), accounting for 11.04 ± 1.20% of the total connectivity variance in the discovery cohort. Visual inspection of the global histogram revealed a notable expansion in the lowest and highest portions of this striatal gradient in CUD group (Fig. 2B). This expansion was further quantified, showing a substantial correlation between the gradient values in the CUD group and the absolute differences from the controls group, with a correlation coefficient of 0.818 (p < 0.001). In the replication cohort, the second-order gradient of striatal FC exhibited a similar spatial distribution to that observed in the discovery cohort, accounting for 10.40 ± 1.50% of the total connectivity variance (Fig. 3A). However, the expansion of the ventromedial-to-dorsolateral striatal gradient in the CUD group was not as pronounced as in the discovery cohort.

Fig. 3. Ventromedial-to-dorsolateral striatal gradient in patients with CUD and controls in the replication cohort.

A The ventromedial-to-dorsolateral striatal gradient mode was replicated in both groups in the replication cohort. B Voxel-wise statistical comparisons between the CUD and controls groups in the replication cohort (TFCE-FWE corrected p < 0.05). Higher gradient values in CUD are presented as warm colors. Arrows in the raincloud plot indicate the direction of the significant differences between the CUD patients and controls. The areas involving increased dark red voxels in CUD patients were located very close to the discovery and replication cohorts. CUD cocaine use disorder, VS ventral striatum.

As shown in Fig. 2B significant differences in the range (t = 3.308, Cohen’s d = 0.718, p < 0.01) and spatial variance (t = 3.650, Cohen’s d = 0.792, p < 0.001) of the ventromedial-to-dorsolateral striatal gradient were observed between the CUD and controls groups in the discovery cohort, after controlling for age, sex, and years of tobacco use (details in Table S3). Partial correlation analysis further revealed a positive correlation between the spatial variance of the gradient and the severity of dependence symptoms in CUD, even after adjusting for age, sex, and years of tobacco use (r = 0.332, p = 0.047, two-tailed permutation test; Fig. S1). Although the range and spatial variance of the striatal gradient did not show statistical differences between CUD and HC in the replication cohort (Table S4), we found the trend to significant correlations between gradient variance and dependence symptoms in both the replication and rTMS-treatment cohorts (replication: r = 0.249, p = 0.097; rTMS-treatment: r = 0.291, p = 0.064; both assessed using a two-tailed permutation test; Fig. S2).

Regionally, the CUD group exhibited higher gradient values in the left ventral striatal regions, including NAc and anterior caudate, while lower values were observed in the left DS, specifically the posterior caudate, compared to the controls group in the discovery cohort (Fig. 2B, p < 0.05, TFCE-FWE corrected). Additionally, we found that the average weekly cocaine dose over the past year was associated with reduced gradient values in the left dorsal striatal regions (r = −0.370, p = 0.024, two-tailed permutation test; Fig. S1) and, to some extent, correlated with increased gradient values in the left ventral striatal areas (r = 0.293, p = 0.078, two-tailed permutation test; Fig. S1). In the replication cohort, regional comparisons largely replicated these main findings, identifying higher gradient values in CUD, particularly in the left anterior putamen and anterior caudate and NAc (p < 0.05, TFCE–FWE corrected; Fig. 3B). However, the lower gradient values observed in the left dorsal striatal regions were not replicated, likely due to the lower severity of drug dependence in the replication cohort (mean ± SD = 4.491 ± 1.577) compared to the discovery cohort (mean ± SD = 6.256 ± 1.069). As data on weekly cocaine intake was not available in the replication cohort, we did not assess the relationship between cocaine consumption and striatal gradient alterations in this sample.

To further validate our findings, we controlled for additional potential confounding variables (e.g., years of education and mean FD) when comparing the ventromedial-to-dorsolateral striatal gradient. The CUD-related alterations in the ventromedial-to-dorsolateral striatal gradient remained highly consistent with the main results across various analyses (details in Supplementary Results, Table S5, and Fig. S7).

Notably, the exploratory analysis demonstrated that abnormal areas within the ventromedial-to-dorsolateral striatal gradient showed aberrant connectivity to prefrontal regions in CUD (see Supplementary Results and Fig. S3).

Effects of rTMS targeting the dlPFC on modulating abnormal striatal gradient in CUD

We investigated whether rTMS targeting the left dlPFC could change abnormal striatal functional gradient. A significant group × time interaction effect was observed (F _(1,42) = 7.588, p = 0.009, partial η² = 0.153, Fig. 4A) in the left VS. Post-hoc tests revealed that the active group exhibited significantly reduced gradient values in the left VS compared to the sham group at post-treatment (t _(1,42) = 2.202, p = 0.033, Fig. 4B), bringing them closer to the values observed in controls (also see Supplementary Results). These results remained significant after controlling for age, sex, and education years. However, no significant alterations were observed in the left DS (Fig. S4). Previous work has demonstrated that active 5-Hz rTMS over two weeks resulted in greater reductions in craving, as measured by VAS, and impulsivity, as assessed by BIS-11, compared to sham stimulation [37]. In this work, we found that changes in ventral striatal gradients correlated with changes in craving VAS significantly in active group (r = 0.385, p = 0.043, one-tailed; Fig. 4C) after controlling for age, sex, and education years. However, this correlation was not observed in the sham group (r = 0.080, p = 0.620, one-tailed). Additionally, in two groups, we found no significant associations between changes in ventral striatal gradients and changes in impulsivity.

Fig. 4. Gradient alterations in ventral striatum for the 2-week acute phase.

A Plot on the left shows the ventral striatum ROI. The raincloud plot on the right shows the gradient values in each time point, where each data point represents the gradient value in each subject. B The violin plots illustrate the gradient values for each group at two respective time points. C Relationship between changes in ventral striatal gradient (i.e., post-treatment—baseline) and changes in cocaine craving (i.e., post-treatment—baseline). Of note, the data point presents residual data after regressing age, sex, and education years. VS ventral striatum, n.s. not significant, *p < 0.05.

Exploratory analysis further showed that following rTMS, the active group (mean connectivity strength = 0.170 ± 0.173 SD) exhibited higher FC between left VS and dlPFC than the sham group (mean connectivity strength = 0.083 ± 0.235 SD), after controlling gender, age and education years (p = 0.017, partial η² = 0.137) (Fig. S5).

Cognitive characterization of altered ventral striatal regions in CUD

Meta-analytical decoding confirmed that significant characteristics of CUD-related gradient alterations were significantly associated with reward and motivation functions, as indicated by correlations with meta-analytic terms such as ‘prediction’, ‘loss’, ‘gains’, and ‘sensitivity’ (see Fig. S6, size of the font scaled by correlation strength).

Discussion

This study is the first to employ a gradient-based connectivity analysis combined with rsfMRI to characterize ventromedial-to-dorsolateral striatal functional topography in CUD. Moving beyond traditional discrete parcellation approaches, which assume discrete subregional boundaries, we identified continuous spatial gradient maps reflecting overlapping functional organization. Individuals with CUD who exhibited heavier drug dependence and use showed more pronounced disruptions along the ventromedial-to-dorsolateral striatal axis, marked by increased gradient variance, elevated values in the NAc and ventromedial caudate, and reduced values in the dorsomedial caudate. Furthermore, these gradient abnormalities were associated with dysregulated frontostriatal circuitry in CUD. Critically, we provide novel evidence that 5-Hz rTMS over the left dlPFC modulates aberrant gradients in ventral striatal regions by addressing corresponding frontostriatal circuit dysregulation. This modulation was also correlated with reductions in cocaine craving, highlighting the potential of rTMS as a promising therapeutic intervention for CUD.

Our findings underscore the significant role of disrupted ventromedial-to-dorsolateral hierarchical striatal topography in CUD, aligning with prior studies linking dorsal-ventral striatal differentiation to impaired cognitive and reward processing [53, 54]. The expanded gradient mode and spatial variance in CUD are likely to indicate a heightened divergence between ventromedial and dorsolateral striatal signaling. This spatial disorganization parallels rsfMRI and task-based fMRI studies reporting hyperconnectivity in ventral striatal-anterior prefrontal/orbitofrontal corcuits [9, 55], and hypoconnectivity in dorsal striatal-dorsal prefrontal pathways [56, 57] among CUD patients. Moreover, the association between greater gradient variance and higher dependence severity across three cohorts highlights this metric’s potential as a biomarker for CUD diagnosis and progression. Regionally, gradient abnormalities were localized to ventral striatal regions (e.g., NAc and anterior caudate), key hubs for reward and incentive processing [51, 53], and dorsal striatal regions governing cognitive control [12, 51]. Cocaine consumption severity exacerbated alterations in these areas, suggesting dose-dependent neurobiological impairment. Complementary seed-based FC analyses revealed hypo-connectivity between VS and dlPFC, implicated in impaired reward learning and the extent of drug craving [58], as well as between the DS and dorsomedial prefrontal cortex (dmPFC), linked to inhibitory deficits [13, 59, 60]. These findings further emphasize dysregulation in networks subserving reward valuation and executive function in CUD [61–63].

Previous studies have demonstrated that dlPFC-targeted neuromodulation exhibits therapeutic potential for reducing cocaine use and craving [35, 38, 64–71]. Theoretical and imaging studies have further shown that rTMS of dlPFC induces alterations in striatal activity [39, 40, 72]. However, whether dlPFC-targeted stimulation affects striatal circuits in CUD and whether these potential neurobiological effects are associated with reductions in craving remain under investigated. Our findings demonstrated that left dlPFC-targeted 5-Hz rTMS normalizes aberrant striatal functional topography, subsequently altering craving in CUD populations. Specifically, it “corrects” the elevated gradient values observed in ventral striatal areas. This regional specificity is consistent with anatomical tracing studies in non-human primates showing that ventromedial and central striatal areas receive direct projections from the dlPFC, whereas dorsolateral striatal regions are innervated by premotor and motor cortices [15]. Earlier rTMS studies further support this dissociation, reporting that cocaine users retain intact dorsal striatal blood oxygen level-dependent (BOLD) responses to dlPFC stimulation, with no significant differences compared to controls [73]. Importantly, we found that these changes in the striatum were correlated with craving performance, with greater decreases in ventral striatal gradient values corresponding to larger reductions in current craving. This aligns with evidence implicating dopamine dysregulation in the VS as a driver of cocaine-seeking behavior [39, 74]. Prior work has demonstrated that dlPFC stimulation enhanced dopamine excretion in mesolimbic and mesostriatal pathways through its interconnection with the ventral tegmental area and VS [75–77], potentially addressing drug-induced dopaminergic dysfunction [78]. Exploratory analysis also revealed that rTMS strengthened FC between the VS and dlPFC. These results complement work by Angeles-Valdez et al. [37], who reported increased connectivity between the left dlPFC-vmPFC and vmPFC-right angular gyrus following rTMS. According to the competing neurobehavioral decision systems model [79], excessive activity in the limbic system (encompassing the VS and vmPFC [80]) in CUD may be counterbalanced by enhancing executive control network activity, which is pathologically diminished in this population. Our findings thus support the hypothesis that rTMS exerts therapeutic effects in CUD by directly modulating executive control network activity and indirectly dampening limbic system hyperactivity.

Notably, ventral striatal abnormalities were consistently replicated across all cohorts. Our meta-analysis further confirmed that these alterations localized to regions subserving reward processing and incentive-based decision-making, highlighting a disrupted reward valuation system - primarily mediated by ventral striatal dysfunction - as a core pathophysiological feature of CUD.

Despite promising findings, this study has several limitations. First, the sample was predominantly male, which hinders sex-specific analyses. Second, the cross-sectional design precludes the examination of striatal gradient alterations across different stages of CUD. Future longitudinal studies should track these changes over time to provide a more comprehensive understanding. Finally, we focused on striatal changes due to their central role in CUD. Future work could investigate gradient characteristics and rTMS effects in other regions (e.g., insula, dACC).

In conclusion, by applying a gradient-based framework, we delineated fine-grained disruptions in the ventromedial-to-dorsolateral striatal axis in CUD, offering a novel perspective on CUD-related neural reorganization. Our findings position striatal gradient variance as a potential diagnostic biomarker and underscore the effect of rTMS over left dlPFC to modulate ventral striatal dysfunction, advancing mechanistic insights into neuromodulation therapies.

Author contributions

RZ, and QH conceptualized and designed the work. RZ, and TZ preprocessed the data. RZ, and XX performed data analysis. RZ, XX, YH, YY and QH were responsible for the interpretation of the data. DD, TX, FZ, and YQ provided important suggestions during the formal analysis. RZ drafted the manuscript. DD and FZ provided suggestions during manuscript writing. YH, YY, and QH revised the paper. QH supervised the project.

Data availability

The datasets used in the discovery cohort and rTMS-treatment cohort are publicly available. The SUDMEX-CONN dataset is available from Zenodo (https://zenodo.org/record/5123331). The SUDMEX-TMS dataset is also available from Zenodo (https://zenodo.org/record/7126853). The data of replication cohort supporting the findings of this study are available from the corresponding author upon reasonable request.

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at 10.1038/s41386-025-02098-z.