Optimizing brain functional–structural architecture with cTBS to reduce relapse in alcohol use disorder: a randomized controlled trial

Abstract

Alcohol use disorder (AUD) is characterized by high relapse rates, necessitating novel interventions. Continuous theta-burst stimulation (cTBS) targeting the right dorsolateral prefrontal cortex (rDLPFC) has been proposed to reduce relapse risk by modulating prefrontal control networks implicated in craving and self-regulation, but its efficacy and mechanisms remain to be established. In a randomized, double-blind, sham-controlled trial at the First Special Hospital of Harbin, 50 AUD patients (aged 18–60 years) received active cTBS or sham stimulation over 2 weeks (10 treatment days; two applications/day). Alcohol Use Disorders Identification Test (AUDIT) scores were assessed before and after intervention, and relapse status was monitored during the 1-year follow-up. Resting-state fMRI data were analyzed using non-negative matrix factorization (NMF) and Tabular Prior-data Fitted Network (TabPFN) to identify relapse-associated neural patterns. Network control theory (NCT) was used to quantify the control energy required to regulate frontoparietal network (FPN) to subcortical network (SUB) transitions, constrained by structural connectomes. Transcriptomic analysis and TransBrain-based cross-species phenotype mapping were used to link control energy changes to gene expression and investigate the evolutionary conservation of the identified network-control targets. Compared with sham treatment, active cTBS was associated with a lower risk of relapse over one year (hazard ratio = 0.426, 95% CI [0.189, 0.964], p = 0.041). NMF revealed an abstinence-associated pattern in the prefrontal-subcortical regions, enhanced by cTBS (p = 0.042). NCT showed a reduction in FPN-to-SUB control energy (p = 0.003), mediating changes in AUDIT-assessed alcohol-use severity (95% CI [−0.302, −0.014]). Changes in the control energy correlated with gene expression profiles enriched for genes related to neuroplasticity (p < 0.001, FDR corrected) and mapped specifically to the murine infralimbic and prelimbic areas. rDLPFC-targeted cTBS may be associated with lower relapse risk through modulation of FPN-to-SUB regulation, a network-level effect linked to neuroplasticity-related gene expression and evolutionarily conserved prefrontal circuits. These findings support further evaluation of network-informed neuromodulation strategies for AUD.

Introduction

Alcohol use disorder (AUD) is a brain disease characterized by high prevalence, high relapse rates, and frequent psychiatric comorbidities [1, 2]. Approximately 1.4% of adults worldwide are affected by AUD [3], with relapse rates reaching 60% within 6 months and nearly 90% within one year of a withdrawal intervention [4, 5]. This high relapse rate highlights the urgent need for new interventions [6, 7]. From the perspective of the dual-process model of addiction, AUD reflects dysfunction across large-scale brain networks, particularly an imbalance between the top-down frontoparietal network (FPN), which governs cognitive control, and bottom-up systems related to stimulus salience and reward processing, including the ventral attention network (VAN) and the subcortical network (SUB) [8–10]. This specific network dysfunction provides a theoretical basis for targeted neuromodulation.

In this context, continuous theta-burst stimulation (cTBS)—a high-efficiency variant of repetitive transcranial magnetic stimulation (rTMS)—has gained widespread attention [11]. Unlike conventional rTMS, which requires weeks of daily sessions, cTBS needs only 40 seconds per session, offering significant advantages in duration and efficiency suitable for accelerated clinical protocols [12–14]. In addition, cTBS induces more sustained neuromodulatory effects and can elicit plasticity changes akin to long-term depression, demonstrating the potential to modulate aberrant cortical activity linked to addictive behaviors [11, 15, 16]. Although AUD is generally characterized by reduced prefrontal control, this hypoactivity is neither spatially nor functionally uniform [10]. Accumulating evidence indicates that the right dorsolateral prefrontal cortex (rDLPFC) is preferentially involved in craving-related attentional bias and exhibits aberrant involvement in alcohol-related information processing, with its functional connectivity and engagement patterns being predictive of relapse risk [17–19]. We therefore focused on the rDLPFC as a putative node of right-lateralized craving-related network dysfunction. Accordingly, we applied an inhibitory cTBS protocol targeting the rDLPFC, hypothesizing that downregulating its aberrant network involvement would attenuate craving-related circuitry and facilitate the rebalancing of the FPN.

Given the promising clinical applications of cTBS [20], an unanswered question remains regarding its application in AUD intervention: Are functional changes induced by stimulation of cortical targets transmitted to downstream networks through the brain’s physical architecture (e.g., white matter tracts)? The hierarchical organization of brain networks underlies behavior and cognition, with structural connectivity constraining functional dynamics [21, 22]. Cross-sectional studies of AUD have revealed network-level imbalances [23–25], which are often underpinned by widespread white-matter microstructural damage. Specifically, reduced integrity in key white matter tracts, such as the corpus callosum and superior longitudinal fasciculus, has been shown to impair the efficient transmission of top-down cognitive control signals, thereby increasing vulnerability to relapse [26, 27]. Clarifying how functional alterations propagate along this structurally compromised hierarchy is therefore essential for understanding AUD pathophysiology and optimizing neuromodulatory interventions.

To bridge this gap and capture the complexity of relapse mechanisms, we adopted a multiscale framework integrating functional states, control dynamics, molecular architecture, and cross-species phenotype mapping. First, we applied amplitude of low-frequency fluctuation (ALFF) combined with non-negative matrix factorization (NMF), together with a Tabular Prior-data Fitted Network (TabPFN), to identify spatially distinct, data-driven neural oscillation patterns associated with relapse risk and behavioral inflexibility in AUD [28, 29]. While ALFF quantifies the local intensity of spontaneous neural activity—consistent with prior addiction studies suggesting that aberrant low-frequency amplitude co-occurs with subjective craving-related measures and symptom burden [30–32]—it does not explain how the brain transitions between these states. This is where network control theory (NCT) becomes critical [33]. Clinically, high control energy reflects a ‘less cost-effective’ pathological state [34] and is associated with greater cognitive impairment [35]. NCT allows us to quantify this ‘control energy’—the theoretical effort required to guide the brain out of these states based on its white matter architecture [36–38]. Therefore, linking ALFF (state intensity) with NCT (transition cost) provides a mechanistic explanation for the clinical difficulty in regulating craving. Finally, given that macroscale functional organization is fundamentally constrained by molecular architecture [39–41], and that relapse vulnerability is biologically underpinned by specific neurocircuitry adaptations [42], we further integrated transcriptomic analysis using the Allen Human Brain Atlas (AHBA) and cross-species phenotype mapping via the TransBrain framework [43]. This allowed us to interpret our network-level findings within the biological context of relapse, explore their associations with underlying molecular mechanisms [44], and assess the evolutionary conservation of network-control targets [45, 46].

On the basis of this multiscale framework, we posit that cTBS acts by enhancing the driver capability of the FPN, effectively ‘flattening’ the energy landscape and reducing the energetic barriers for state transitions. Accordingly, we propose the following hypotheses: 1) cTBS of the rDLPFC specifically modulates neural oscillation patterns, yielding changes in brain dynamics that are associated with relapse risk. 2) Beyond local modulation, cTBS reduces the control energy required for the FPN to regulate downstream systems, thereby potentially improving global network control efficiency. 3) Such modulation of network control dynamics is associated with individual-level changes in clinical severity measures (e.g., AUDIT scores) and prospective relapse-related outcomes. 4) Individual differences in these control effects are shaped by underlying spatial gene expression architecture, and the identified human network-control targets map onto homologous rodent circuits implicated in top-down control. In this study, we aimed to elucidate the network-level mechanisms through which cTBS influences relapse in patients with AUD and to provide a mechanistic basis for the development of individualized, precision neuromodulation strategies.

Methods and materials

Study design and participants

This single-center, randomized, double-blind, sham-controlled clinical trial was conducted at the First Special Hospital of Harbin. The intervention was delivered on weekdays (Monday–Friday) over two weeks (10 treatment days). Each treatment day comprised two cTBS applications (morning and afternoon, separated by a 6-hour interval), resulting in 20 total cTBS applications. Participants were randomly assigned to receive either active cTBS or sham stimulation. Randomization and blinding procedures (computer-generated allocation with concealment) are detailed in the Supplementary Methods.

We recruited patients with AUD who were aged between 18 and 60 years, had completed junior high school or higher, and met the diagnostic criteria for a substance use disorder (SUD) in the Diagnostic and Statistical Manual of Mental Disorders, 5th Edition (DSM-5). The exclusion criteria included a history of epilepsy, intellectual disability, or other serious mental illnesses; any contraindications for magnetic resonance imaging (MRI) or rTMS; or being in the acute phase of withdrawal with severe symptoms.

The study protocol was approved by the Research Ethics Committee of the First Special Hospital of Harbin in accordance with the Declaration of Helsinki. The trial was monitored by a local safety monitoring board and was registered with the Chinese Clinical Trial Registry (registration number: ChiCTR2200065453). All participants provided voluntary, written informed consent before the study began.

Continuous TBS

cTBS was delivered with a figure-eight coil (Model B9076) on a CCY-I stimulator (Wuhan Yiruide Medical Co., Ltd., Wuhan, China). The rDLPFC was targeted at F4 using the Yiruide TMS Positioning Cap according to the International 10–20 system [47]. Each session comprised 50 Hz triplets repeated at 5 Hz (200 ms interburst interval) for 40 s, totaling 600 pulses [11]. The stimulation intensity was set to 100% of the resting motor threshold (RMT), which was determined from the dominant-hand first dorsal interosseous (FDI) muscle using a standardized procedure [48]. Detailed procedures for RMT determination, coil positioning, and sham stimulation are provided in the Supplementary Methods.

Clinical assessments

The Alcohol Use Disorders Identification Test (AUDIT) and alcohol craving (assessed via a visual cue-reactivity paradigm using a 0–100 visual analog scale [VAS]) were administered at baseline (pre-intervention) and immediately post-intervention (after the final day of active or sham stimulation). Similarly, additional questionnaires characterizing clinical features, including the Perceived Stress Scale-10 (PSS-10), Barratt Impulsiveness Scale-11 (BIS-11), Clinical Institute Withdrawal Assessment for Alcohol-Revised (CIWA-Ar), Pittsburgh Sleep Quality Index (PSQI), and the Chinese version of the Habit, Reward, and Fear Scale (HRFS-C), were assessed at both time points. Trained staff conducted telephone follow-ups at 1, 3, 6, 9, and 12 months post-intervention to verify identity, collect drinking history, and rate relapse severity on a standardized 6-level scale (0–5). To ensure accuracy, reports were cross-checked with a family member or designated contact. The details of the clinical assessment procedures are provided in the Supplementary Methods.

MRI data acquisition

All MRI data were acquired on a 3.0T GE DISCOVERY MR750 scanner (GE Healthcare, Milwaukee, WI, USA) using a 12-channel head coil. High-resolution T1-weighted anatomical images, diffusion tensor imaging (DTI) data and rs-fMRI images were acquired for each participant at both baseline and post-intervention timepoints. Each rs-fMRI scan lasted for 8 min, resulting in the collection of 240 volumes. Detailed acquisition parameters for the T1-weighted, DTI and rs-fMRI sequences are provided in the Supplementary Methods.

fMRI data processing and ALFF calculation

Data were preprocessed using fMRIPrep (v23.1.4) [49] and subsequently post-processed with XCP-D (v0.9.0) [50] to generate ALFF maps. The key post-processing steps included a 36-parameter nuisance regression, band-pass filtering (0.01–0.08 Hz), and spatial smoothing (6 mm FWHM kernel). Finally, we extracted the mean ALFF signal from 246 regions of the Brainnetome Atlas (BNA246) [51] and concatenated each participant’s pre- and post-intervention vectors to form the input matrix for NMF. A comprehensive description of this workflow is available in the Supplementary Methods.

Non-negative matrix factorization

To identify global patterns of neural activity, we applied NMF to the ALFF data. ALFF is strictly non-negative by definition, eliminating the need for thresholding prior to NMF. This algorithm decomposed the regional ALFF data, comprising all pre- and post-intervention scans, into a set of spatial components (H matrix) and a corresponding matrix of expression weights for each scan (W matrix). The optimal number of components, k, was determined by systematically evaluating values from 1 to 10. For each k, we balanced model stability (consistency across 100 random initializations) against reconstruction error (data-fitting accuracy). The value of k = 5 was selected because it maximized the stability-to-error ratio. This optimal five-component solution was then carried forward to the subsequent relapse prediction analysis. The detailed methodology for NMF, including matrix construction and the component selection process, is described in the Supplementary Methods.

TabPFN-SHAP-based prediction model

To identify the neural oscillation patterns predicting relapse outcomes, we constructed a classification model using the TabPFN framework [29, 52]. The model utilized the spatial component matrix obtained from the NMF analysis as input features, and the classification label was defined as relapse status at follow-up assessments (yes vs. no). Its performance was evaluated using receiver operating characteristic (ROC) analysis. Given the limited sample size, we leveraged TabPFN’s prior-fitted architecture (which requires no hyperparameter tuning) to structurally mitigate overfitting, and implemented a Stratified 5-Fold Cross-Validation framework to rigorously estimate performance. To interpret the model’s predictions and determine the importance of each neural component in driving the outcome, we subsequently applied SHapley Additive exPlanations (SHAP) analysis [53]. We defined the top-ranked component according to the SHAP importance ranking as the global relapse-predictive oscillation pattern. A detailed description of the TabPFN architecture and the SHAP analysis implementation is available in the Supplementary Methods.

Structural connectome construction

Individual structural connectomes were constructed using standardized pipelines implemented in QSIPrep (v0.24.1) and QSIRecon (v1.0.1) [54]. The process involved comprehensive preprocessing of each participant’s T1-weighted and diffusion MRI data.

Using the BNA246 Atlas, whole-brain deterministic tractography was performed in MRtrix3 to estimate structural connectivity, defined as the streamline count between each pair of the 246 atlas regions. This procedure yielded a 246 × 246 weighted structural connectivity matrix for each participant at each time point (baseline and post-intervention), which was used in all subsequent NCT analyses. A detailed, step-by-step description of the image preprocessing, reconstruction, and tractography parameters is provided in the Supplementary Methods.

Network control theory analysis

After defining global changes in brain function, we leveraged the NCT to quantify the theoretical control energy required for brain state transitions. This approach models brain dynamics as a linear system whose trajectories are constrained by each participant’s structural connectome. To ensure comparability across participants, brain states were defined using the standardized group-level spatial pattern (H matrix) of the NMF component of interest. Specifically, we simulated state transitions originating from the FPN. We confirmed empirically that the FPN represents the primary component within the rDLPFC stimulation target (see Supplementary Methods for ROI quantification), allowing us to model network-level effects anchored in this FPN-dominant site. We calculated the control energy required to transition from the FPN to eight specific target states. These comprised the subcortical system (SUB) and the seven networks defined by the Yeo parcellation [55]: Visual (VIS), Somatomotor (SMN), Dorsal Attention (DAN), Ventral Attention (VAN), Limbic (LIM), Frontoparietal (FPN), and Default Mode (DMN).

This analysis yielded a minimal control energy value for each transition, quantifying the theoretical effort for the FPN to regulate each of the other functional systems, with non-target regions set to zero to selectively isolate specific circuit dynamics. The detailed mathematical framework of the NCT model, including state-space equations and the specific definitions of the initial and target state vectors, is provided in the Supplementary Methods.

Transcriptional analysis and cross-species phenotype mapping

To explore the molecular underpinnings of the observed changes in network control, we investigated the spatial relationship between our neuroimaging findings and brain-wide gene expression patterns from the AHBA [44, 56]. Following established imaging-transcriptomic protocols [39, 57, 58], we employed partial least squares regression (PLSR) to identify a transcriptomic signature that was spatially correlated (across 246 brain regions) with a group-level t-map representing the group difference (active vs. sham) in the cTBS-related changes in control energy. To interpret the biological significance of this relationship, the genes most strongly associated with the primary PLS component were subjected to gene enrichment analysis using the Metascape platform [57]. The detailed procedures for deriving the control energy t-map, preprocessing the AHBA data, and implementing the PLS model are available in the Supplementary Methods.

Furthermore, to investigate the evolutionary conservation of these network-control targets, we performed cross-species phenotype mapping using the TransBrain framework [43]. We aligned the group-level t-map—capturing the spatial topography of the cTBS-related reductions in control energy—with the Allen Mouse Brain Common Coordinate Framework (CCFv3) via manifold embedding algorithms. This approach enabled the spatial projection of the human network-control targets onto murine brain architecture to systematically explore potential anatomical homology. The specific implementation procedures for this cross-species phenotype mapping analysis are detailed in the Supplementary Methods.

Statistical analysis

Within the framework of the NCT, the control energy required to transition from the FPN to each of the seven Yeo functional networks and the SUB was calculated using individual-level structural connectivity matrices. The effect of the intervention was assessed using repeated-measures analysis of variance (ANOVA), with time (pre vs. post) as the within-subject factor and group (cTBS vs. sham) as the between-subject factor. Significant interactions were decomposed using simple effects analyses to characterize the direction of group differences over time. The p-values for post hoc simple effects were adjusted using the same multiple-comparison correction applied to the corresponding test family. When multiple pairwise comparisons were required (e.g., NCT analyses involving multiple transitions across functional networks), post hoc pairwise tests were adjusted using Bonferroni correction to control the family-wise error rate. A power analysis (detailed in the Supplementary Material) confirmed that the final analytical sample (N = 50) exceeded the a priori minimum (N = 34) for the primary repeated-measures ANOVA. We evaluated the risk of relapse during follow-up using a Cox proportional hazards model. To assess clinical relevance, we examined the association between pre–post changes in control energy and changes in AUDIT score using multiple linear regression analyses. In these models, the change in AUDIT score was entered as the dependent variable, the change in control energy as the independent variable, and age and baseline AUDIT score were included as covariates to control for potential confounding effects and regression-to-the-mean. The change score (Δ) was defined as the post-intervention value minus the pre-intervention value. Additionally, duration of abstinence (days) was included as a covariate in a sensitivity analysis. Furthermore, to examine the mediating role of this neural measure, a mediation analysis was performed to test whether the change in control energy mediated the association between intervention assignment (cTBS vs. sham) and the change in AUDIT score. The mediation analysis was conducted using a nonparametric bootstrapping approach to estimate the 95% confidence interval of the indirect effect. Effect sizes were reported as partial eta-squared (${\eta }_p^2$) for ANOVA and standardized path coefficients for mediation analyses. At the molecular level, the control energy difference map was spatially matched to regional gene expression data from the AHBA, and PLS was performed to identify genes whose spatial expression patterns covaried with the control energy changes. The first PLS component (PLS1), which explained the greatest variance, was selected for downstream analysis, with statistical significance corrected using the false discovery rate (FDR). Finally, to investigate the evolutionary conservation of the identified network-control targets, we utilized the TransBrain framework to map their spatial topography onto the murine brain architecture via manifold embedding. The overall study design and analysis pipeline are shown in Fig. 1.

Fig. 1. Study design and analysis pipeline.

(A) Schematic of the study design. This study was a double-blind, randomized controlled trial (DB-RCT) of patients with alcohol use disorder (AUD). At baseline, all participants underwent clinical assessments and functional magnetic resonance imaging (fMRI) scans. Participants were subsequently randomly assigned to one of two groups: (1) the active group, which received continuous theta-burst stimulation (cTBS) targeting the right dorsolateral prefrontal cortex (rDLPFC); or (2) the sham group. The intervention for both groups lasted for 2 weeks and comprised 10 treatment days (weekdays over 2 weeks) with two cTBS applications per day (20 total applications). Following the intervention, all participants completed the same assessments as at baseline. (B) Diagram of the data analysis pipeline. This flowchart illustrates the complete analysis workflow for the multimodal data. First, resting-state fMRI (rs-fMRI) data were processed to calculate the amplitude of low-frequency fluctuation (ALFF), and the ALFF values were extracted for each brain region using the Brainnetome Atlas (BNA246). Next, non-negative matrix factorization (NMF) was used to decompose the ALFF region-of-interest (ROI) matrix from all patients into a set of basic spatial components and corresponding patient-specific weights. In the subsequent multivariate analysis, first, a Cox proportional hazards model was used to analyze the risk of relapse between the active and sham groups during follow-up, and then the above NMF-derived components and weights were used as features and input into a Tabular Prior-data Fitted Network (TabPFN) classification model to perform the relapse classification prediction task. On the basis of the first spatial component from the NMF analysis, brain network states for network control analysis were defined; the initial state was the activity pattern of the frontoparietal network (FPN), and the target states were the seven Yeo functional networks and the subcortical network. The network control theory (NCT) framework was then applied to calculate the control energy required to steer brain activity from the initial state to each target state on the basis of the individual’s structural connectome. Finally, to explore the molecular underpinnings of these energy changes, a control energy t-map was generated and linked with gene expression data from the Allen Human Brain Atlas using partial least squares (PLS) regression, followed by a gene enrichment analysis of the identified key genes. Additionally, cross-species phenotype mapping was performed via the TransBrain framework to assess the evolutionary conservation of the identified network-control targets.

Results

Demographic and clinical behavior data

Participant flow and exclusions are summarized in the Supplementary Materials (Figure S1). The final analysis included 29 participants in the active group and 21 in the sham group. Statistical analysis revealed no systematic differences between the two groups in terms of demographics, clinical history, or baseline clinical behavior scale scores before the intervention (Table 1 and Table S4).

Table 1.

Demographic characteristics of participants.

Characteristics	Active Group n = 29	Sham Group n = 21	p Value
Gender, n
Male	28	18	0.30a
Female	1	3
Age, Years, Mean ± SD	46.00 ± 9.69	49.62 ± 9.28	0.19b
Education years, Mean ± SD	9.93 ± 1.71	10.05 ± 1.72	0.81b
Drink Years, Mean ± SD	26.02 ± 9.92	24.67 ± 10.90	0.65b
Duration of abstinence (Days), Mean ± SD	6.97 ± 6.90	8.75 ± 6.59	0.37b
AUDIT, Mean ± SD	31.21 ± 6.93	27.10 ± 11.15	0.15b
CIWA-Ar, Mean ± SD	25.25 ± 11.57	22.48 ± 10.80	0.40b
PSS-10, Mean ± SD	29.79 ± 9.54	28.29 ± 10.19	0.60b
BIS-11, Mean ± SD	52.04 ± 12.53	54.37 ± 11.51	0.51b
PSQI, Mean ± SD	10.10 ± 3.88	9.19 ± 3.75	0.41b
Habit, Mean ± SD	4.83 ± 1.74	4.28 ± 2.08	0.31b
Reward, Mean ± SD	5.32 ± 1.05	5.01 ± 1.28	0.35b
Fear, Mean ± SD	3.90 ± 1.12	3.76 ± 1.06	0.65b
Craving, Mean ± SD	46.21 ± 24.56	53.81 ± 29.41	0.34b

The values are presented as n or mean ± SD. There were no significant differences between the active and sham groups for any demographic or clinical characteristics at baseline.

AUDIT, Alcohol Use Disorders Identification Test, BIS-11, Barratt Impulsiveness Scale-11, CIWA-Ar, Clinical Institute Withdrawal Assessment for Alcohol-Revised; PSQI, Pittsburgh Sleep Quality Index, HRFS-C, Chinese version of the Habit, Reward, and Fear Scale, PSS-10, Perceived Stress Scale-10, VAS, Visual Analog Scale.

^ap-value from the Fisher’s exact test.

^bp-value from independent samples t-test.

With respect to the intervention effect, repeated-measures ANOVA revealed a significant Time × Group interaction for the PSS-10 (F(1, 48) = 4.656, p = 0.036, ${\eta }_p^2$ = 0.088). To characterize the direction of this interaction, simple effects analysis indicated that PSS-10 scores significantly decreased in the active cTBS group (p = 0.017), whereas the sham group showed no significant change (p = 0.475), suggesting a post-intervention reduction in perceived stress in the active cTBS group. In contrast, no significant interaction effects were observed on clinical scales such as the AUDIT (F(1, 48) = 1.871, p = 0.178, ${\eta }_p^2$ = 0.038) or on craving scores (F(1, 48) = 3.029, p = 0.088, ${\eta }_p^2$ = 0.059; see Supplementary Materials, Table S1). In the active cTBS group, mean AUDIT scores changed minimally from 31.21 to 30.90, indicating limited short-term change in AUDIT-assessed alcohol-use severity. For the relapse risk after intervention, the time to relapse was modeled using a Cox proportional hazards model. The risk of relapse within one year was significantly lower in the active cTBS group than in the sham group (HR = 0.426, 95% CI [0.189, 0.964], p = 0.041; Fig. 2A). Consistent with this longitudinal outcome, cross-sectional analyses revealed numerically lower relapse rates in the active cTBS group across all time points (Supplementary Table S2), with significant group differences observed only at 6 months (p = 0.044) and 9 months (p = 0.042) under the strict abstinence definition (Supplementary Results and Table S3).

Fig. 2. NMF-identified brain network activity patterns predict relapse and are modulated by cTBS.

A Adjusted relapse-free probability curves from a Cox proportional hazards model, showing a significantly lower risk of relapse (p = 0.041) in the active cTBS group than in the sham group, with the number at risk shown below. The number at risk decreases primarily due to relapse events and administrative censoring; attrition was minimal (n = 2) and occurred only in the active group. B The top five spatial components (C1–C5) identified from amplitude of low-frequency fluctuation (ALFF) data using non-negative matrix factorization (NMF). The heatmap at the top displays the weights of major brain networks and anatomical regions within each component. The brain maps below provide a spatial visualization of these five components, with warmer colors indicating higher regional weights. C Receiver operating characteristic (ROC) curve for the TabPFN model, in which the NMF components were used to predict relapse. The model achieved an area under the curve (AUC) of 0.906 and a prediction accuracy of 0.880, indicating good classification performance. D SHAP (SHapley Additive exPlanations) analysis of feature importance. The bar plot shows the mean absolute SHAP value (global importance), indicating that Component 1 was the most important feature. The beeswarm plot provides a detailed illustration of each component’s impact on the model’s output for individual subjects, where each point represents a subject; its position on the x-axis indicates the magnitude and direction of its impact, while its color represents the component’s expression weight. E Comparison of the changes in the expression weight of the most important component (Component 1) before (Pre) and after (Post) the intervention for both the active cTBS group and the sham group. The active cTBS intervention significantly increased the expression weight of Component 1 (p = 0.042), whereas no significant change was observed in the sham group.

cTBS modulates a specific abstinence-associated neural oscillation pattern

To identify key resting-state neural oscillation patterns in a data-driven manner, we applied NMF to the ALFF data and extracted five typical components (C1–C5), which exhibited high bilateral symmetry and clear spatial heterogeneity. For example, Component 1 (C1) was prominently expressed in the SUB and ventral attention network (VAN), including the medial frontal gyrus (MFG), cingulate gyrus (CG), basal ganglia (BG), and thalamus (Thal). Component 2 (C2) was expressed primarily in the visual network (VIS), especially in the medial and lateral occipital cortices (MVocC, LocC) (Fig. 2B). To assess the clinical relevance of these components, we constructed a relapse prediction model using C1–C5. The model demonstrated good performance (AUC = 0.906, accuracy = 88%) (Fig. 2C). Crucially, to verify the robustness of this finding and rule out overfitting, we further evaluated the model using Stratified 5-Fold Cross-Validation, which confirmed a stable predictive performance (mean AUC = 0.878, mean accuracy = 84.4%) (Figure S4). A feature importance analysis revealed C1 as the dominant predictor (mean|effect | = 0.099). SHAP analysis further confirmed that higher values of C1 contributed negatively to relapse risk prediction (associated with reduced risk) (Fig. 2D). Given the predictive value of C1, we examined whether it was modulated by cTBS. A significant time-by-group interaction was observed for C1 expression weights (F(1, 48) = 4.186, p = 0.046, ${\eta }_p^2$ = 0.080) (Fig. 2E), indicating a differential pre–post change between the active and sham groups. This interaction was further assessed in outlier sensitivity analyses addressing the potential influence of the C1 outlier (see Supplementary Results). To characterize the direction of this interaction, we conducted planned simple-effects comparisons. C1 increased following active cTBS (F(1, 48) = 4.359, p = 0.042, ${\eta }_p^2$ = 0.083), whereas no change was observed in the sham group (F(1, 48) = 0.828, p = 0.368, ${\eta }_p^2$ = 0.017). Thus, active cTBS induced a significant modulation of a neural oscillation pattern that is strongly associated with long-term abstinence.

cTBS modulates network control energy associated with changes in AUDIT-assessed alcohol-use severity

To elucidate the neural mechanisms underlying cTBS modulation, we employed NCT to assess changes in the control energy required for transitions from the FPN to major functional networks (Yeo’s 7 networks and the SUB). Among all the transition paths, cTBS selectively influenced the FPN-to-VAN and FPN-to-SUB transitions. For the FPN-to-VAN transition, a significant time-by-group interaction was observed (F(1, 48) = 11.060, p = 0.002, ${\eta }_p^2$ = 0.187, Bonferroni corrected). Simple effect analysis revealed a significant decrease in control energy in the active cTBS group (F(1, 48) = 8.974, p = 0.004, ${\eta }_p^2$ = 0.158, Bonferroni corrected), with no significant change in the sham group (F(1, 48) = 3.303, p = 0.075, ${\eta }_p^2$ = 0.064) (Fig. 3A). Similarly, a significant interaction was found for the FPN-to-SUB transition (F(1, 48) = 14.007, p < 0.001, ${\eta }_p^2$ = 0.226, Bonferroni corrected). The control energy significantly decreased after the intervention in the active group (F(1, 48) = 9.516, p = 0.003, ${\eta }_p^2$ = 0.165, Bonferroni corrected) but increased in the sham group (F(1, 48) = 5.241, p = 0.027, ${\eta }_p^2$ = 0.098) (Fig. 3B). To assess clinical relevance, we performed multiple linear regression analysis, which indicated that the reduction in the FPN-to-SUB control energy (ΔFPN-to-SUB) was significantly associated with AUDIT score change (ΔAUDIT) after adjusting for age and baseline AUDIT score (β = 0.824, t = 2.376, p = 0.025; Fig. 3C). This association was robust to influence diagnostics (see Supplementary Results). Mediation analysis further indicated that ΔFPN-to-SUB mediated the association between intervention assignment and changes in AUDIT-assessed alcohol-use severity (indirect effect = -0.168, 95% CI [−0.302, −0.014]) (Fig. 3D). Crucially, beyond the targeted FPN, no significant changes were observed for transitions to any of the other networks after correction for multiple comparisons, confirming the specificity of the modulation (Figure S5).

Fig. 3. cTBS-related reductions in network control energy are associated with changes in AUDIT-assessed alcohol-use severity.

A–B For the two state transition pathways from the frontoparietal network (FPN) to the ventral attention network (VAN) (A) and to the subcortical network (SUB) (B), the active cTBS intervention significantly decreased the required control energy. C Scatter plot illustrating the positive association between the change in FPN-to-SUB control energy (ΔFPN-to-SUB control energy, post to pre) and AUDIT score change (ΔAUDIT, post to pre) (r = 0.420, p = 0.023). Multiple linear regression analysis indicated that this association remained significant (β = 0.824, t = 2.376, p = 0.025) after adjusting for age and baseline AUDIT score. D Mediation analysis indicated that the change in FPN-to-SUB control energy mediated the association between intervention assignment (cTBS vs. sham) and changes in AUDIT-assessed alcohol-use severity (indirect effect = −0.168, 95% CI [−0.302, −0.014]).

Molecular basis and cross-species phenotype mapping of the FPN-to-SUB network control energy changes

To identify the molecular basis of the changes in the FPN-to-SUB network control energy, we obtained a whole-brain gene expression matrix from the AHBA and applied a PLS to link gene expression patterns with the group-difference t-map of the cTBS-related control energy changes (Fig. 4B, top). PLS1 explained 11.41% of the variance in group differences in control energy, which was significantly greater than chance (spin test, p_spin = 0.034), and exhibited an anterior–posterior spatial gradient consistent with previous findings (Fig. 4A). Moreover, the PLS1 scores were significantly positively correlated with the group-difference t-map of cTBS-related control energy changes (r = 0.277, p_spin = 1.55 × 10⁻⁵) (Fig. 4C).

Fig. 4. Transcriptomic correlates and cross-species phenotype mapping of the change in the FPN-to-SUB network control energy.

A The percentage of variance explained by each component in the partial least squares (PLS) regression model. The first PLS component (PLS1) explained 11.41% of the total variance. The inset brain map displays the spatial distribution of PLS1 scores, representing the gene expression pattern most strongly associated with the FPN-to-SUB control energy change t-map. B The top panel shows the input variable for the PLS regression: the t-map represents the group difference in the change in FPN-to-SUB control energy. Here, negative t-values indicate a greater reduction in energy in the active group (reflecting a stronger neuromodulatory effect), while positive t-values indicate a smaller reduction or relative resistance to modulation. The bottom panel illustrates cross-species phenotype mapping via the TransBrain framework. The sagittal view (ML = 0.2 mm) demonstrates that the identified human network-control targets map predominantly to the murine infralimbic (ILA) and prelimbic (PL) areas, with the bar plot ranking the top 10 regions by mapping weight. C The scatter plot demonstrates a significant positive spatial correlation (r = 0.277, p_spin = 1.55 × 10⁻⁵) between the PLS1 scores and the t values for control energy changes, where the gray band represents the 95% confidence interval. D List of the top-ranked genes contributing most significantly to the PLS1 component, ranked by their Z scores. The list is separated into genes with positive (red) and negative (blue) contributions. E–F Network visualizations of the gene enrichment analysis results for genes positively (E) and negatively (F) associated with the PLS1 component. The positively associated genes were enriched primarily in pathways related to nervous system development and synaptic function, such as “regulation of synapse organization” and “neuron projection development”. The negatively associated genes were enriched primarily in more fundamental cellular processes, such as “ribosome biogenesis” and “DNA repair”.

To characterize the biological function of this transcriptomic signature, we ranked all the genes by their standardized PLS1 weight (Z score). Gene enrichment analysis was performed on the top 1000 genes with the strongest positive association (PLS1⁺) and the bottom 1000 genes with the strongest negative association (PLS1⁻) (Fig. 4D). Positively associated genes (PLS1⁺) were enriched in pathways related to nervous system development and function, including “regulation of synapse organization,” “neuroactive ligand–receptor interaction,” and “behavior” (Fig. 4E; all p < 0.001, FDR corrected). In contrast, negatively associated genes (PLS1⁻) were enriched in fundamental cellular processes, such as “ribosome biogenesis,” “DNA repair,” and “RNA splicing” (Fig. 4F; all p < 0.001, FDR corrected).

Finally, to assess the biological translatability of this effect, cross-species phenotype mapping revealed that the spatial phenotype of cTBS-related reductions in control energy was not uniformly distributed but was highly concentrated in the murine medial prefrontal cortex (Fig. 4B, bottom). Among all 68 mouse brain regions, the infralimbic (ILA) and prelimbic (PL) areas exhibited the highest mapping weights (ILA = 1.018, PL = 0.997), which may constitute the evolutionary homologs of the identified human network-control targets.

Discussion

This study provides preliminary evidence that rDLPFC-targeted cTBS may be associated with lower long-term relapse risk and reduced perceived stress (PSS-10), although it did not yield clinically meaningful reductions in AUDIT-assessed alcohol-use severity or significantly greater reductions in craving compared with sham. To elucidate its underlying mechanism, this study employed a multiscale, multimodal analytical framework. We applied NCT constrained by individual structural connectivity to examine whether cTBS-related effects were accompanied by changes in brain control dynamics. Specifically, cTBS significantly reduced the ‘control energy’ required for the FPN—the core of cognitive control—to regulate craving-related SUB. This enhancement in network control efficiency mediated the association between intervention assignment and changes in AUDIT-assessed alcohol-use severity. Further molecular analysis and cross-species phenotype mapping indicated that this macroscopic optimization of network efficiency is associated with a gene expression profile enriched in neuroplasticity pathways, and specifically aligns with homologous circuits (ILA/PL) implicated in top-down control in the rodent brain. These findings offer insights into how cTBS more efficiently leverages the brain’s evolutionarily conserved structure to improve its function.

Modulatory effects of cTBS on an abstinence-associated neural oscillation pattern

A central challenge in AUD neuromodulation is linking a macroscopic target like the rDLPFC with core pathological circuits. To address this, the present study applied NMF and TabPFN-based classification to ALFF maps, aiming to identify key patterns of spontaneous neural oscillations. C1 emerged as the strongest predictor of relapse risk and exhibited a spatial distribution consistent with addiction-related circuits, including the MFG, CG, BG, and Thal—regions central to reward processing, habitual behavior, and emotional regulation [42, 59, 60]. This data-driven component aligns with theoretical models, highlighting its relevance to AUD pathophysiology. A key finding is that inhibitory cTBS to the rDLPFC significantly increased the expression of the subcortical C1 pattern, reflecting cTBS-related reorganization of the ALFF within this circuit. This cTBS-related neural reorganization was accompanied by a prospective time-to-relapse finding, with the active cTBS group showing a lower risk of relapse over one year. This finding suggests the potential clinical relevance of cTBS, but should be interpreted cautiously given the limited change in AUDIT-assessed alcohol-use severity and the absence of significant pointwise group differences in relapse-related outcomes at the 12-month follow-up. The cortical inhibition-induced enhancement of deep-brain oscillatory patterns may reflect a complex network-level mechanism. By transiently disrupting maladaptive cortical control, cTBS may allow SUB to restore more adaptive low-frequency dynamics. Crucially, our comparative analyses indicated that conventional univariate metrics derived from raw ALFF lacked sensitivity to detect intervention-related effects and did not show prognostic value for relapse outcomes. In contrast, the NMF-derived Component 1 disentangles a relapse-relevant covariance pattern from nonspecific inter-individual variance. SHAP analysis supported a protective role of C1: higher expression was linked to lower relapse risk (negative SHAP values). Consequently, the cTBS-related increase in the expression of this abstinence-associated pattern is consistent with modulation of neural dynamics that may support abstinence maintenance. Accordingly, NMF serves as an important data-driven filtering step, enabling the subsequent NCT framework to model the energy landscape of a clinically meaningful pathological network state, rather than that of undifferentiated background activity.

Optimization of network control dynamics as a candidate mechanism of cTBS effects

On the basis of the dual-process model of addiction—which posits an imbalance between weakened top-down control (represented here by the FPN) and hyperactive bottom-up drivers (the SUB) [8–10]—we demonstrated that cTBS not only modulated global neural oscillatory activity in the brain but also altered the local transmission effects of structural connections. The cTBS intervention significantly modulated state transitions from the FPN to both the VAN and the SUB. Although NMF identified these networks as co-active elements within the same spatial pattern, modeling the directional trajectory allows us to probe the structural capacity of the FPN—the direct recipient of the cTBS perturbation—to propagate control signals to downstream targets (specifically the VAN and SUB). The improved FPN-to-VAN efficiency likely reflects enhanced top-down attentional modulation. Existing studies have shown that stimulation of key nodes within the FPN by TMS can causally modulate activation levels in the VAN [61, 62]. Crucially, cTBS significantly reduced the control energy of the FPN-to-SUB pathway, unlike in the sham group. This finding indicates optimized neural efficiency in the top-down control of subcortical circuits (specifically the SUB), reflecting a reduced energetic cost to modulate these addiction-related hubs. In other words, the intervention effectively made it ‘easier’ for prefrontal networks to exert top-down modulation, stabilizing the otherwise volatile interaction with impulsive subcortical drivers [34, 63]. Such efficiency may stem from improved FPN–SUB integration or reduced neural transmission noise. Studies have shown that rTMS or cTBS applied to prefrontal regions (e.g., the DLPFC/mPFC) can significantly reduce craving in individuals with SUD, a mechanism associated with the modulation of functional connectivity in the prefrontal-subcortical reward circuit (e.g., the ventral striatum and amygdala) [64]. Mediation analysis indicated that improved FPN-to-SUB control efficiency mediated the association between the cTBS intervention and AUDIT score changes, suggesting that cTBS may modulate the dynamic efficiency of key control circuits in a way that is associated with individual-level changes in AUDIT-assessed alcohol-use severity. This interpretation aligns with emerging theoretical frameworks suggesting that high control energy reflects a ‘cost-ineffective’ pathological state [34], often associated with cognitive impairment [35]. In this context, the elevated energy costs in the sham group likely signify that the addicted brain is trapped in rigid, maladaptive dynamics. By lowering this energy barrier, cTBS effectively ‘flattens’ the energy landscape [63], facilitating the escape from these rigid constraints. Crucially, such energy reduction has been linked to increased neural signal diversity (entropy) rather than functional loss [65], confirming that our findings reflect the restoration of a flexible and metabolically efficient control system. Finally, this optimization appeared specific to the stimulation target and addiction-related circuits, as no surviving effects were observed for transitions to any of the other networks.

Transcriptional profiling and cross-species mapping of changes in network control

AUD arises from complex pathological processes across molecular, cellular, and systems levels, culminating in deficits in cognitive control [66, 67]. By integrating network control dynamics with transcriptomic profiles, we found that stimulation-induced changes in FPN-to-SUB control energy were significantly associated with cortical gene expression patterns. The PLS1⁺ genes, which are enriched in synaptic signaling and neuroplasticity pathways (e.g., synapse organization and neuroactive ligand–receptor interaction), were preferentially expressed in anterior regions with weaker intervention responses, potentially reflecting a spatial vulnerability to synaptic remodeling processes that may be preferentially engaged in the context of chronic alcohol exposure [68, 69]. Conversely, PLS1⁻ genes, enriched in ribosome biogenesis and RNA splicing, were expressed in regions with greater post-intervention control energy reductions. These biological processes are essential for cellular homeostasis and may provide a molecular baseline supporting adaptive network reorganization, given that robust ribosomal function is critical for mechanisms such as activity-dependent plasticity and synaptic remodeling [70, 71]. Specifically, regions showing weaker cTBS-related energy changes were enriched for PLS1⁺ genes, potentially reflecting reduced plasticity capacity, whereas PLS1⁻-enriched regions were associated with greater energy optimization, implying a baseline molecular architecture conducive to network restructuring. Beyond molecular associations, TransBrain-based cross-species mapping identifies the murine ILA and PL areas as evolutionary homologs of human FPN-driven control circuits. These regions mediate top-down regulation of subcortical reward circuits via synaptic receptors (e.g., mGluR2, 5-HT2A/2C) [72–74]. This convergence establishes a translational framework in which optogenetic manipulation of ILA/PL may causally validate the human network mechanisms.

Limitations and future research

Several limitations of this study warrant consideration. First, the relatively modest sample size and the unequal group distribution—partially due to COVID-19 disruptions—may limit the statistical power for long-term follow-up and the generalizability of the findings. The cohort was also predominantly male; given sex-related neuroendocrine influences on cortical excitability, future sex-balanced studies are necessary to confirm whether the observed network control mechanisms generalize across sexes. Second, our network control model utilizes a linear framework, which may not fully capture the brain’s complex, nonlinear state transitions. Additionally, while the 10–20 targeting system facilitates clinical scalability, it does not account for individual anatomical variability. This underscores the need for future research to elucidate how individual brain network topology—rather than just anatomical localization—modulates the translation of physical electromagnetic stimulation into large-scale network control effects. Third, the molecular associations reported here reflect baseline spatial gene expression patterns in healthy donors from the AHBA rather than disease-specific or dynamic transcriptional alterations in AUD. Furthermore, while TransBrain-based mapping supports the evolutionary conservation of the FPN-to-SUB regulatory mechanism, species-specific functional complexities between murine ILA/PL areas and their human homologs remain to be fully elucidated. Finally, despite a rigorously implemented sham protocol, stimulation-related sensory experiences may differ between groups, and blinding integrity was not formally evaluated, precluding definitive verification of participant-level blinding. Additionally, the NCT-based analytical pipeline was not pre-registered. These mechanistic insights should therefore be considered exploratory. Future research incorporating larger multicenter cohorts, maintenance stimulation paradigms, and multimodal integration—such as exploring the role of the glymphatic system in AUD recovery—will be critical to refining these network-based neuromodulation strategies and establishing individualized, mechanistically grounded treatment schedules.

Conclusions

In conclusion, our findings suggest that cTBS applied to the rDLPFC may enhance the top-down regulation of subcortical circuits in patients with AUD. cTBS significantly reduced the theoretical minimum control energy required for the FPN to regulate the SUB, and this reduction was associated with individual-level changes in AUDIT-assessed alcohol-use severity. These findings provide a circuit-level account of candidate mechanisms underlying cTBS-related effects. From the perspective of brain network control dynamics, this study suggests that cTBS may modulate the top-down regulatory pathway from the FPN to the subcortical system, thereby influencing AUD-related functional states. Cross-species analyses further suggest that the identified network-control targets may be anatomically conserved across humans and rodents, highlighting the translational potential of network-based targets. These findings expand the application of NCT in psychiatric disorders and inform the design of future precision neuromodulation strategies based on network dynamics. Taken together, these results support further evaluation of network-informed cTBS strategies for AUD, particularly in larger trials designed to determine whether the lower relapse risk observed here translates into clinically meaningful improvements in alcohol-use outcomes.

Author contributions

MW, HZ, and ZKW conceptualized and designed the study. ZKW, JNZ, NNZ, LJCH, YYL, and GHD performed the investigation. ZKW, JNZ, HZ, NNZ, and MW conducted the formal analysis and data curation. ZKW wrote the original draft. HZ, NNZ, and MW reviewed and edited the manuscript. MW, HZ, and NNZ supervised the project, and MW also handled project administration. JZD and GHD acquired the research funding. All authors read and approved the final manuscript.

Funding

This work was supported by the Noncommunicable Chronic Diseases–National Science and Technology Major Project (2025ZD0550504), the National Natural Science Foundation of China (82370901), the National Key Research and Development Program “Medical Equipment and Biomedical Materials” (2022YFC2407004), the Shanghai Science and Technology Committee (22dz1204700), and the Shanghai Municipal Health Commission (202140098) to J.Z.D.; and the Yunnan Provincial Department of Science and Technology Basic Research Project (No. 202405AC350075 and No. 202501AS070042) to G.H.D.

Data availability

The behavioral and neuroimaging data generated and analyzed during the current study are not publicly available because they contain information that could compromise participant privacy. The data are available from the corresponding author upon reasonable request, subject to institutional ethical guidelines. Human gene expression data used in this study are publicly available through the Allen Human Brain Atlas at https://human.brain-map.org/static/download.

Code availability

Network control analyses were performed using publicly available code adapted from https://github.com/ursbraun/network_control_and_dopamine. Metascape is an open-access resource available at http://metascape.org. Custom analysis scripts 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 https://doi.org/10.1038/s41380-026-03747-5.