An Epicenter-guided computational framework for personalized target of repetitive transcranial magnetic stimulation in schizophrenia Authors: Yingjie Tang, Sisi Jiang, Huan Huang, Xueqing Sun, Chao Mu, Mingjun Duan, Roberto Rodríguez-Labrada, Dezhong Yao, Cheng Luo Published: 2026 ABSTRACT BackgroundThe clinical efficacy of repetitive transcranial magnetic stimulation (rTMS) in schizophrenia is limited by significant inter-individual heterogeneity, largely due to the lack of personalized targeting.HypothesisTo develop and validate an epicenter-guided computational framework that identifies personalized rTMS targets by integrating individual neuroanatomical and connectome profiles.Study designWe established an epicenter-guided framework in a discovery cohort including 532 schizophrenia and 526 healthy controls. The pipeline involved: (1) identifying subject-specific pathological epicenters by correlating regional gray-matter volume reductions with normative functional connectivity profiles; (2) mapping the voxels within the left dorsolateral prefrontal cortex that exhibited maximal functional connectivity to these epicenters; and (3) computing the harmonic centroid of these voxels to define the optimal personalized target. The framework was retrospectively validated in a cohort of schizophrenia receiving a 4-week rTMS treatment.Study resultsThe epicenter-guided framework defined a personalized candidate target for each participant with an identifiable individual epicenter. Individual epicenters converged within higher-order association cortices and subcortical structures, consistent with previous findings in schizophrenia. In the retrospective treatment cohort, nodal integration capability within individual epicenter regions increased significantly after rTMS. Smaller target distance was associated with greater concordance between activity-flow-predicted and observed FC changes and with greater improvement in PANSS negative symptoms.ConclusionsWe developed an epicenter-guided framework to derive personalized DLPFC targets and, in retrospective validation, found that target proximity relates to both FC normalization and symptom improvement. These findings support further prospective testing of connectivity-informed personalized targeting for rTMS in schizophrenia. Introduction Schizophrenia is a severe mental illness characterized by positive symptoms, negative symptoms, cognitive impairments, and so on (Tandon et al., 2009; Andreasen, 1982). Typically emerging in late adolescence or early adulthood, schizophrenia imposes significant disability, morbidity, and financial burdens on patients and their families (Kahn and Keefe, 2013). Currently, antipsychotic drug therapy is the primary clinical treatment to alleviate the symptoms of schizophrenia. However, approximately one-third of patients experience minimal or no therapeutic response, leading to the development of treatment-resistant schizophrenia (Shergill et al., 1998). Consequently, there is an urgent need to explore non-pharmacological treatment options for these patients. Non-invasive brain stimulation provides an adjunctive treatment option for schizophrenia (Wang et al., 2017; Osoegawa et al., 2018; He et al., 2017; Aleman et al., 2018). rTMS was first established as a treatment for depression (Berlim et al., 2014), and the left DLPFC is a common stimulation target because of its role in cognitive control (Lawson McLean, 2019; Niendam et al., 2012). Evidence for rTMS efficacy in schizophrenia remains mixed across symptom domains and stimulation protocols (Hoffman et al., 1999; Hoffman et al., 2000; Hoffman et al., 2003). A systematic review similarly found heterogeneous evidence for positive symptoms (Marzouk et al., 2020). This variability may reflect differences in brain circuitry and the lack of patient-specific DLPFC localization (Homan et al., 2012; Guerra et al., 2020). Conventional targeting commonly uses the 5–6-cm rule from the motor cortex or the EEG F3 position. These scalp-based rules do not account for individual network anatomy. Neuronavigation and alternative stimulation protocols may improve targeting and delivery (Guan et al., 2020; Singh et al., 2020; Kumar et al., 2020; Guleken et al., 2020), but the most relevant site within the broad DLPFC region remains uncertain. The depression studies discussed below differ in both the stimulation protocol and the method used to select the target. Connectivity guidance refers to site selection, whereas 10-Hz rTMS and intermittent theta-burst stimulation (iTBS) refer to pulse delivery. Importantly, FC was not used to select the treated site in every study. Fox et al. retrospectively related reported DLPFC sites to subgenual cingulate connectivity and did not deliver a new intervention (Cash et al., 2019; Weigand et al., 2018; Fox et al., 2012). In the THREE-D trial, both 10-Hz rTMS and iTBS were delivered to the same MRI-neuronavigated group coordinate derived from prior outcome and connectivity data, rather than to targets derived from each participant's FC (Blumberger et al., 2018; Li et al., 2020). Li et al. compared prolonged iTBS, 10-Hz rTMS, and sham stimulation while separately comparing the 5-cm rule with anatomical MRI-guided positioning; target selection was not FC-guided (Blumberger et al., 2018; Li et al., 2020). The early site-connectivity findings also came from rTMS cohorts whose treatment sites were not prospectively selected by FC. Cash et al. evaluated Beam-F3-targeted 10-Hz rTMS and calculated subgenual connectivity after the treatment site had been selected (Cash et al., 2019; Weigand et al., 2018; Fox et al., 2012). Weigand et al. evaluated conventional rTMS delivered to sites selected by the 5.5-cm rule or task-perfusion mapping, and then tested normative subgenual connectivity of the recorded sites (Cash et al., 2019; Weigand et al., 2018; Fox et al., 2012). Both studies associated stronger negative site-to-subgenual connectivity with better antidepressant outcome. Later evidence was less consistent. A methodological review concluded that the clinical value of personalized targeting remains uncertain (Lynch et al., 2023; Cash and Zalesky, 2024; Khosravani et al., 2026). Khosravani et al. likewise studied naturalistic rTMS initiated with 10-Hz left-DLPFC stimulation. Sites were selected by scalp-based or anatomical rules rather than FC, and the primary association with normative subgenual connectivity was null (Lynch et al., 2023; Cash and Zalesky, 2024; Khosravani et al., 2026). We therefore use this literature to support the biological plausibility of circuit-informed targeting, not to claim established clinical superiority. Prefrontal functional organization varies across individuals (Doucet et al., 2019; Mueller et al., 2013; Finn et al., 2015; Rajkowska and Goldman-Rakic, 1995; Bruno et al., 2022), and cytoarchitectonic locations also vary (Doucet et al., 2019; Mueller et al., 2013; Finn et al., 2015; Rajkowska and Goldman-Rakic, 1995; Bruno et al., 2022). This variability motivates patient-level localization in schizophrenia. More broadly, psychiatric disorders involve distributed networks rather than isolated regions (Fornito et al., 2017; Shafi et al., 2012). Heterogeneous neuroimaging findings in schizophrenia complicate the identification of robust biomarkers for diagnosis and treatment response. The epicenter model tests whether the connectivity profile of a candidate region can account for a distributed disease-related alteration pattern (van den Heuvel and Seoane, 2024; Zhou et al., 2012). This model has been applied to schizophrenia and epilepsy (Shafiei et al., 2020; Jiang, 2024; Lariviere et al., 2020). Previous schizophrenia studies inferred group-level epicenters by linking gray-matter alterations to distributed connectivity patterns (Shafiei et al., 2020; Jiang, 2024; Lariviere et al., 2020), but group averages do not capture patient-level variation. The resulting epicenter set was distributed across cortical and subcortical regions. Some candidate epicenters, including deeply embedded cortical and subcortical nodes, cannot be stimulated directly with conventional rTMS. We therefore retained these regions as pathological network anchors and used patient-specific functional connectivity to identify a connected site within the accessible left DLPFC. This strategy seeks network-mediated cortical access to the identified anchors; it does not relocate an epicenter, nor does it assume that the DLPFC is the pathological origin or the only possible cortical entry point. In this study, we developed an epicenter-guided computational framework to derive individualized candidate rTMS coordinates in schizophrenia. Each participant's whole-brain GMV-deviation pattern was first used to identify candidate pathological epicenters against a normative FC architecture. Rather than prioritizing a single seed, the retained epicenters were treated as a distributed set of pathological anchors. Patient-specific native-space FC then linked each anchor to a candidate voxel within the left DLPFC, and harmonic-centrality fusion produced one coordinate near the consensus configuration of those candidates. This epicenter-first sequence personalizes both the disease-related starting point and the accessible cortical route. We hypothesized that a smaller native-space distance between the framework-derived coordinate and the canonical F3-associated reference would be associated with greater imaging and clinical change. We tested this proximity hypothesis retrospectively in a treatment cohort with longitudinal pre/post MRI and PANSS measurements. The overall workflow is summarized in Fig. 1.Fig. 1Workflow for deriving personalized candidate rTMS targets. (A) MNI-space visualization of patient-specific epicenters identified by relating individual GMV-deviation maps to a normative 254 × 254 FC template. (B) Native-space target derivation: within each participant, the DLPFC voxel with maximal connectivity to each epicenter was selected, and the candidate coordinates were fused by harmonic centrality. The canonical F3 coordinate was defined in MNI space and transformed to native space before distance calculation; personalized targets were projected to MNI space only for group visualization.Fig. 1 Workflow for deriving personalized candidate rTMS targets. (A) MNI-space visualization of patient-specific epicenters identified by relating individual GMV-deviation maps to a normative 254 × 254 FC template. (B) Native-space target derivation: within each participant, the DLPFC voxel with maximal connectivity to each epicenter was selected, and the candidate coordinates were fused by harmonic centrality. The canonical F3 coordinate was defined in MNI space and transformed to native space before distance calculation; personalized targets were projected to MNI space only for group visualization. Methods Participants and cohorts Discovery cohort: Data for algorithm development (the discovery cohort) were acquired from five independent sites and comprised 526 healthy adults (mean age 34.03 ± 12.61 years; 296 males) and 532 patients with schizophrenia (mean age 35.46 ± 12.65 years; 305 males). Retrospective treatment cohort: The treatment cohort included 27 inpatients with schizophrenia recruited from the Clinical Hospital of Chengdu Brain Science Institute. A same-scanner healthy-control comparison group of 88 participants was used to calculate GMV-deviation maps and identify candidate epicenters. Detailed participant characteristics and MRI acquisition parameters for the discovery cohort and retrospective treatment cohort are provided in Supplementary Methods S1. Data preprocessing Functional and structural MRI data were preprocessed using established pipelines (NIT (Dong et al., 2018) and CAT12). Full preprocessing parameters and QC procedures are detailed in Supplementary Methods S2. Personalized target derivation Individual-level epicenter identification For each participant, voxel-wise GMV-deviation scores were aggregated to a 254-region vector. For each candidate region, we correlated its healthy-control normative whole-brain FC profile with the participant's GMV-deviation vector. Regions with r > 0 and spatially corrected Pspin <0.01 were retained as candidate epicenters (Fig. 1A). This operation follows the patient-tailored epicenter method applied to individual GMV-deviation maps in schizophrenia (Shafiei et al., 2020; Jiang, 2024; Lariviere et al., 2020). The GMV-deviation vector supplied the patient-specific disease-related pattern, whereas healthy-control FC supplied the reference intrinsic-connectivity architecture used to evaluate every candidate region. The healthy-control template was not treated as the participant's own FC. The participant's pre-treatment FC was introduced only in the subsequent native-space target-derivation step. Full computational details, parcellation mapping, and spin-test procedures are reported in Supplementary Methods S3. Candidate DLPFC coordinate selection and fusion Because the retrospective treatment cohort had received left-DLPFC rTMS, the cortical search was restricted to the left DLPFC. This restriction reflected the treatment space available for retrospective evaluation, rather than an assumption that the DLPFC was the origin of the pathological process. Each participant's retained epicenters were used as network anchors. Patient-specific native-space FC was then used to identify the DLPFC voxel most strongly connected to each epicenter before the candidate coordinates were fused. We combined information across all subject-specific epicenter-derived FC maps to determine personalized rTMS targets. After identifying each subject's set of epicenters, all subsequent FC calculations were performed in the subject's native space to avoid interpolation artifacts arising from inter-individual differences in head size and anatomy. Using pre-treatment resting-state data, we computed voxel-wise FC between every voxel within the left DLPFC mask and each epicenter, yielding an ordered series of voxel-wise FC maps V1, …, VM. For each map Vk, a candidate coordinate Pk was defined by identifying the voxel that exhibited the maximum absolute connectivity strength. This step produced M candidate coordinates, each representing the peak functional link between the DLPFC and a specific epicenter. To integrate these M candidate coordinates into a single, robust stimulation target T, we employed a Harmonic Centrality fusion algorithm. This approach was specifically chosen for its ability to identify the “topological center” of the coordinate cluster while maintaining high robustness against spatial outliers. Unlike a simple arithmetic mean, which can be easily skewed by atypical coordinates resulting from noise, the harmonic algorithm utilizes the sum of reciprocal distances. In this method, the contribution of a candidate point Pk diminishes rapidly as its distance from the potential target increases; consequently, spatially isolated outliers contribute negligibly to the total score. This ensures that the final target T is naturally anchored within the most consistent “consensus cluster” formed by the majority of epicenter-linked points. This measure reflects the global integration of candidate voxels with respect to all identified epicenters, indicating their potential to influence distributed pathological networks. The final personalized target T was defined as the voxel x within the DLPFC that maximizes the harmonic score Hx:Hx=∑k=1M1x−Pk+1T=argmaxx∈DLPFCHxwhere x−Pk denotes the Euclidean distance, and a constant of 1 is added to prevent division by zero. (Fig. 1B). Group-level epicenter overlap To identify pathological epicenters with high spatial consistency across the schizophrenia cohort, we employed a non-parametric bootstrap framework (1000 iterations with replacement). First, the overlap probability for each of the 254 ROIs was calculated, representing the proportion of subjects in which a specific region was identified as a significant hub. To ensure the robustness of our findings and mitigate the influence of individual outliers, we estimated the resampling distribution of these probabilities at both the individual ROI level and across eight cognitive networks (Yeo-7 networks plus a subcortical network). Final results are reported as the bootstrap mean and the 95% confidence interval (CI), derived from the 2.5th and 97.5th percentiles. This procedure ensures that the identified epicenters reflect stable, reproducible pathological characteristics of the population rather than idiosyncratic variation. PCC amplitude-defined analysis of spontaneous resting-state BOLD activity We compared the functional context of personalized coordinates and the F3-associated reference using a PCC amplitude-defined analysis of spontaneous resting-state BOLD activity. The PCC was selected because it is a central hub of the default mode network (Fransson and Marrelec, 2008). The default mode network supports internally directed and self-referential cognition (Buckner et al., 2008). The PCC time series served as a predefined endogenous reference signal; the analysis did not assume causal influence from the PCC. Event-based resting-state fMRI studies have shown that brief high-amplitude BOLD events retain information about intrinsic network organization (Liu and Duyn, 2013). Related work has detected spontaneous seed-region BOLD excursions and modeled spontaneous events with HRF-convolved GLMs to obtain statistical maps of spontaneous activity (Tagliazucchi et al., 2011; Petridou et al., 2013). For each participant, time points in the highest and lowest 5% of the PCC BOLD amplitude distribution were defined as PCC-high and PCC-low events, respectively. Because the selected time points represented observed BOLD extrema, two TRs were subtracted from each onset index to approximate the preceding neural event. Separate first-level models for the PCC-high and PCC-low events were estimated in SPM12, with zero-duration event regressors convolved with the canonical hemodynamic response function (Friston et al., 1998). For participant-level activity-flow mapping, the within-participant difference between the PCC-high and PCC-low beta maps defined the observed whole-brain map of spontaneous brain activity; for coordinate-level analysis, the same spontaneous brain activity contrast was sampled at the personalized coordinate and the F3-associated reference. Target-level values were harmonized across sites with ComBat and compared between the personalized coordinates and the F3-associated reference. Full model specifications are provided in Supplementary Methods S4. Retrospective validation: Activity-flow mapping, ΔFC, and clinical associations Longitudinal study design and treatment cohort We retrospectively analyzed longitudinal pre/post changes associated with high-frequency rTMS in patients with schizophrenia (Fig. 2A). MRI and PANSS assessments were obtained at baseline (t1) and approximately 1 month later (t2). Twenty-seven patients (20 male and 7 female) received 4 weeks of high-frequency rTMS while continuing a stable antipsychotic regimen. The prescribed antipsychotic regimen was maintained throughout the treatment period. The antipsychotic medications used in the retrospective cohort were clozapine, quetiapine, risperidone, and aripiprazole, administered alone or in combination. Chlorpromazine-equivalent dose is reported in Table 2. The study was approved by the Ethics Committee of Chengdu Mental Health Centre, and all participants provided written informed consent.Fig. 2Retrospective validation experiment design and analysis workflow. (A) Timeline of the validation study (baseline t1 and 1-month follow-up t2) with fMRI and PANSS assessments. (B) Example subject showing the personalized target (yellow) and the canonical F3-associated reference (purple). Distance denotes the native-space Euclidean distance (mm) between the two coordinates. (C) Activity-flow validation schematic: regression-based activity-flow produces a predicted change-in-FC matrix (ΔFC₁) from a simulated epicenter perturbation, which is compared to the empirically observed post−pre ΔFC (ΔFC₂) at epicenter-related edges; representative heatmaps show observed and predicted edge responses pre/post.Fig. 2 Retrospective validation experiment design and analysis workflow. (A) Timeline of the validation study (baseline t1 and 1-month follow-up t2) with fMRI and PANSS assessments. (B) Example subject showing the personalized target (yellow) and the canonical F3-associated reference (purple). Distance denotes the native-space Euclidean distance (mm) between the two coordinates. (C) Activity-flow validation schematic: regression-based activity-flow produces a predicted change-in-FC matrix (ΔFC₁) from a simulated epicenter perturbation, which is compared to the empirically observed post−pre ΔFC (ΔFC₂) at epicenter-related edges; representative heatmaps show observed and predicted edge responses pre/post. rTMS protocol and F3-associated reference The treatment cohort was previously acquired, and the present study is a retrospective analysis of those longitudinal rTMS data rather than a prospective personalized-targeting trial. Stimulation parameters and coil placement were determined by the routine clinical protocol used at the time of treatment, not by the epicenter-guided framework proposed here (Schneider et al., 2008; Wobrock et al., 2015). rTMS was delivered with a YRDCCY1 stimulator (Yiruide Medical Equipment New Technology Co., Ltd., Wuhan, China) and a figure-of-eight coil (B9076). The clinical target was the left DLPFC localized with the 10–20 EEG F3 scalp heuristic, a practical scalp-based approximation widely used in TMS practice (Herwig et al., 2003; Beam et al., 2009; Mir-Moghtadaei et al., 2015). During treatment, the intersection of the two coil loops was aligned with the F3 position of the 10–20 EEG system (Turriziani et al., 2010). Individual neuronavigated coil coordinates, coil orientation, and session-wise displacement were not recorded. For the retrospective spatial analysis, we represented the F3 localization rule used during treatment with the canonical MNI coordinate [−43, 58, 40]. Cash et al. tabulated this coordinate as an EEG F3 representation based on the cranio-cerebral correspondence reported by Okamoto et al. (Okamoto et al., 2004; Cash et al., 2021a). This coordinate was specified before the outcome analysis and was not used to derive the personalized target. It was transformed into each participant's native space with the participant-specific registration transform before the distance calculation. Stimulation was delivered at 10 Hz and 110% of the resting motor threshold. The resting motor threshold was determined before treatment with a standardized five-step procedure (Schutter and van Honk, 2006). Participants received 20 sessions over 4 weeks (Schneider et al., 2008; Wobrock et al., 2015). Each session delivered 1000 pulses in 20 trains, with 50 pulses per train, a train duration of 5 s, an inter-train interval of 30 s, and a total duration of 11 min 40 s. To evaluate whether the identified targets produce biologically meaningful network-level effects, we employed an independent validation approach. We assume that rTMS perturbs local nodes and that such perturbations propagate through the brain's FC topology to produce distributed changes in activation. To quantify this propagation, we applied activity flow mapping, a validated framework that models target activations as the weighted sum of activations in other regions with weights defined by FC (Cole et al., 2016). This method has been validated in both healthy subjects and schizophrenia, and has also been applied to resting-state data (Ito et al., 2017; Hearne et al., 2021). Therefore, this study utilizes activity flow mapping to characterize the propagation of resting-state brain activity between regions. Activity-flow prediction procedure For each subject, the spontaneous brain activity contrast map described above served as the empirical whole-brain activity pattern, and we computed a subject-wise 254 × 254 z-transformed FC matrix. The activity-flow prediction for any target region is the FC-weighted sum of activity in all other regions (Fig. 2B). We operationalized two summary metrics: (1) group-level concordance (Spearman correlation between group-mean predicted and observed activation maps) and (2) Nodal integration capability (IC), which quantifies, for each region, how well distributed inputs predict that region's activation across subjects. Full equations and definitions are in Supplementary Methods S4. Simulation of expected FC changes: Regression-based fitting To simulate how perturbing an epicenter would alter network connectivity, we fit separate linear regression models that map pre-TMS co-activation in each epicenter region to co-activation patterns in healthy controls; the learned β weights were treated as the simulated FC linking that epicenter to other regions (Fig. 2C). Detailed protocols are provided in Supplementary Methods S5. Comparing simulated and observed ΔFC Previous research utilized a calculation method similar to the individual epicenter analysis described in this paper to identify epicenters in a large sample of schizophrenia patients. In that study, 46 key regions were selected as the epicenters representing the schizophrenic population (Jiang, 2024). In the analyses below, we treated these 46 regions as the target epicenter set. This choice was motivated by two primary considerations: (1) it establishes a common, fixed feature dimensionality across all subjects, facilitating subsequent group-level comparisons and multivariate analyses; and (2) it reduces variability and idiosyncratic bias introduced by individualized epicenter selection, thereby improving statistical power. We defined ΔFC₁ as the difference between the simulated (expected) FC and the subject's pre-TMS FC, and ΔFC₂ as the empirically observed difference between post-TMS FC and pre-TMS FC. For each subject we therefore concatenated the columns corresponding to the 46 epicenter regions into a single edge vector (46 × 254 = 11,684 elements), where each element represents the connection strength between one epicenter region and one target region in the 254-region atlas.∆FC1=FCsimulated−FCpre∆FC2=FCpost−FCpre For each subject, we computed the Pearson correlation between ΔFC₁ and ΔFC₂ computed over population epicenter-related edges; this Pearson similarity quantifies how closely the expected FC reproduces the empirically observed FC change. We then assessed whether this similarity was related to deviation distance (i.e., whether closer targeting yielded ΔFC₂ that more closely matched the simulated normalization ΔFC₁) (Fig. 2C). Clinical validation: Target proximity and PANSS change Spearman correlation analysis was used to assess the relationship between changes in PANSS scores and deviation distances of schizophrenia subjects before and after intervention to investigate the relationship between the intervention target and the intervention effect. Results Participant demographics and clinical characteristics The discovery cohort used for framework development comprised 526 healthy controls and 532 patients with schizophrenia. Demographic and clinical characteristics are summarized in Table 1.Table 1Demographic information of the discovery cohort.SZMean (SD)HCMean (SD)p valueSex (male/female)305/227296/2300.73aAge (years)35.46(12.65)34.03(12.61)0.07bPositive17.82 (6.83)–Negative18.54 (7.31)–General33.35(9.14)–Total69.71(18.65)– Table 1Demographic information of the discovery cohort.SZMean (SD)HCMean (SD)p valueSex (male/female)305/227296/2300.73aAge (years)35.46(12.65)34.03(12.61)0.07bPositive17.82 (6.83)–Negative18.54 (7.31)–General33.35(9.14)–Total69.71(18.65)– Demographic information of the discovery cohort. The retrospective treatment cohort comprised 27 inpatients with schizophrenia who underwent high-frequency rTMS and 88 healthy control participants. Demographic and clinical characteristics are summarized in Table 2.Table 2Demographic information for the retrospective treatment cohort.rTMS-SZMean (SD)HCMean (SD)p valueSex (male/female)20/735/530.004aAge (years)46.70(11.73)44.03(12.04)0.32bBaseline PANSS scorePositive11.70 (4.16)–Negative23.56 (6.70)–General30.41(6.13)–Total65.67(12.70)–1-month follow upPANSS scorePositive9.96(3.44)–0.02cNegative18.07(5.06)–<0.001cGeneral25.81(6.09)–<0.001cTotal53.85(11.69)–<0.001cCPZ-equivalent dose291.63(106.91)–PANSS, Positive and Negative Syndrome Scale; control; SD, standard deviation.a. P values were obtained from the Chi-square.b. P values were obtained from the two-sample t-test.c. P values were obtained from the paired-sample t-test. Table 2Demographic information for the retrospective treatment cohort.rTMS-SZMean (SD)HCMean (SD)p valueSex (male/female)20/735/530.004aAge (years)46.70(11.73)44.03(12.04)0.32bBaseline PANSS scorePositive11.70 (4.16)–Negative23.56 (6.70)–General30.41(6.13)–Total65.67(12.70)–1-month follow upPANSS scorePositive9.96(3.44)–0.02cNegative18.07(5.06)–<0.001cGeneral25.81(6.09)–<0.001cTotal53.85(11.69)–<0.001cCPZ-equivalent dose291.63(106.91)–PANSS, Positive and Negative Syndrome Scale; control; SD, standard deviation.a. P values were obtained from the Chi-square.b. P values were obtained from the two-sample t-test.c. P values were obtained from the paired-sample t-test. Demographic information for the retrospective treatment cohort. PANSS, Positive and Negative Syndrome Scale; control; SD, standard deviation. a. P values were obtained from the Chi-square. b. P values were obtained from the two-sample t-test. c. P values were obtained from the paired-sample t-test. Group-level spatial consistency and network distribution of epicenters (discovery) Individual epicenters were computed for each participant, and only subjects satisfying the epicenter selection criterion (r > 0 and Pspin <0.01) were retained. A total of 76 subjects failed to satisfy this criterion and were excluded, yielding a final discovery sample of 450 subjects. Network-level preference: The spatial distribution of epicenters exhibited a distinct non-random pattern, as evidenced by the markedly higher overlap probabilities in higher-order associative networks compared to primary sensory networks. Among canonical functional systems, the Limbic Network showed the highest overlap probability (Mean = 0.135, 95% CI [0.120, 0.153]). Elevated overlap probabilities were also observed in the Salience/Ventral Attention Network (Mean = 0.118, 95% CI [0.105, 0.130]) and the Executive Control Network (Mean = 0.110, 95% CI [0.098, 0.121]); the Default Mode Network displayed a comparable level (Mean = 0.108). By contrast, the Visual (Mean = 0.080) and Somatomotor (Mean = 0.070) networks exhibited the lowest epicenter probabilities. These results indicate preferential involvement of higher-order associative networks rather than primary sensory systems in the population-level epicenter distribution (Fig. 3B).Fig. 3Spatial consistency and network-level distribution of group-level epicenters in the discovery cohort. (A) Surface and subcortical maps of epicenter overlap probability. Epicenter overlap maps show the probability that each region was identified as an epicenter across subjects; color scale indicates overlap probability (yellow: low, high: purple). (B) Network-level overlap probability of epicenters. Radial plot of mean epicenter overlap probability for canonical functional systems (filled polygons) with 95% confidence intervals (thin outlines). Representative cortical surface renderings illustrate the spatial distribution of each canonical functional network. (C) ROI-level ranked epicenter probabilities. Bars show mean epicenter probability for each ROI sorted descending (Threshold >15%); vertical whiskers denote 95% confidence intervals estimated by bootstrap (1000 resamples).Fig. 3 Spatial consistency and network-level distribution of group-level epicenters in the discovery cohort. (A) Surface and subcortical maps of epicenter overlap probability. Epicenter overlap maps show the probability that each region was identified as an epicenter across subjects; color scale indicates overlap probability (yellow: low, high: purple). (B) Network-level overlap probability of epicenters. Radial plot of mean epicenter overlap probability for canonical functional systems (filled polygons) with 95% confidence intervals (thin outlines). Representative cortical surface renderings illustrate the spatial distribution of each canonical functional network. (C) ROI-level ranked epicenter probabilities. Bars show mean epicenter probability for each ROI sorted descending (Threshold >15%); vertical whiskers denote 95% confidence intervals estimated by bootstrap (1000 resamples). ROI-level distribution: At the regional level, high-consistency epicenters in the discovery cohort were distributed across cortical association regions and selected subcortical structures. The right pre-supplementary motor area, a key medial node of the Salience/Ventral Attention network, demonstrated the greatest population-level consistency (Mean = 0.180, 95% CI [0.148, 0.212]). This was followed by the right medial orbitofrontal cortex (Mean = 0.180, 95% CI [0.147, 0.212]) and the right hippocampal head (Mean = 0.178, 95% CI [0.145, 0.212]). Furthermore, cortical epicenters were also identified in the posterior cingulate cortex (Mean = 0.170) and the anterior insula (Mean = 0.145). Subcortical structures, specifically the ventroposterolateral nucleus of the thalamus (Mean = 0.150) and the dorsal putamen (Mean = 0.123), also showed overlap and formed part of this distributed pattern. The positive 95% confidence intervals for the leading ROIs supported the stability of these overlap estimates in the discovery sample (Fig. 3C). Detailed regional statistics are provided in Supplementary Table S1. Group-level spatial consistency and network distribution of epicenters (validation) Applying the same epicenter selection criterion (r > 0 and Pspin <0.01), 3 subjects did not satisfy the criterion and were excluded, resulting in a final validation sample of 24 subjects. Network-level preference: In the retrospective treatment cohort, the spatial distribution of epicenters demonstrated a distinct non-random pattern across functional systems. The Salience/Ventral Attention Network exhibited the highest overlap probability (Mean = 0.149, 95% CI [0.098, 0.202]), followed by the Default Mode Network (Mean = 0.043, 95% CI [0.024, 0.066]) and the Executive Control Network (Mean = 0.021, 95% CI [0.007, 0.040]). Similar to the discovery cohort, the primary sensory networks, including the Visual (Mean = 0.001) and Somatomotor (Mean = 0.003) systems, showed the lowest epicenter probabilities. These results show that pathological epicenters in the treatment cohort are preferentially localized within higher-order cognitive and attentional networks (Fig. 4B).Fig. 4Spatial consistency and network-level distribution of population-level epicenters in the retrospective treatment cohort. (A) Surface and subcortical maps of epicenter overlap probability. (B) Network-level overlap probability of epicenters. (C) ROI-level ranked epicenter probabilities (threshold >25%).Fig. 4 Spatial consistency and network-level distribution of population-level epicenters in the retrospective treatment cohort. (A) Surface and subcortical maps of epicenter overlap probability. (B) Network-level overlap probability of epicenters. (C) ROI-level ranked epicenter probabilities (threshold >25%). ROI-level Distribution: Regional analysis further pinpointed the most consistent epicenters within subcortical hubs and specific salience nodes. The top three regions with the highest population-level consistency were all localized in the subcortical nuclei: the left dorsal putamen (Mean = 0.633, 95% CI [0.444, 0.815]), the right anterior globus pallidus (Mean = 0.597, 95% CI [0.444, 0.778]), and the left dorsal lateral thalamus (Mean = 0.588, 95% CI [0.407, 0.778]). Notably, within the cortical surface, the left frontoparietal opercular insula (Mean = 0.448, 95% CI [0.259, 0.630]) emerged as the most prominent cortical epicenter. (Fig. 4C). Detailed regional statistics are provided in Supplementary Table S2. Distribution and functional comparison of personalized coordinates and the F3 reference By analysing the spatial coordinates of personalized targets from the discovery set, we found that personalized targets are distributed with substantial heterogeneity in standard space (Fig. 5A). This heterogeneity indicates that relying on a single F3-associated reference coordinate overlooks inter-individual differences in functional organization, and that the framework proposed here can capture those subject-specific structural and functional details. We observed the same heterogeneous distribution in the validation set (n = 24; three subjects without identifiable individual epicenters were excluded) (Fig. 7A). We also computed the Euclidean distance between each personalized coordinate and the F3-associated reference in both datasets.Kernel Density Estimation plots show a high degree of overlap between the two cohorts' distance distributions (Fig. 5B). The mean displacement was 6.23 ± 2.30 mm for the discovery set and 5.82 ± 2.00 mm for the validation set; the Mann–Whitney U test indicates no significant difference between these distributions (p = 0.5489).Fig. 5Spatial distribution and functional comparison of personalized coordinates and the F3 reference, and their spatial separation. (A) In the discovery cohort, projected locations of personalized coordinates (yellow markers) and the F3-associated reference (purple marker) in standard space. (B) Kernel density estimates of native-space Euclidean distances between each personalized coordinate and the F3-associated reference for the discovery cohort (blue) and retrospective treatment cohort (red); vertical lines mark group means. Mean ± SD distance: discovery cohort, 6.23 ± 2.30 mm; retrospective treatment cohort, 5.82 ± 2.00 mm. The Mann-Whitney U test showed no significant difference between the two distributions (p = 0.5489). (C) Schematic overview of the PCC amplitude-defined spontaneous BOLD event analysis. Red and blue dots indicate time points with the highest and lowest PCC activity, respectively. The purple mask shows the PCC ROI used for time-series extraction. The central brain maps illustrate instantaneous states at extreme PCC activity, and the right brain maps show the corresponding spontaneous brain activity patterns. (D) Paired comparisons of spontaneous brain activity contrast estimates at the personalized coordinates and the F3-associated reference in the discovery cohort. Thin lines show participant-level paired data; boxes show the median and interquartile range; violin plots show the marginal distributions. Personalized candidate coordinates were calculated in each participant's native space and projected to standard space only for visualization. Participants without an identifiable individual epicenter were excluded from the corresponding analyses.Fig. 5 Spatial distribution and functional comparison of personalized coordinates and the F3 reference, and their spatial separation. (A) In the discovery cohort, projected locations of personalized coordinates (yellow markers) and the F3-associated reference (purple marker) in standard space. (B) Kernel density estimates of native-space Euclidean distances between each personalized coordinate and the F3-associated reference for the discovery cohort (blue) and retrospective treatment cohort (red); vertical lines mark group means. Mean ± SD distance: discovery cohort, 6.23 ± 2.30 mm; retrospective treatment cohort, 5.82 ± 2.00 mm. The Mann-Whitney U test showed no significant difference between the two distributions (p = 0.5489). (C) Schematic overview of the PCC amplitude-defined spontaneous BOLD event analysis. Red and blue dots indicate time points with the highest and lowest PCC activity, respectively. The purple mask shows the PCC ROI used for time-series extraction. The central brain maps illustrate instantaneous states at extreme PCC activity, and the right brain maps show the corresponding spontaneous brain activity patterns. (D) Paired comparisons of spontaneous brain activity contrast estimates at the personalized coordinates and the F3-associated reference in the discovery cohort. Thin lines show participant-level paired data; boxes show the median and interquartile range; violin plots show the marginal distributions. Personalized candidate coordinates were calculated in each participant's native space and projected to standard space only for visualization. Participants without an identifiable individual epicenter were excluded from the corresponding analyses. As a secondary functional characterization, we compared spontaneous brain activity contrast estimates at the personalized coordinates and the F3-associated reference. In the discovery cohort, contrast estimates were higher at personalized coordinates than at the F3-associated reference (paired t-test, p < 0.001; Fig. 5D). This difference indicates stronger differentiation of spontaneous brain activity between the predefined PCC amplitude states at the personalized coordinates; it does not indicate higher absolute resting-state activity. Retrospective associations between target proximity, network change, and symptom outcome First, using activity flow mapping we evaluated whether model predictive accuracy improved following rTMS. At the group level, the Spearman correlation between the group-mean predicted and group-mean observed activation maps increased from r = 0.60 before rTMS (p = 5.64 × 10−26) to r = 0.75 after rTMS (p = 1.31 × 10−46), yielding Δr = 0.15; this change exceeded the permutation null distribution (permutation p < 0.001) (Fig. 6A) (Supplementary Fig. S1). At the region-level, Nodal IC increased from 0.44 ± 0.17 (pre) to 0.49 ± 0.18 (post), and this difference was statistically significant (p < 0.005) (Supplementary Fig. S1). Furthermore, we observed that Nodal IC within the epicenter regions was significantly elevated following rTMS intervention compared to baseline (p = 0.045, Fig. 6B). For this analysis, the epicenters were defined as the union of individual epicenter maps, restricted to regions appearing with a frequency of at least two across the cohort (n = 59ROIs). Overall, the observed increases in Nodal IC together with the enhanced group-level Spearman concordance indicate an increase in integrative capacity following rTMS, potentially reflecting greater engagement of distributed network interactions after stimulation.Fig. 6Group-level and region-level integration effects revealed by activity-flow mapping before and after rTMS. (A) Top panels show group-level spatial concordance between activity-flow predicted and empirically observed activation patterns. For each condition (pre-rTMS and post-rTMS), cortical maps display the group-mean predicted activation pattern (left) and the corresponding group-mean observed activation pattern (right). Scatter plots depict correspondence between predicted and observed group-mean activations, quantified using Spearman rank correlation. Group-level concordance was significantly higher following rTMS, as assessed by a paired-label permutation test (n = 5000, p < 0.001). (B) Nodal IC changes within epicenter regions following rTMS intervention.Fig. 6 Group-level and region-level integration effects revealed by activity-flow mapping before and after rTMS. (A) Top panels show group-level spatial concordance between activity-flow predicted and empirically observed activation patterns. For each condition (pre-rTMS and post-rTMS), cortical maps display the group-mean predicted activation pattern (left) and the corresponding group-mean observed activation pattern (right). Scatter plots depict correspondence between predicted and observed group-mean activations, quantified using Spearman rank correlation. Group-level concordance was significantly higher following rTMS, as assessed by a paired-label permutation test (n = 5000, p < 0.001). (B) Nodal IC changes within epicenter regions following rTMS intervention. Then, we simulated expected FC interventions (ΔFC₁) derived from the activity-flow regression models and compared them to empirically observed FC changes (ΔFC₂). Across ROI models, average model errors were MSE = 0.01 and MAE ≈ 0.0764. Importantly, the similarity between ΔFC₁ and ΔFC₂ was negatively correlated with deviation distance (Spearman r = −0.45, p = 0.03; Fig. 7B). Subjects whose stimulation sites were closer to their personalized targets showed greater concordance between simulated and observed FC changes, indicating that more accurate targeting was associated with FC changes closer to the normative pattern.Fig. 7Target distance and functional and clinical outcomes. (A) In the retrospective treatment cohort, the personalized coordinates (yellow markers) and the canonical F3-associated reference (purple marker) are projected to standard space for visualization. (B) Association between native-space Euclidean distance (mm; x-axis) and ΔFC similarity (y-axis), where ΔFC similarity denotes the participant-wise concordance between activity-flow-predicted ΔFC and empirically observed ΔFC over epicenter-related edges. Spearman r = −0.45, p = 0.03. (C) Association between native-space distance (mm; x-axis) and relative improvement in negative symptoms (ΔNegative / baseline; y-axis). Spearman r = −0.42, p = 0.04. In panels B and C, the yellow line and shaded band show the fitted line and confidence band for visualization.Fig. 7 Target distance and functional and clinical outcomes. (A) In the retrospective treatment cohort, the personalized coordinates (yellow markers) and the canonical F3-associated reference (purple marker) are projected to standard space for visualization. (B) Association between native-space Euclidean distance (mm; x-axis) and ΔFC similarity (y-axis), where ΔFC similarity denotes the participant-wise concordance between activity-flow-predicted ΔFC and empirically observed ΔFC over epicenter-related edges. Spearman r = −0.45, p = 0.03. (C) Association between native-space distance (mm; x-axis) and relative improvement in negative symptoms (ΔNegative / baseline; y-axis). Spearman r = −0.42, p = 0.04. In panels B and C, the yellow line and shaded band show the fitted line and confidence band for visualization. Finally, we tested whether targeting accuracy (deviation distance) predicted clinical benefit. Spearman correlation revealed a significant negative association between deviation distance and improvement in negative symptoms (ΔPANSS; r = −0.42, p = 0.04), indicating that larger spatial deviations from the personalized target were associated with smaller symptom reductions (Fig. 7C). Together, these retrospective findings show that smaller distance between the canonical F3-associated reference and the framework-derived coordinate was associated with greater concordance between simulated and observed FC change and with greater improvement in negative symptoms. Cortical maps illustrate the Nodal IC specifically within the identified epicenter regions for both the pre-rTMS and post-rTMS conditions. The raincloud plot compares Nodal IC at pre-rTMS and post-rTMS for n = 59 epicenter regions. These regions were defined as the union of individual epicenter maps, restricted to those appearing with a frequency of at least two across the cohort. Nodal IC was significantly elevated post-intervention compared to baseline (p = 0.045). Individual dots connected by gray lines represent paired changes for specific ROIs. Half-violin and box plots show the data distribution and summary statistics (median and IQR) for each condition. Discussion Currently, the clinical efficacy of rTMS in schizophrenia remains suboptimal, with a substantial proportion of patients failing to achieve meaningful symptom remission. In this study, we employed an epicenter-mapping approach to develop a computational framework for personalized rTMS target selection. In the discovery cohort, personalized coordinates showed higher spontaneous brain activity contrast estimates than the F3-associated reference. In the retrospective treatment cohort, smaller native-space distance between the canonical F3-associated reference and the framework-derived coordinate was associated with greater improvement in negative symptoms and greater concordance between simulated and observed FC changes. Taken together, these findings provide retrospective support for the potential relevance of epicenter-informed target selection in schizophrenia and motivate prospective studies of connectivity-informed neuromodulation. The prefrontal and parietal cortices show substantial inter-individual variation in functional-network organization (Doucet et al., 2019; Mueller et al., 2013; Finn et al., 2015; Rajkowska and Goldman-Rakic, 1995; Bruno et al., 2022). Connectivity-guided approaches have therefore been developed to localize stimulation targets within the DLPFC (Cash et al., 2021b), and individualized functional networks have informed both non-invasive and invasive stimulation studies (Cash et al., 2019; Weigand et al., 2018; Fox et al., 2012). The present framework differs in its starting point. It first derives multiple patient-specific candidate epicenters from the whole-brain distribution of the individual's GMV deviation. Rather than prioritizing a single seed, the retained epicenters are treated as a distributed set of pathological anchors. Native-space FC then links each anchor to an accessible DLPFC candidate, and harmonic-centrality fusion produces one cortical candidate near their consensus configuration. Personalization therefore begins with the patient's disease-related pattern rather than only with localization relative to a fixed group seed or predefined symptom circuit. The PCC amplitude-defined analysis provided a complementary functional characterization of the resulting coordinates. Higher spontaneous brain activity contrast estimates at the personalized coordinates indicate stronger differentiation between the predefined PCC amplitude states. This measure does not indicate higher absolute activity, causal influence from the PCC, or a TMS-evoked response. It also does not establish clinical superiority. The regional distribution was not uniformly subcortical across cohorts. In the discovery cohort, high-consistency epicenters included cortical association regions and selected subcortical structures; in the retrospective treatment cohort, the three top-ranked regions were subcortical, and the frontoparietal opercular insula was the highest-ranked cortical region. The insula is a salience-network hub involved in interoception, affect, and salience detection, and structural and functional abnormalities of this region have been reported in schizophrenia (Palaniyappan and Liddle, 2012; Wylie and Tregellas, 2010). The insula has also been explored as a target in deep-TMS studies (Ibrahim et al., 2023). Thalamocortical dysconnectivity is a robust network finding in schizophrenia (Woodward et al., 2012), and altered putamen activity has been associated with treatment-response signals (Wu et al., 2019). The cohort-specific regional findings do not mean that any epicenter was relocated to the DLPFC. For anchors beyond direct rTMS reach, the framework asks whether an accessible cortical site is functionally connected to them. Interleaved TMS-fMRI studies show that cortical stimulation can modulate connected subcortical structures and distributed networks (van Holstein et al., 2018; Oathes et al., 2021). Connectome modeling further implicates polysynaptic white-matter routes between DLPFC stimulation sites and the subgenual cingulate cortex (Seguin et al., 2026). Direct TMS-iEEG and TMS-fMRI evidence also shows that FC-guided parietal stimulation can engage and modulate hippocampal activity (Li et al., 2026). These studies support the possibility that connected cortical stimulation can influence deep network nodes, but they do not establish that the DLPFC is the optimal entry point for this cohort. The framework itself is not tied to the DLPFC: a prospective application could use another accessible cortical search space selected for the disorder, symptom, and stimulation modality. The present data do not establish the DLPFC as the only or universally optimal entry point. There is currently a viewpoint suggesting that the broad and subtle variations in FC may have significant effects on behavior and clinical states (Hearne et al., 2021; Cole et al., 2014; Krienen et al., 2014; Spronk et al., 2021). In our final analysis, we simulated putative treatment-related FC changes for schizophrenia based on hypothesized FC changes to strengthen the rationale for target selection. Extending the framework of activity flow mapping, we examined whether FC alterations could facilitate the “neuro-normalization” of aberrant activation patterns. Previous research suggested that limited predictive performance in certain brain regions within activity flow mapping has been attributed to highly localized processing, which the distributed processing assumption fails to capture (Hearne et al., 2021; Ito et al., 2020). Our results indicate that rTMS was associated with significant increases in Nodal integrative capacity together with enhanced group-level spatial concordance. This indicates that, after stimulation, group-mean activation maps were more predictable from distributed activity-flow models, suggesting an enhanced capacity for distributed signal propagation across the brain. At a mechanistic level, this pattern is consistent with increased distributed input to regions that were previously more locally driven, as well as strengthened coupling between higher-order integrative hubs and distributed sources, which together may underlie the therapeutic effects of rTMS. Furthermore, simulating FC changes required to normalize brain activation demonstrated that closer alignment of rTMS intervention with optimal targets led to better outcomes, evidenced by FC patterns resembling those of healthy individuals. Traditional stimulation strategies have largely relied on fixed, population-based coordinates that do not account for inter-individual differences in network architecture. The epicenter-guided approach instead prioritizes patient-specific network pathology and offers a principled route to tailor stimulation to each individual's deviation pattern. The same personalization principle might be examined beyond the DLPFC at symptom-relevant accessible cortical regions, such as the temporoparietal junction (Hoffman et al., 1999; Hoffman et al., 2000; Hoffman et al., 2003), but this possibility requires direct testing. The computational sequence might also be evaluated in other psychiatric disorders characterized by distributed network abnormalities, but this remains a hypothesis. In major depressive disorder, individualized resting-state FC has been used to select left-DLPFC sites according to their anticorrelation with the subgenual anterior cingulate cortex in a randomized TMS trial (Cole et al., 2022). In obsessive-compulsive disorder, a randomized deep-TMS trial targeted the medial prefrontal and anterior cingulate cortices on the basis of a cortical-striatal-thalamic-cortical circuit rationale (Carmi et al., 2019). These examples support the feasibility of diagnosis-specific circuit-informed targeting, but they do not establish that the schizophrenia epicenter framework, target, or clinical effect generalizes to those disorders. Any extension would require a disorder-specific deviation phenotype, candidate pathological anchors, accessible cortical search space, stimulation protocol, and outcome measure, followed by prospective validation. The present results therefore define a testable procedure for schizophrenia rather than a disorder-independent target. Limitations This study has several limitations that should be acknowledged. First, although we localized stimulation targets using group-level DLPFC coordinates derived from prior studies, personalized real coil locations were not available. This reliance on approximated coordinates may introduce bias relative to true personalized targeting. Second, our verification of expected FC relied on a relatively simple regression framework; more sophisticated machine learning approaches, such as deep neural networks, could provide a more flexible means to model nonlinear structure–function relationships. In addition, we did not explicitly model the individual electric field induced by stimulation; future prospective studies combining neuronavigation with electric-field modeling will be important for testing whether target engagement can be further improved. Finally, we did not perform direct validation of epicenter-guided targeting in the current study. Future work incorporating real epicenter-informed rTMS protocols will be needed to establish causal efficacy. A prospective trial is needed to test the framework directly. Patients should be randomized to neuronavigated framework-derived targeting or conventional F3-based targeting while stimulation frequency, intensity, pulse number, treatment schedule, concomitant medication requirements, and assessment timing are matched. Coil position and orientation should be recorded at every session, clinical raters should be blinded where feasible, and clinical and imaging outcomes should be preregistered. Such a design would distinguish target-delivery fidelity from clinical efficacy. Conclusion In this study, we developed an epicenter-guided computational framework for personalized rTMS targeting in schizophrenia using multimodal neuroimaging in a large discovery cohort. The framework integrates individual GMV-deviation patterns with normative FC to identify patient-specific candidate pathological epicenters and uses native-space FC to derive candidate coordinates within the DLPFC. In the discovery cohort, personalized coordinates showed higher spontaneous brain activity contrast estimates than the F3-associated reference, indicating stronger differentiation between the predefined PCC amplitude states at those coordinates. This contrast characterizes one predefined intrinsic-network context rather than a TMS-evoked response. In the retrospective treatment cohort, smaller native-space distance between the framework-derived coordinate and the canonical F3-associated reference was associated with stronger concordance between simulated and observed FC changes and with greater improvement in negative symptoms. Together, these findings provide a functional characterization of the derived coordinates within this predefined resting-state context and motivate prospective testing of their therapeutic value. CRediT authorship contribution statement Yingjie Tang: Writing – review & editing, Writing – original draft, Software, Methodology, Formal analysis. Sisi Jiang: Software, Methodology, Conceptualization. Huan Huang: Validation, Software, Methodology. Xueqing Sun: Visualization, Validation, Software, Data curation. Chao Mu: Project administration, Investigation. Mingjun Duan: Project administration, Investigation. Roberto Rodríguez-Labrada: Supervision, Resources, Project administration. Dezhong Yao: Resources, Project administration, Investigation, Conceptualization. Cheng Luo: Writing – review & editing, Resources, Funding acquisition, Conceptualization. Funding This work was partly funded by grants from the 10.13039/501100012166National Key Research and Development Program of China (2024YFE0215100), the 10.13039/501100001809National Natural Science Foundation of China (grant number: 82371560,62401124, 62571106), the CAMS Innovation Fund for Medical Sciences (CIFMS) (No. 2019-I2M-5-039), and the Natural Science Foundation of Sichuan (2024YFFK0362, 2023NSFSC0037), and the Sichuan Provincial Program of Traditional Chinese Medicine(2021ZD017). Declaration of competing interest The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.