Associations between neuronavigated rTMS‐mediated functional connectome plasticity and neurotransmitter receptors in Alzheimer's disease

Abstract

INTRODUCTION

Cognitive improvement is mediated by repetitive transcranial magnetic stimulation (rTMS) in Alzheimer's disease (AD), and connectome gradient analysis has important potential to explore more comprehensive information for treatment efficacy. However, it remains unclear whether and how rTMS mediates connectome gradient reconstruction and its spatial neurotransmitter associations in subjects on the AD spectrum.

METHODS

A total of 84 subjects on the AD spectrum underwent 4 weeks of conventional neuronavigated rTMS targeting the left angular gyrus (20 Hz, 20 sessions over 4 weeks, 20 minutes per session), including 60 subjects with true stimulation and 24 with sham stimulation. This study identified functional connectome gradients and investigated their neuroplasticity between pre‐ and post‐rTMS intervention. A support vector regression model was subsequently used to explore the cognitive prediction value of baseline measures. Finally, connectome‒neurotransmitter association analysis was used to investigate neurotransmitter profiles related to rTMS therapeutic efficacy.

RESULTS

These findings showed that rTMS treatment mainly mediated the decreased gradient values in the somatomotor regions and such baseline values of these regions showed preliminary predictive value for memory improvement after treatment (R ²: 0.40–0.63, mean absolute error: 1.54–3.19, root mean square error: 1.98–3.94, p = 0.0002). Moreover, connectome–neurotransmitter association analysis suggested spatial associations between these gradient changes and normative metabotropic glutamate receptor 5 (mGluR5) and 5‐hydroxytryptamine receptor 2A (5HT2a) receptor distributions.

DISCUSSION

rTMS mediates plasticity in high‐order metric levels in the brain and is linked to the spatial distributions of specific neurotransmitter receptors in subjects on the AD spectrum.

Highlights

• This study suggests that repetitive transcranial magnetic stimulation (rTMS) mediates changes in the macroscale organization of high‐order metric levels in the brain and is linked to the spatial distributions of specific neurotransmitter receptors in subjects on the Alzheimer's disease (AD) spectrum.

• rTMS mediated the decreased gradient values in the somatomotor regions, and such baseline measures can excellently predict memory improvement after treatment in patients with AD.

• Connectome–neurotransmitter association analysis revealed a spatial correspondence between the reconstructed gradient of the somatomotor regions and neurotransmitter profiles enriched in metabotropic glutamate receptor 5 (mGluR5) and 5‐hydroxytryptamine receptor 2A (5HT2a) receptors in subjects on the AD spectrum.

1. INTRODUCTION

Effectively improving cognitive impairment has been a key focus and major challenge in the field of Alzheimer's disease (AD), and repetitive transcranial magnetic stimulation (rTMS) has been explored as a non‐invasive brain stimulation technique for the treatment of AD. ¹ , ² , ³ Its potential mechanism involves multilevel neural regulation and functional reorganization, including rTMS‐induced changes in neuronal excitability, synaptic transmission and plasticity. ⁴

rTMS treatment can reconstruct the integrity of the memory network, ⁵ default mode network (DMN), ⁶ and multinetwork connectivity ⁷ by regulating functional connections between target brain regions and other brain regions in patients with AD. Increasing evidence has suggested that actual brain functional activity is in a state of high‐order metric level, that each brain region and network does not exist in isolation, and that the relationships among brain regions in a network and their impact on the overall network should be considered. ⁸ , ⁹ The strategy for high‐order brain network analysis can compensate for the shortcomings of traditional neuroimaging methods. Importantly, a novel modeling framework for investigating brain high‐order metric‐level architecture from connectome gradient analysis has been developed. ¹⁰ , ¹¹ Gradient analysis uses dimensionality reduction techniques to reveal the continuous changes in brain functional organization from a global perspective, capturing the functional connectivity patterns between brain regions and their continuous transitions in space. A previous study revealed that patients with AD exhibited global connectome gradient alterations compared to normal controls. ¹² However, there are currently no studies on rTMS‐mediated connectome gradient architecture changes in patients with AD.

The regulation of excitatory and inhibitory neurotransmitter balance has gradually received increasing attention with respect to the therapeutic efficacy of rTMS. ¹³ , ¹⁴ The underlying mechanism revealed that neurotransmitter receptors drive synaptic plasticity, modify neural states, and ultimately shape network‐wide communication, suggesting that changes in neurotransmitters play important roles in promoting neural plasticity and brain network remodeling. ¹⁴ Therefore, connectome‒neurotransmitter association studies have bridged the gap between the microlevel transcriptome profile and the macroscale brain network. ¹⁵ Although there are currently few studies in the field of AD, many valuable discoveries in other disease areas have been reported. ¹⁶ , ¹⁷ Therefore, we postulated that neurotransmitters also play an important role in rTMS‐mediated connectome gradient changes in patients with AD. ¹⁸

This study aimed to investigate the rTMS‐mediated neuroplasticity of functional connectome gradients in subjects on the AD spectrum and their associations with cognitive improvements and neurotransmitter profiles. Here, the neurotransmitter data were derived from a previously published whole‐brain normative atlas based on aggregated positron emission tomography (PET) data, which includes 19 receptors across nine neurotransmitter systems. These maps represent standardized spatial distributions derived from independent datasets. ¹⁹ This study outlines the analytical workflow shown in Figure 1.

FIGURE 1.

This study outlines the analytical workflow. Four groups of participants were included (pre‐ and post‐ true stimulation and pre‐ and post‐ sham stimulation). The brain network gradients were first extracted, followed by interaction effect analysis. SVR was subsequently used to evaluate the ability of rTMS treatment to predict cognitive improvement. Moreover, the results were associated with neurotransmitter maps to explore the underlying molecular mechanisms. FC, functional connectivity; fMRI, functional magnetic resonance imaging; rTMS, repetitive transcranial magnetic stimulation; SVR, support vector regression.

RESEARCH IN CONTEXT

1. Systematic review: A strategy for high‐order brain analysis can compensate for the shortcomings of traditional neuroimaging methods, and neurotransmitters play an intermediate role in promoting neural plasticity and brain network remodeling. However, it remains unclear whether and how repetitive transcranial magnetic stimulation (rTMS) mediates connectome gradient reconstruction and its spatial neurotransmitter associations in subjects on the Alzheimer's disease (AD) spectrum.

2. Interpretation: A total of 84 participants at risk for AD were recruited and subjected to 4 weeks of rTMS true or sham stimulation. This study identified functional connectome gradients and investigated their neuroplasticity before and after rTMS intervention. A support vector regression model was subsequently used to explore cognitive prediction on the basis of rTMS‐mediated gradient changes. Finally, connectome–neurotransmitter spatial association analysis was used to investigate neurotransmitter profiles related to rTMS therapeutic efficacy. These findings showed that rTMS treatment mainly mediated the decreased gradient values in the somatomotor regions and that such baseline values of these regions can predict memory improvement efficacy after treatment. Moreover, a spatial association between the gradient reconstruction of somatomotor regions and the spatial distributions of neurotransmitter receptors enriched in metabotropic glutamate receptor 5 (mGluR5) and 5‐hydroxytryptamine receptor 2A (5HT2a) receptors was found in subjects on the AD spectrum.

3. Future directions: Future studies with larger independent cohorts, the integration of AD‐related biomarkers including molecular imaging and cerebrospinal fluid measures of amyloid beta and tau pathology, and subject‐specific electric‐field modeling are needed to further validate and extend these findings.

2. MATERIALS AND METHODS

2.1. Participant characteristics

The study protocol was approved by the ethics committee of Nanjing Drum Tower Hospital. Prior to enrollment, all potential participants underwent an initial clinical screening to exclude individuals who did not meet the eligibility criteria. A total of 84 participants on the AD spectrum were recruited and subjected to 4 weeks of conventional 20 Hz high‐frequency neuronavigated rTMS true or sham stimulation. After enrollment, participants were randomly assigned to either the true stimulation group (n = 60) or the sham stimulation group (n = 24) according to a predefined randomization schedule. No significant differences in sex, age, or years of education were observed between the two groups at baseline. The allocation was implemented by research staff. Clinicians responsible for neuropsychological assessments were blinded to group assignment, whereas the operators delivering rTMS were aware of the stimulation condition due to the nature of the procedure. Patients included in the study were diagnosed with either mild cognitive impairment (MCI) or AD, were aged between 55 and 80 years, and were right handed. The details of diagnostic criteria are shown in Section 1 in supporting information.

2.2. Neuronavigated rTMS procedure and neuropsychological assessment

All participants received rTMS treatment via a magnetic stimulator (CCY‐IV model; Yirid, Inc., Wuhan, China) equipped with a 70‐mm figure‐eight coil. The stimulation target was localized to the left angular gyrus (Montreal Neurological Institute coordinates: −45, −67, 38) on the basis of our previous seed‐based functional connectivity study in an AD spectrum cohort. rTMS cannot directly stimulate the hippocampus, and we calculated this area on the basis of an independent sample of patients with cognitive impairment and found that there was a strong positive functional connection between this point and the hippocampus. The details of rTMS procedure are shown in Section 2 in supporting information. The main stimulation parameters were as follows: each session consisted of 1600 pulses delivered in 40 trains of 40 pulses each. The stimulation was delivered at a frequency of 20 Hz, with each train lasting 2 seconds. The conventional 20 Hz high‐frequency rTMS protocol included a total of 20 sessions administered over a 4‐week period (five sessions per week, Monday through Friday), with each session lasting 20 minutes. An interval of 28 seconds was applied to prevent excessive cortical excitability, and the stimulation intensity was set at 100% of the resting motor threshold. In the sham stimulation group, a specially designed insulating panel was placed between the coil and the scalp at the target site to block magnetic field transmission and eliminate effective cortical stimulation. All participants successfully completed a comprehensive neuropsychological assessment to ensure the reliability and validity of the cognitive measurements (Table 1).

TABLE 1.

Participant demographics and neuropsychological assessments at baseline and follow‐up.

	rTMS true		rTMS sham		rTMS true	rTMS sham	mean ± SD
Items	Pre (n = 60)	Post (n = 60)	Pre (n = 24)	Post (n = 24)	Pa value (pre vs. post)	Pb ‐value (pre vs. post)	ΔrTMS true (post‐pre)	ΔrTMS sham (post‐pre)
Demographics
Sex (male/female)	(21 / 39)	‐	(12 / 12)	‐	‐	‐	‐	‐
Age (years)	66.37 ± 6.43	‐	65.58 ± 8.29	‐	‐	‐	‐	‐
Education (years)	11.68 ± 3.27	‐	11.83 ± 3.66	‐	‐	‐	‐	‐
MMSE	26.70 ± 3.67	27.67 ± 2.94	27.08 ± 3.30	27.63 ± 2.86	<0.001 ***	0.196	0.97 ± 1.99	0.54 ± 1.84
MoCA	22.35 ± 4.29	24.40 ± 4.22	23.21 ± 4.61	22.46 ± 4.26	<0.001 ***	0.054	2.05 ± 2.55	−0.75 ± 2.92
Memory assessments
AVLT‐SR	4.50 ± 3.23	5.98 ± 3.18	4.54 ± 3.28	5.70 ± 2.93	<0.001 ***	0.019 *	1.48 ± 2.24	0.92 ± 2.50
AVLT‐LR	3.93 ± 2.84	5.55 ± 3.29	4.21 ± 2.90	5.04 ± 3.10	<0.001 ***	0.066	1.62 ± 1.88	0.83 ± 2.06
AVLT‐IR	15.23 ± 5.02	18.00 ± 5.62	14.71 ± 5.15	16.71 ± 4.79	<0.001 ***	0.002 **	2.77 ± 3.92	2.00 ± 2.84
AVLT‐R	19.48 ± 3.28	20.02 ± 3.48	19.38 ± 3.35	19.50 ± 3.72	0.108	0.751	0.53 ± 2.41	0.13 ± 2.94
VR‐DR	6.25 ± 4.36	8.03 ± 4.43	6.71 ± 4.41	8.33 ± 4.29	<0.001 ***	0.017 *	1.78 ± 2.92	1.63 ± 2.89
Executive function
Stroop‐C (raw score)	38.79 ± 17.40	35.04 ± 15.15	49.50 ± 34.10	41.54 ± 29.72	0.044 *	0.013 *	−3.75 ± 17.21	−7.96 ± 2.92
TMT‐B (raw score)	143.70 ± 69.62	123.68 ± 95.39	169.63 ± 120.40	137.68 ± 82.63	<0.001 ***	0.134	−20.02 ± 60.60	−31.95 ± 99.7
DST‐back (raw score)	4.37 ± 1.37	4.60 ± 1.38	5.46 ± 2.30	5.46 ± 2.23	0.104	0.974	0.23 ± 1.20	0.00 ± 1.25
Processing speed
TMT‐A (raw score)	71.89 ± 31.48	67.77 ± 35.47	77.59 ± 45.97	65.79 ± 27.63	0.030 *	0.092	−4.12 ± 33.6	−11.80 ± 32.57
Stroop‐A (raw score)	23.75 ± 14.80	22.52 ± 18.32	26.59 ± 11.80	25.21 ± 15.34	0.109	0.188	−1.23 ± 7.97	−1.39 ± 9.95
Stroop‐B (raw score)	25.79 ± 10.47	23.01 ± 9.30	31.18 ± 22.42	33.08 ± 33.58	<0.001 ***	0.782	−2.78 ± 7.35	1.90 ± 15.21
Language function
BNT	49.90 ± 7.81	51.48 ± 7.11	50.33 ± 7.11	51.13 ± 6.98	<0.001 ***	0.207	1.58 ± 2.56	0.79 ± 3.13
CVF	15.90 ± 4.53	16.62 ± 4.70	15.67 ± 4.45	15.25 ± 5.32	0.178	0.701	0.72 ± 4.34	−0.42 ± 5.18
Visuospatial functional
VR‐C	13.07 ± 2.12	13.23 ± 2.13	13.63 ± 1.13	13.92 ± 0.41	0.166	0.180	0.17 ± 0.89	0.29 ± 1.08

Notes: Data were expressed as mean ± SD; comparisons within groups (pre‐ vs. post‐intervention) were analyzed using the Wilcoxon signed‐rank test (denoted as P ^a for true‐rTMS and for P ^b sham‐rTMS).

Abbreviations: AVLT‐IR, Auditory Verbal Learning Test immediate recall; AVLT‐LR, Auditory Verbal Learning Test long recall (3–5 minutes vs. 20‐minute delay), AVLT‐R, Auditory Verbal Learning Test recognition; AVLT‐SR, Auditory Verbal Learning Test short recall; BNT, Boston Naming Test; CVF, Category Verbal Fluency; DST‐back, Digit Span Test backward; MMSE, Mini‐Mental State Examination; MoCA, Montreal Cognitive Assessment; rTMS, repetitive transcranial magnetic stimulation; SD, standard deviation; Stroop‐A, B, and C, Stroop Color and Word Test Parts A, B, and C; TMT‐A and ‐B, Trail Making Test Parts A and B; VR‐C, Visual Reproduction copy; VR‐DR, Wechsler Memory Scale Visual Reproduction delayed recall.

* p < 0.05, ** p < 0.01, *** p < 0.001.

2.3. Neuroimaging data acquisition and preprocessing

Imaging data were acquired via a 3.0T Philips scanner. The detailed scanning parameters and structural preprocessing steps are provided in Sections 3 and 4 in supporting information. The detailed functional magnetic resonance imaging (fMRI) preprocessing procedures (such as slice‐timing correction, motion correction, spatial normalization, and smoothing) have been described in our previous study. ²⁰ Both T1 and fMRI data were collected from all participants at both baseline and 4 weeks after the intervention.

2.4. Macroscale connectome gradient analysis

To capture the continuous macroscale organization of functional connectivity in the cerebral cortex, we performed gradient analysis via the BrainSpace toolbox (https://brainspace.readthedocs.io/), which is widely used for connectome‐based gradient mapping. The functional connectivity network was constructed based on the Schaefer 400‐parcel atlas, ²¹ which divides the cortical surface into 400 functionally homogeneous regions and annotates them according to the seven large‐scale functional systems. ²² Functional connectivity matrices were constructed for each participant, and dimensionality reduction was applied to derive principal gradients. ²³ , ²⁴ , ²⁵ To enable group‐level comparisons, individual gradients were aligned to a common template via Procrustes alignment, ²⁶ and summary metrics (i.e., explained ratio, variance, and range) were computed. The full processing details of the connectome gradient analysis are provided in Section 5 in supporting information.

After individual gradient maps were constructed and spatial alignment was performed, we conducted regionwise statistical analyses to further explore the effects of rTMS on local functional topography. Specifically, on the basis of the Schaefer 400 atlas, we extracted gradient values (Gradient 1, Gradient 2, and Gradient 3) from each of the 400 cortical parcels for each participant, which were used for subsequent comparisons. An analysis of variance (ANOVA) was further performed, with a within‐subject repeated factor (i.e., group: pre‐ and post‐rTMS) and an across‐subject factor (i.e., rTMS: true intervention and sham intervention), with the subject as a nested random variable within the group. We used the afex package in R to construct ANOVA models to evaluate the interaction effects in each brain region. To control for false positives caused by multiple comparisons, we applied false discovery rate (FDR) correction, using a threshold of FDR‐adjusted p < 0.05 to determine statistical significance. For regions showing significant effects, we further performed post hoc analyses via the emmeans (estimated marginal means) package. All post hoc comparisons were adjusted for multiple testing via the Bonferroni correction, with a Bonferroni‐adjusted p < 0.05 considered statistically significant. Furthermore, motivated by the observed groups × rTMS interaction effects in Gradient 1, we aimed to examine whether baseline gradient values could predict post‐intervention cognitive performance using support vector machine analysis (Section 6 in supporting information).

2.5. Connectome–neurotransmitter association analysis

To further investigate whether rTMS‐mediated gradient changes are spatially associated with the distribution patterns of specific neurotransmitter systems, we conducted partial least squares (PLS) regression analysis via Gradient 1 interaction effects and neurotransmitter maps. Specifically, we first calculated the difference in gradient values (“post‐intervention minus pre‐intervention”) for each participant in each brain region and then performed independent‐sample t tests between groups based on these difference scores, resulting in t values for all 400 Schaefer parcels. These t values were used as statistical representations of the Gradient 1 interaction effect and served as the dependent variable (Y vector) in the PLS analysis. For extended analysis, the same procedure was also applied to Gradient 2 and Gradient 3 to extract interaction‐effect t values and conduct separate PLS models.

The independent variables (X) consisted of spatial distribution maps of 19 neurotransmitters across the Schaefer 400 parcellation. These X and Y matrices were then entered into the PLS model to extract latent components that explain variance in the gradient interaction effects, with the number of components set to 15. To assess whether the explained variance of each PLS component exceeded random expectations and to account for spatial autocorrelation inherent in cortical maps, we conducted statistical inference via spin‐based permutation testing, a spherical rotation procedure that preserves the spatial structure of the brain surface.

In addition, we further conducted a statistical evaluation of the following metrics: (1) X scores: the coordinates of the independent variable X projected onto the latent component space, representing the distribution of samples across the latent variables. We computed the correlation between the X scores and the t values and assessed statistical significance via a spin test to obtain p values. ²⁷ (2) X loadings: the linear combination weights of each neurotransmitter in the PLS latent components, indicating the degree to which each neurotransmitter is related to the interaction effect. We also applied spin test permutation analysis to evaluate the statistical significance of these loadings. All significance thresholds were set at a spin test p < 0.05.

3. RESULTS

3.1. Demographic characteristics and clinical outcomes

At baseline, no significant differences were found in terms of sex, age, or education level between the true pre and sham pre groups (P _sex = 0.206, P _age = 0.886, P _education = 0.868), and these two groups presented similar cognitive impairments in various cognitive domains (Table 1). However, we found more significant longitudinal improvements in the true stimulation group than in the baseline group after 4 weeks of rTMS treatment, including improvements in general cognition (i.e., Montreal Cognitive Assessment and Mini‐Mental State Examination), memory function (i.e., Auditory Verbal Learning Test [AVLT]‐SR [short recall], AVLT‐LR [long recall], AVLT‐IR [immediate recall], and Wechsler Memory Scale Visual Reproduction delayed recall [VR‐DR]), executive function (i.e., Stroop Color and Word Test Part C and Trail‐Making Test Part B), processing speed (i.e., Trail‐Making Test Part and Stroop Color and Word Test Part B), and language function (i.e., Boston Naming Test)(p < 0.05). The sham stimulation group only showed a few sporadic changes in the scales, which may be due to placebo‐like responses. The rTMS‐mediated improvement in memory dysfunction is a key focus of subsequent predictive analysis. These details are shown in Table 1.

3.2. Functional connectome gradient patterns

Gradient analysis was performed to test functional connectome gradient changes between pre‐ and post‐rTMS treatment in subjects on the AD spectrum. The principal gradient (i.e., Gradient 1) showed the highest eigenvector; therefore, the focus of this study was on the principal gradient, and other gradients (i.e., Gradient 2 and Gradient 3) can be found in supporting information (Figure S1 and S2). The principal gradient explained ratios were 16.5%, 16.4%, 16.2%, and 16.5% of the whole‐brain connectome variance in the true_pre, true_post, sham_pre, and sham_post groups, respectively (Figure 2). The spatial patterns of the four group‐averaged principal gradient maps were similar (Figure 3). In line with previous studies, ⁵ , ¹⁰ the principal gradient increased organizationally along a gradual axis from the sensory systems (i.e., visual network and somatomotor network) to the association network (i.e., DMN; Figure 2 and Figure 3).

FIGURE 2.

Functional connectivity gradient mapping results of the four participant groups. (A) Scatter plot of Gradient 1 (i.e., principal gradient) versus Gradient 2 for the true_pre group. Gradient 1 extended between the sensory systems (visual network and somatomotor network, blue and red) and the association network (default mode network, green). The top right inset displays the eigenvalue spectrum obtained from the diffusion embedding algorithm. The explanation ratio of connectome‐level variance is explained by the top 15 components. The bottom right shows the cortical surface visualization of the gradient distribution. (B‐D) Corresponding gradient mapping results for the true_post, sham_pre, and sham_post groups, respectively.

FIGURE 3.

Spatial distribution of the principal gradient across the four participant groups. (A) The top panel shows the spatial distribution of Gradient 1 on the cortical surface, with colors representing the gradient score of each brain region along the gradient axis. The middle panel presents a histogram of the overall distribution of gradient scores across all regions. The bottom panel displays the network‐level segmentation of Gradient 1, which is derived via a community‐based z score analysis of gradient scores. Gradient 1 increased organizationally along a gradual axis from the sensory systems (i.e., the VIN and SMN) to the association network (i.e., the DMN). (B‐D) illustrate the same analysis pipeline and results for the true_post, sham_pre, and sham_post groups, respectively. Yeo 7 networks are used to map the Schaefer 400 regions of interest. DAN, dorsal attention network; DMN, default mode network; FPCN, frontal‒parietal control network; LIM, limbic region; SMN, somatomotor network; VAN, ventral attention network; VIN, visual network.

3.3. rTMS treatment modulates gradient reconstruction

In this study, the Yeo 7 networks were constructed via the Schaefer 400 regions of interest. A groups × rTMS ANOVA was further performed, and the gradients at the global level and local level were statistically compared. (1) Gradient global index of the Yeo 7 networks: Among the four groups, ANOVA and post hoc comparisons were used to show the changes in the gradient range, variability and explanation ratio. However, no differences in these parameters were found between the groups and pre‐post the statistical comparisons (p > 0.05). (2) Gradient local index of Schaefer 400 regions: the patterns of groups × rTMS ANOVA interactions and the distributions of brain regions, including SomMotA_30 and SomMotA_31 (FDR corrected, p < 0.05; Figure 4A and Table S1 in supporting information), were confirmed. Moreover, post hoc analysis revealed decreased gradient values in these two somatomotor regions between the true_post and true_pre rTMS groups, whereas the sham stimulation group was associated with the opposite trend (Bonferroni corrected, p < 0.05; Figure 4A). These findings revealed that rTMS preferentially mediated somatomotor region gradient reconstruction of the principal gradient of brain function in AD spectrum subjects.

FIGURE 4.

Brain regions with rTMS‐mediated gradient changes and corresponding predictive models. (A) Brain regions showing significant group × rTMS interaction effects in Gradient 1 after FDR correction are displayed (i.e., SomMotA_30 and SomMotA_31). Post hoc analysis revealed decreased gradient values in these two somatomotor regions between the true_post and true_pre rTMS groups, whereas the sham stimulation group was associated with the opposite trend (Bonferroni corrected, p < 0.05). (B) On the basis of the interaction effect statistics, the combination of the baseline gradient values of the somatomotor regions and the baseline memory scale, a machine learning method constructed using the SVR model was used to predict memory improvement after rTMS treatment. The results of the SVR model are presented for the following behavioral outcomes: AVLT_LR, AVLT_IR, AVLT_SR, AVLT_R, and VR_DR. In addition, to evaluate the statistical significance of the model under the empirical null distribution, we performed 5000 random label permutations and recorded the model performance metric for each permutation. The resulting 5000 metric values were then plotted as a frequency distribution histogram. AVLT‐IR, Auditory Verbal Learning Test immediate recall; AVLT‐LR, Auditory Verbal Learning Test long recall; AVLT‐R, Auditory Verbal Learning Test recognition; AVLT‐SR, Auditory Verbal Learning Test short recall; FDR, false discovery rate; SVR, support vector regression; rTMS, repetitive transcranial magnetic stimulation; VR‐DR, Wechsler Memory Scale Visual Reproduction delayed recall.

3.4. Associations between rTMS‐mediated gradient changes and cognitive prediction

To better understand the clinical significance of the gradient values, we further explored this association between rTMS‐mediated gradient changes and cognitive prediction in subjects on the AD spectrum. Therefore, the combination of the baseline gradient values of somatomotor regions and the baseline neuropsychological scale and a machine learning method constructed using the support vector regression (SVR) model were used to predict the changes in this scale after rTMS treatment. We found that these rTMS‐mediated improvements in memory function can be used to obtain accurate predictions in the true stimulation group. Specifically, the outcomes of AVLT_LR (R²  = 0.6330, mean absolute error [MAE] = 1.5398, root mean square error [RMSE] = 1.9735, permutation test p = 0.0002), AVLT_IR (R²  = 0.5021, MAE = 3.1895, RMSE = 3.9352, permutation test p = 0.0002), AVLT_SR (R²  = 0.5630, MAE = 1.6429, RMSE = 2.0851, permutation test p = 0.0002), AVLT_R (R²  = 0.4008, MAE = 2.1381, RMSE = 2.6685, permutation test p = 0.0002), and VR_DR (R²  = 0.5601, MAE = 2.3415, RMSE = 2.9135, permutation test p = 0.0002) were detected in this study (Figure 4B). R ² can measure the explanatory power of the model for the target, for example, values > 0.5 indicate strong prediction; MAE/RMSE represents the difference between the predicted and actual values. These findings suggest that gradient values may have predictive value for therapeutic efficacy, with baseline gradient values in somatomotor regions potentially associated with memory improvement after rTMS treatment in subjects on the AD spectrum.

3.5. Associations between rTMS‐mediated gradient changes and neurotransmitter profiles

To explore the spatial associations between rTMS‐mediated gradient changes and the 19 receptors across nine different neurotransmitter systems, PLS regression analysis was conducted across all true stimulation of subjects on the AD spectrum. The proportions of variance explained by the PLS component of the principal gradient (i.e., Gradient 1) and the PLS1 of the principal gradient loading on various neurotransmitters are shown in Table S2 in supporting information. Figure 5 depicts the detailed spatial correspondence between rTMS‐mediated gradient changes and normative neurotransmitter systems. First, rTMS‒intervention gradient interaction effects were positively related to neurotransmitter maps (r = 0.309, spin test p = 0.035). Second, the pre–post changes in the gradient values of the somatomotor regions were significantly correlated with the spatial distributions of two neurotransmitter receptors, namely, metabotropic glutamate receptor 5 (mGluR5) receptors (loading = 18.2825, spin test p = 0.0319) and 5‐hydroxytryptamine receptor 2A (5HT2a) receptors (loading = 15.4082, spin test p = 0.041). These other spatial correlation results are summarized in Figures S3 and S4 in supporting information.

FIGURE 5.

PLS analysis of the spatial associations between neurotransmitter maps and rTMS‐mediated gradient changes. (A) Explained variance of the top 15 components obtained from the PLS model. The boxplots represent the null distribution generated from spatial spin permutation testing. (B) Variance explanation curve with the spin test–derived null distribution removed, highlighting the actual variance explained by each component. (C) From left to right: cortical distribution of PLS1 X scores; brain map of T statistics for the Gradient 1 groups × rTMS interaction effect; and spatial correlation map between PLS1 X scores and T values across the cortex. (D) Loadings of the 19 receptors and transporters across nine different neurotransmitter systems on the first latent component (PLS1), reflecting each neurotransmitter's weight in the latent variable. Statistical significance was assessed via spin permutation testing, with a threshold of p < 0.05 for the spin test. PLS, partial least squares; rTMS, repetitive transcranial magnetic stimulation.

4. DISCUSSION

This study suggests that rTMS mediates changes in the macroscale organization of high‐order metric levels in the brain and is linked to the spatial distributions of specific neurotransmitter receptors in subjects on the AD spectrum. A connectome–neurotransmitter association analysis revealed a spatial association between the reconstructed gradient of the somatomotor regions and neurotransmitter profiles enriched in mGluR5 and 5HT2a receptors.

Previous traditional neuroimaging has shown that the characteristic patterns of somatomotor region disruption were observed both in AD cross‐sections ²⁸ and in longitudinal studies. ²⁹ Although the stimulation target in this study was the left angular gyrus, the significant gradient changes were primarily observed in somatomotor regions. This spatial pattern may reflect the network‐level propagation effects of rTMS stimulation, as the angular gyrus is a hub within the parietal association cortex that is functionally connected with multiple large‐scale brain systems. From the perspective of connectome gradient organization, somatomotor regions may exhibit greater sensitivity to nerveregulation perturbations. The gradient features in this study were consistent with those of previous high‐order metric‐level studies, ³⁰ , ³¹ revealing that the pattern of the primary gradient showed a continuous transition from low‐level visual and somatomotor function to high‐level cognitive function and was the core dimension of the hierarchical organization of brain functions. The principal connectome gradient reflects a hierarchical axis extending from primary sensory and somatomotor regions to higher order association networks. Within this framework, decreased gradient values may indicate reduced functional distance between lower level and higher order systems, suggesting increased network integration. Such large‐scale network reorganization may contribute to improved communication between memory‐related regions and distributed cortical systems after rTMS. The somatomotor regions of the primary gradient changed after rTMS treatment in the present AD spectrum subjects, whereas other networks did not significantly change. It is necessary to discuss the functional characteristics of the somatomotor regions: (1) First, the somatomotor regions are responsible for processing low‐level perceptual motor functions and are among the basic functional characteristics. The somatomotor regions have high structural plasticity, ³² and they also have strong functional connections with other networks (i.e., the DMN) at the functional level. ³³ Therefore, these regions respond quickly to external rTMS stimuli and undergo functional reorganization. (2) Second, rTMS mainly stimulates neurons in the vicinity of the target area (i.e., the left angular gyrus, which is relatively close to the somatomotor regions) through magnetic field‐induced currents. ³⁴ (3) Finally, the distribution of amyloid and tau proteins in the brains of patients with AD is uneven, with the DMN usually being more severely affected, whereas the somatomotor regions are less affected. ³⁵ Therefore, rTMS may be more effective for somatomotor regions with lower pathological burdens but has limited effectiveness for the DMN, which has a greater pathological burden.

Importantly, prognostic modeling is used to predict disease progression through longitudinal data analysis and risk prediction models. ³⁶ Therefore, the fact that rTMS not only improves cognitive ability but also baseline gradients can help predict who benefits. Our SVR analysis suggests that the baseline gradient values of the somatomotor regions can be used to predict rTMS treatment efficacy in subjects on the AD spectrum, especially for predicting improvements in memory impairment. Notably, the baseline gradient maps of patients significantly predict symptomatic improvement after treatment, which has also been observed in patients with major depressive disorder . ³⁷ This predictive ability may help identify patients more likely to benefit from rTMS treatment. These findings support the use of connectome gradient reconstruction in subjects on the AD spectrum and its linkage with clinical management, highlighting the possibility of further deepening and expanding gradient application studies in the future. However, given the relatively modest sample size, the predictive results should be interpreted with caution, as the possibility of model instability or overfitting cannot be fully excluded despite the use of cross‐validation and permutation testing. Future studies with larger and independent cohorts are needed to further validate these findings.

The previous studies of aging and major depressive disorder have shown that neuronal activity and cortical thickness in the human cerebral cortex are regulated by neurotransmitters and that there is a close correlation between cortical functional abnormalities and corresponding neurotransmitter system expression levels in different diseases. ³⁸ , ³⁹ , ⁴⁰ , ⁴¹ In this study, the connectome–neurotransmitter spatial association analysis revealed that rTMS‐mediated decreases in somatomotor region gradient values in subjects on the AD spectrum were significantly positively correlated with the spatial profiles of mGluR5 receptors and 5HT2a receptors. Both mGluR5 and 5HT2a are involved in synaptic plasticity, interactions with amyloid beta (Aβ) and tau pathology, and the regulation of neuroinflammation in AD, ⁴² , ⁴³ and the somatomotor regions of the brain widely express these two neurotransmitters. ¹⁹ The differences between the two in terms of AD mechanisms include that mGluR5 mainly participates in glutamate signaling, whereas 5HT2a receptors are involved mainly in serotonin signaling; mGluR5 mediates neurotoxicity through direct binding to Aβ, whereas 5HT2a receptors disrupt neuronal excitability through Aβ interference. ⁴⁴ , ⁴⁵ These similarities and differences reveal the complexity and diversity of mGluR5 and 5HT2a receptors in the pathological mechanism of AD. We hypothesized that rTMS‐related changes may involve neurotransmitter levels associated with mGluR5 and 5HT2a receptors, thereby promoting remodeling of the somatomotor region gradient and participating in the overall neural circuit regulating memory function.

In summary, this study revealed the neuroimaging mechanism by which rTMS regulates the homeostasis of higher order brain networks in subjects on the AD spectrum. In addition, we also found that the therapeutic effect of rTMS may be spatially associated with specific neurotransmitter receptors. Although several issues need to be acknowledged (see Section 7 in supporting information), these findings provide new insight into the network reorganization associated with rTMS treatment in subjects on the AD spectrum.

AUTHOR CONTRIBUTIONS

W.Y. and F.B. contributed to the conception and design of the study. W.Y., T.L., W.Z., and F.B. performed the analyses, obtained the findings, and provided guidance on the interpretation of the results. W.Y., T.L., W.Z., Y.L., X.S., Y.Z., S.Y., and B.Y. were involved in the data collection. W.Y. and F.B. wrote the manuscript and supervised the study. Artificial intelligence–assisted tools were used solely for language editing and not for generating academic content. All core data, research results, and scientific conclusions presented in this manuscript were generated and verified by the authors.

CONFLICT OF INTEREST STATEMENT

The authors declare no conflicts of interest. Author disclosures are available in the Supporting Information.

CONSENT STATEMENT

This study was approved by the ethics committees of Nanjing Drum Tower Hospital of Nanjing University Medical School. All participants provided informed consent before the experiment.

Supporting information

Supporting Information: trc270294‐sup‐0001‐SuppMat.doc

Supporting Information: trc270294‐sup‐0002‐ICMJE.pdf