Applications and potential mechanisms of transcranial magnetic stimulation in autism spectrum disorders

Abstract

Background

Transcranial magnetic stimulation (TMS) is a non-invasive neuromodulation technique widely applied in clinical interventions for various neurological disorders. Its safety profile, ease of operation, and potential therapeutic value have prompted exploration in autism spectrum disorder (ASD). However, TMS efficacy in ASD exhibits marked heterogeneity, primarily due to the lack of robust scientific theoretical support for existing therapeutic approaches—this severely hinders the standardisation of TMS in ASD clinical practice and the improvement of therapeutic outcomes.

Main Body

The present narrative review first provides an in-depth synthesis of existing TMS research in ASD, focusing on the safety validation of different stimulation protocols, the scientific rationale for protocol selection, and the observed differences in efficacy across recent studies. It also explores the suitability of stimulation coil types and the rationality of target localisation, offering practical guidance for harnessing TMS’s therapeutic potential. Building on this, the core of the review focuses on summarising both potential and proposed mechanisms of TMS in ASD, encompassing key dimensions such as ion channels, excitatory-inhibitory imbalance, synaptic plasticity, neural oscillations, neuroinflammation, and the gut microbiome, while elucidating the interrelationships among these factors.

Conclusion

This narrative review systematically synthesises the proposed mechanisms by which TMS may affect ASD, aiming to provide a foundation for optimising TMS-based therapeutic regimens for ASD and advancing the development of TMS as a more effective and reliable treatment option.

Introduction

Autism spectrum disorder (ASD) is a range of neurodevelopmental disorders characterized by persistent deficits in social communication and social interaction, as well as restricted, repetitive patterns of behavior, interests, or activities [1, 2]. ASD affects individuals worldwide, regardless of their race, ethnicity, or socioeconomic status. According to the latest 2025-released data from the U.S. CDC’s Autism and Developmental Disabilities Monitoring Network (2022 surveillance year), approximately one in 31 (3.2%) children aged 8 years across 16 monitoring sites in the U.S. has ASD, with prevalence varying by region and demographic characteristics [3]. ASD also affects approximately 2.2–2.3% of adults, with prevalence estimates rising from 1.1% in 2008 to 2.3% in 2018 [4]; while updated adult-specific data remain limited, current evidence suggests rates may be higher than previously reported, partly reflecting improved awareness and diagnostic practices. The rising number of diagnoses and the need for lifelong care and support in areas such as education, healthcare, and social services places a significant burden on patients, families, society, and the country [5]. Despite the challenges presented by this burden to the social system, current treatments remain limited due to the clinical heterogeneity of ASD manifestations [6]— specifically, interindividual differences in language abilities, the high prevalence of comorbid conditions (e.g., attention-deficit/hyperactivity disorder, epilepsy), and variability in age at onset — as well as dynamic symptom changes across different developmental stages. Currently, no drugs are available to address the core deficits of ASD; existing pharmacotherapies only target maladaptive behaviors and comorbidities [7, 8]. Beyond pharmacotherapy, mainstream non-pharmacological interventions for ASD include Applied Behavior Analysis and cognitive behavioral therapy. However, these non-pharmacological strategies have notable limitations, such as inconsistent efficacy across patient subgroups, restricted accessibility in resource-poor regions, poor suitability for specific populations (e.g., minimally verbal individuals) [9, 10], and associated ethical concerns [11]. Thus, exploring the pathogenesis of ASD is an urgent priority, as it can provide a theoretical basis for developing effective novel interventions or optimizing existing therapeutic modalities.

In recent years, various neuromodulation techniques have been developed to provide new therapeutic options for central nervous system (CNS) diseases. One of these techniques, transcranial magnetic stimulation (TMS), has gained significant attention. TMS is a classical non-invasive neurostimulation and neuromodulation technique [12]. Due to its safety, non-invasiveness, and ease of use, this treatment has been widely applied in clinical settings. It has undergone extensive clinical research and has shown potential results. The core neuromodulation mechanism of TMS involves reversible regulation of cortical excitability via electromagnetic induction that propagates to deep brain regions, with long-term therapeutic effects mediated by induced synaptic plasticity, modulated neurotransmitter release (e.g., dopamine, 5-hydroxytryptamine), and modulation of brain network connectivity [13]. Supported by this mechanism, the U.S. Food and Drug Administration has granted clearance for TMS to treat multiple indications, including major depressive disorder (recently extended to adolescents), obsessive-compulsive disorder, migraine, and smoking cessation [13–16]. A recently published systematic review analyzed clinical research reports on TMS interventions for ASD over a five-year period. Its findings indicated that TMS may have potential in improving the clinical core symptoms of ASD, particularly in repetitive behaviors and social communication [17]. Furthermore, the same literature indicates that TMS is a promising and safe therapeutic option for patients with ASD [18, 19], thereby supporting the conduct of additional large-scale clinical trials on the application of TMS in patients with ASD in the future. However, while the demand for ASD treatment is pressing, existing research remains constrained by issues such as small sample sizes, inadequate control of placebo effects, and reliance on subjective clinical assessments [20, 21], and there remains scope for improvement in the integration of analyses. Specifically, on the one hand, previous research focusing on the critical aspect of TMS intervention efficacy still has room for improvement in terms of systematicity and comprehensiveness. More importantly, there remains a lack of a unified theoretical model to guide TMS intervention design for ASD, and the mechanistic links between specific stimulation parameters (e.g., frequency, pulse number, target region) and relevant biomarkers (e.g., synaptic plasticity indices, E-I balance markers) have not been fully established. On the other hand, the majority of studies either emphasise clinical outcome observation or focus on fundamental mechanism exploration, yet lack cross-dimensional integration of the latest clinical trial data and cutting-edge basic research conclusions. Consequently, the relationship between intervention effects and potential mechanisms remains inadequately elucidated.

Against this backdrop, the present narrative review seeks to synthesize current research findings regarding the application of TMS in ASD, while attempting to address the aforementioned research gaps in a targeted manner. A key innovation of this review lies in systematically integrating clinical evidence with mechanistic insights to bridge the existing disconnect between clinical outcomes and underlying mechanisms. Building on a comprehensive narrative synthesis of its application status (including the development of optimized stimulation protocols and individualized target selection strategies), and combined with the latest clinical efficacy data, this study will provide a preliminary analysis of the potential specific pathways underlying the therapeutic effects of TMS—with a focus on integrating individualized targeting strategies with multi-level mechanisms—from the perspectives of classic mechanisms such as ion channels, excitation-inhibition (E-I) balance, synaptic plasticity, neural oscillations, neuroinflammation, and gut microbiota (refer to Fig. 1). This integrative approach, which combines individualised target selection with multi-level mechanistic analysis, provides a coherent framework linking intervention effects to underlying mechanistic explanations. The findings of this study may provide modest insights for the optimisation of subsequent treatment strategies for individuals with ASD, thereby laying certain theoretical groundwork for the comprehensive exploration of TMS therapeutic potential.

Fig. 1.

Shows that TMS may exert its effects through classic mechanisms. These mechanisms include ion channels, excitatory-inhibitory balance, synaptic plasticity, neural oscillations, neuroinflammation, and gut microbiota

Application of TMS in ASDs

TMS works primarily based on the theory of electromagnetic induction by generating a brief pulse of high current through a coil placed tangentially on the surface of the scalp. Flux lines that are perpendicular to the plane of the coil generate a magnetic field. This field does not attenuate due to the tissues surrounding the brain, such as skin and bone. Instead, it generates a phasic electric field in the target tissues [22, 23]. The electric field depolarizes excitable structures (such as neurons) within the brain, and action potentials are triggered when the electric field is strong enough to cause depolarization of the neuron’s membrane potential above a certain threshold [23]. Furthermore, through the implementation of various protocols, TMS can generate diverse stimulation effects that serve distinct purposes, such as diagnosing and predicting diseases, as well as providing therapeutic benefits [24]. TMS effects are dependent on various physical and biological parameters, including the pulse waveform and number, coil shape and orientation, stimulation strength, frequency, and pattern, direction of brain-generated currents, and the stimulated neuronal elements [25, 26]. In this article, we will specifically focus on the effects of three factors: stimulation pattern, coil, and target.

Stimulation patterns

TMS stimulation modalities and their characteristics

Depending on the pulses, TMS can be classified into three main stimulation modes. Single-pulse transcranial magnetic stimulation (sTMS) is a technique that generates transient currents in the cerebral cortex, instantly depolarizing neurons [27]. When sTMS is applied with appropriate intensity to the subject’s primary motor cortex (M1), it leads to motor evoked potentials. These motor evoked potentials are directly responsive to excitability and functional integrity of corticospinal tract with excellent temporal resolution [28]. Additionally, MEPs can assess other indices, such as resting motor threshold [29] and central motor conduction time [30]. Paired-pulse transcranial magnetic stimulation (pTMS) can deliver two consecutive stimulations of different intensities at very short intervals at the same stimulation site, or apply two stimulators to two different sites (also known as double-coil TMS), to study neural facilitation and inhibition by adjusting the intensity and interstimulus interval between the pulse (the previous one) and the subsequent test pulse [31, 32].

Repetitive transcranial magnetic stimulation (rTMS) is a technique that delivers multiple stimulation pulses at various stimulation frequencies (e.g., 1, 5, or 10 Hz) over short time intervals [33]. In general, sTMS and pTMS can be used to explore brain function, while rTMS is used to induce changes in brain activity. rTMS produces longer-lasting changes in neural activity compared to sTMS and pTMS protocols [25]. Additionally, the after-effects of rTMS primarily depend on the stimulation frequency and duration [34]. Depending on the frequency of stimulation, rTMS can achieve therapeutic and temporary excitatory or inhibitory effects on specific cortical functional areas, it is widely recognized as a classic finding that low-frequency (≤ 1 Hz) rTMS reduces neuronal excitability, whereas high-frequency (5–20 Hz) rTMS can increase neuronal excitability [35–38]. In addition to the frequency-dependent stimulation effect, the duration of the after-effects appears to be proportional to the duration of the stimulation. That is, the longer the stimulation, the longer the duration of the after-effects [25].

The traditional rTMS protocol consist of consecutive identical stimulations with fixed interstimulus intervals, and their effect depends on the frequency of the stimulation. Subsequent studies have developed new patterned protocols based on this. Theta burst stimulation (TBS) is a commonly used technique that consists of repetitive high-frequency stimulation pulses (3 pulses at 50 Hz) spaced 200 ms apart (i.e., 5 Hz, theta rhythm in electroencephalogram (EEG) nomenclature) [39]. The intensity is typically set to 80% of the active motor threshold, and the pattern is designed to enhance cortical excitability by mimicking cortical theta wave rhythms to improve synaptic transmission [40]. TBS may be a potential solution for optimizing therapeutic utility and duration of effects. This novel stimulation paradigm can modulate the human cerebral cortex, producing controlled, consistent, powerful, and longer-lasting after-effects on the physiology and behavior of the relevant brain regions with fewer impulses over a shorter period and at lower stimulation intensities [41]. Various TBS patterns elicit distinct effects on cortical excitability. Clinical studies typically employ two types: intermittent theta burst stimulation (iTBS) and continuous theta burst stimulation (cTBS). iTBS involves 2-second TBS treatments at 8-second intervals over a 192-second period (a total of 600 pulses), which promotes cortical excitability. In contrast, cTBS involves continuous TBS over a 40-second period with a total of 600 pulses, which reduces cortical excitability [42].

Safety and adverse effects in the ASD population

Recent clinical reports on TMS interventions for ASD (refer to Tables 1 and 2), indicate that current TMS protocols demonstrate generally favourable safety profiles: No cases of severe adverse events, including persistent headaches or seizures, were reported in any of the studies. A limited number of studies have documented mild and transient adverse reactions, including temporary discomfort resulting from periorbital muscle twitching, following interventions involving high-frequency conventional rTMS (e.g., 20 Hz protocols) or patterned rTMS (e.g., iTBS). Such discomfort is ordinarily resolved rapidly and does not impede the intervention process. This finding is consistent with the established mechanism of non-invasive modulation of cortical excitability by TMS, and it is also consistent with the safety considerations that are incorporated into current dosage designs, such as segmented pulses and bilateral stimulation. However, it is crucial to note that the vast majority of these clinical trials exhibit limitations in their follow-up design. The lack of long-term data on both efficacy (e.g., sustained improvement in daily functioning and core symptoms) and safety remains a major limitation. This paucity of information may impede a comprehensive evaluation of TMS’s long-term efficacy in ASD intervention. Consequently, further prospective studies with extended follow-up periods are required to further validate these findings. Nevertheless, drawing upon the findings of meta-analyses regarding the application of TMS in other neuropsychiatric disorders, such as depression and anxiety [43], TMS demonstrates relatively high overall safety, even when employing dosage regimens similar to those used in research on ASD, including interventions targeting paediatric populations. It is noteworthy that children and adults with ASD exhibit age-related differences in tolerance, with their developing cerebral cortex and immature neural circuits rendering them more sensitive to the intensity and frequency of stimuli. Consequently, more rigorous dose titration is required compared to adult protocols (e.g., commencing with subthreshold stimulation and gradually escalating to effective doses) [44, 45]. Concurrently, current clinical research incorporates dedicated ethical review measures for paediatric participants. These include the requirement for legal guardians to provide informed consent, where possible, for children who have the capacity to do so. Rigorous risk-benefit assessments must be conducted prior to the commencement of trials, and there is a requirement for real-time, continuous monitoring of adverse reactions throughout the intervention [44, 46].This cross-disease safety data may provide circumstantial evidence regarding the safety of TMS interventions in the ASD field to a certain extent. Furthermore, it provides a framework for optimising TMS dosages and developing long-term intervention protocols for future ASD patients, with a particular focus on the paediatric population.

Table 1.

Major studies of conventional rTMS interventions in ASD

Study Design	Risk of Bias	Intervention Programs	Theoretical Basis	Subject	Pulses	Target	Coil	Adverse Events	Findings/Conclusions	References
Clinical trial (Randomized controlled, double-blind, sham-controlled)	Low	1 Hz rTMS	In accordance with the minicolumnpathy theory, TMS exerts an influence on cortical excitability	ASD (active: n = 15; sham: n = 26; children)	180 pulses × 18 sessions	DLPFC	figure-eight coil	unreported	Following rTMS treatment, DTF connectivity was reduced in the α-band between O1 and T7 as well as between P7 and Fp1. Additionally, DTF values were reduced in the γ-band between Pz and T8	[219]
Clinical trials (Randomized controlled, double-blind, sham-controlled, pilot)	Low	20 Hz rTMS	Enhancement of cortical inhibition by high-frequency rTMS	ASD (active: n = 20; sham: n = 20; adults)	1500 pulses/session (750 pulses per hemisphere, bilateral) × 20 sessions	DLPFC	figure-eight coil	mild and moderate	Improvements in executive functioning	[54]
Clinical trials (Randomized controlled, double-blind, sham-controlled, proof-of-principle)	Low	20 Hz rTMS	rTMS stabilizes hyperplasticity by enhancing brain inhibitory mechanisms	ASD (active: n = 14; sham: n = 15; adults), TD (n = 30; adults)	6000 pulses × 1 session	M1	unreported	unreported	The application of rTMS has the potential to stabilize the excessive LTD observed in ASD	[73]
Clinical trials (Randomized controlled, double-blind, sham-controlled, pilot)	Low	20 Hz rTMS	The altered excitatory and inhibitory neurotransmission associated with ASD	ASD with executive function impairment (active: n = 16; sham: n = 12; adults), TD (n = 19; adults)	1500 pulses/session (750 pulses per hemisphere, bilateral) × 20 sessions	DLPFC	unreported	unreported	rTMS has been demonstrated to modulate glutamatergic levels in patients with ASD, with the direction of change observed to correlate with the patient’s baseline glutamatergic levels	[133]
Clinical trials (Randomized controlled, double-blind, sham-controlled)	Low	5 Hz rTMS	rTMS is a common technique used to enhance the excitability of underactive cortical regions and related networks	ASD (n = 28; adults)	1500 pulses × 10 sessions	dorsomedial prefrontal cortex	HAUT-Coil	unreported	rTMS reduces social-related disorders and social-related anxiety	[80]
Clinical trial (Randomized controlled, feasibility)	Moderate	1 Hz rTMS	Based on minicolumnpathy theory, rTMS over DLPFC can improve E-I ratio	ASD with intellectual disability (active: n = 16; sham: n = 16; children)	180 pulses × 18 sessions	DLPFC	figure-eight coil	unreported	Behavioral and EEG results showed differences before and after treatment	[53]
Clinical trial (Randomized controlled, wait-list controlled)	Moderate	1 Hz rTMS	Based on minicolumnpathy theory, low-frequency TMS restores cortical E-I balance and improves long-range cortical connectivity	ASD (active: n = 20; sham: n = 20; children)	150 pulses × 12 sessions	DLPFC	figure-eight coil	unreported	Error monitoring and corrective function were improved after TMS treatment	[74]
Clinical trial (Randomized controlled, wait-list controlled, pilot)	Moderate	1 Hz rTMS	rTMS over the DLPFC improves E-I ratio	ASD (active: n = 16; sham: n = 16; children)	180 pulses × 18 sessions	DLPFC	figure-eight coil	unreported	The rTMS group demonstrated notable enhancements in both behavioral and functional outcomes	[310]
Clinical trial (Single-group exploratory, pre-post intervention)	High	0.5 Hz rTMS	Alteration of cortical E-I balance by activation of inhibitory GABAergic double bouquet interneurons	ASD (n = 13, aged 9–27 years)	150 pulses × 6 sessions	DLPFC	figure-eight coil	unreported	Significant changes were observed in the early, mid-latency, and late event-related potential components in the frontal, central, parietal, and parieto-occipital regions of interest	[72]
Clinical trial (Single-group exploratory, pre-post intervention)	High	1 Hz rTMS	Low-frequency rTMS has been demonstrated to normalize aberrant gamma oscillations in patients with ASD, and to improve repetitive behaviors and executive functions	ASD (n = 19; children), TD (n = 19; children)	180 pulses × 12 sessions	DLPFC	figure-eight coil	unreported	Following the rTMS intervention, patients with ASD exhibited a notable reduction in gamma responses to task-irrelevant stimuli, a diminished propensity for aberrant behaviors, and a decline in irritability, hyperactivity, and repetitive behavior scores as evidenced by behavioral questionnaires	[311]
Clinical trial (Single-group exploratory, pre-post intervention)	High	1 Hz rTMS	Low-frequency rTMS can reduce cortical excitability	ASD (n = 14; children)	160 pulses × 20 sessions	DLPFC	figure-eight coil	no adverse events occurred	It can alter the brain structure and function of children with ASD, and these changes are correlated with improvements in behavioral symptoms	[312]
Animal study (Preclinical, single-group pre-post intervention)	High	1 Hz rTMS	rTMS has been demonstrated to modulate synaptic plasticity, attenuate neuroinflammation, and inhibit glial cell activation	Sham, rTMS, ASD rat model, ASD rat model + rTMS (n = 8 per group)	900 pulses × 14 sessions	whole brain	small animal circular coil	unreported	The 1 Hz rTMS treatment has been demonstrated to significantly ameliorate abnormal behavior and deficits in synaptic plasticity, as well as excessive neuroinflammation, in ASD model rats	[57]
Animal study (Preclinical, single-group pre-post intervention)	High	10 Hz rTMS	rTMS has been demonstrated to possess antioxidant properties, to enhance BDNF production, and to impact dendrite growth and spine maturation	Sham, rTMS + Healthy rat, ASD rat model, ASD rat model + rTMS (n = 8 per group)	600 pulses × 14 sessions	whole brain	figure-eight coil	unreported	rTMS was observed to improve ASD symptoms for reasons related to antioxidant properties and the capacity to enhance BDNF, SYN levels, and dendritic spine density	[58]
Clinical trial (Single-group exploratory, open-label, pre-post intervention)	High	10 Hz rTMS	Imbalance between excitatory and inhibitory signals and altered functional connectivity within and between different brain regions	ASD and co-morbid major depressive disorder (n = 10; adults)	3000 pulses × 25 sessions	DLPFC	figure-eight coil	The side effects of rTMS are minimal and well tolerated.	A significant improvement was observed in depressive symptoms and core autism symptoms.	[313]
Clinical trials (Single-group exploratory, pre-post intervention)	High	15 Hz rTMS	High-frequency rTMS has the potential to facilitate interactions between parietal and other brain regions	ASD (n = 24; children), TD (n = 24; children)	—	Left parietal lobe	figure-eight coil	unreported	High-frequency rTMS over the parietal lobe may ameliorate core ASD symptoms by enhancing long-range connectivity reorganization	[314]
Clinical trials (Single-group exploratory, pre-post intervention)	High	1 and 10 Hz rTMS	High-frequency rTMS over the left DLPFC has been demonstrated to induce LTP of synaptic transmission in the stimulated area. Conversely, low-frequency rTMS over the right DLPFC has been shown to improve the pattern of abnormal brainwave activity in the gamma bandwidth in patients with ASD	ASD (n = 45; children)	—	left DLPFC with high frequency (10 Hz) and right DLPFC with low frequency (1 Hz)	figure-eight coil	unreported	Improvements in Childhood Autism Rating Scale scores and eye gaze on faces were observed	[315]

Table 2.

Major studies of patterned rTMS interventions in ASD

Study Design	Risk of Bias	Intervention Programs	Theoretical Basis	Subject	Pulses	Target	Coil	Adverse Events	Findings/Conclusions	References
Clinical trial (Randomized controlled, single-blind, sham-controlled)	Low	iTBS	iTBS has been shown to influence LTP in neurons, as well as synaptic plasticity	ASD (active: n = 22; sham: n = 27; children and adolescents )	2400 pulses/session (1200 pulses per hemisphere, bilateral) × 16 sessions	pSTS	figure-eight coil	unreported	Null effect of iTBS on the macro/microstructure of cerebral white matter	[59]
Clinical trial (Randomized controlled, single-blind, sham-controlled, two-phase)	Low	iTBS	iTBS has been demonstrated to influence cortical excitability and induce alterations in neuroplasticity	ASD (active: n = 40; sham: n = 38; children)	2400 pulses/session (1200 pulses per hemisphere, bilateral) × 16 sessions	pSTS	figure-eight coil	minor and transient side effects	Longer therapy sessions are necessary to achieve a therapeutic effect on social deficits in children with ASD	[60]
Clinical trial (Randomized controlled, single-blind, sham-controlled, crossover, pilot)	Low	iTBS	TBS can be delivered continuously or intermittently, producing inhibitory LTD-like or excitatory LTP-like effects, respectively	ASD (active then sham: n = 6; sham then active: n = 7; adults)	2400 pulses/session (1200 pulses per hemisphere, bilateral) × 10 sessions	pSTS	figure-eight coil	unreported	A 5-day course of multi-treatment iTBS shows therapeutic potential for adult patients with ASD	[67]
Clinical trial (Randomized controlled, double-blind, sham-controlled)	Low	iTBS	iTBS has been demonstrated to enhance cortical excitability	autism-like traits (active: n = 16; sham: n = 16; adults)	600 pulses × 5 sessions	pSTS	air-cooled figure-eight coil	The only reported side effect is temporary discomfort caused by muscle twitching around the eyes	iTBS can modulate relevant neural networks to improve patients’ emotional perceptions	[68]
Clinical trial (Randomized controlled, sham-controlled, crossover, pilot)	Low	iTBS	iTBS has been demonstrated to elicit excitatory LTP-like effects	ASD (cross-acceptance of active / sham stimulation: n = 19; adults)	2400 pulses/session (1200 pulses per hemisphere, bilateral) × 1 session	DLPFC, pSTS	figure-eight coil	No adverse reactions other than transient discomfort due to muscle twitching around the eyes have been reported	A single iTBS on bilateral DLPFC may alter neuropsychological functioning in ASD	[75]
Clinical trial (Randomized controlled, double-blind, sham-controlled)	Low	cTBS	An imbalance in the E-I ratio is a common feature of ASD patients	ASD (active: n = 30; sham: n = 30; aged 8–30 years)	600 pulses ×16 sessions	DLPFC	figure-eight coil	The discomfort subsides rapidly	No support for cTBS is valid	[51]
Clinical trial (Randomized controlled, double-blind, sham-controlled)	Low	cTBS	The correction of E-I imbalance can be achieved by the inhibition of cortical excitability	ASD (active: n = 28; sham: n = 27; aged 8-30years)	600 pulses × 16 sessions	DLPFC	figure-eight coil	unreported	The results demonstrated no statistically significant effect of cTBS over the left DLPFC on cerebral white matter macrostructures and microstructure as well as connectivity in patients with ASD	[52]
Clinical trial (Randomized controlled, double-blind, active-controlled)	Low	cTBS	cTBS induces LTD–like effects in cortical areas	ASD (active: n = 23; sham: n = 21; children)	1800 pulses × 20 sessions	active: the site of the left DLPFC that has functional connectivity with the amygdala; sham: standard prefrontal site	figure-eight coil	unreported	Personalized brain stimulation targeting key autism-related brain regions (the amygdala-prefrontal cortex circuit) demonstrates substantially greater therapeutic potential than standard stimulation protocols, with superior outcomes in treatment efficacy, brain structural/functional changes, and neural network modulation.	[96]
Clinical trial (Single-group exploratory, pre-post intervention, open-label, pilot)	High	iTBS	Excitatory and inhibitory (E-I) imbalance	ASD (active: n = 10; adults)	1200 pulses × 1 session	lateral cerebellum	figure-eight coil	no severe adverse events	decrease in functional connectivity within the default-mode network and somatosensory motor network	[316]
Clinical trial (Single-group exploratory, pre-post intervention)	High	iTBS	iTBS has been demonstrated to enhance cortical excitability and elicit LTP-like effects	ASD (active: n = 10; children and adolescents aged 9–17 years)	600 pulses × 15 sessions	DLPFC	unreported	The treatment was found to be well-tolerated, with no serious adverse effects reported	The evidence suggests that iTBS may facilitate improvements in restrictive and repetitive behaviors, obsessive-compulsive behaviors, and neurocognitive functioning	[71]

Dose parameter optimization strategies

Furthermore, regarding dosage, as shown in Tables 1 and 2, conventional low-frequency rTMS interventions for ASD in early clinical trials demonstrated significantly lower single-pulse and total pulse doses compared to high-frequency conventional rTMS. A comparison of pulses per session in patterned rTMS vs. conventional low-frequency protocols shows patterned rTMS uses more pulses per session. This higher single dose stems from its targeted design. For example, most iTBS studies use bilateral segmented protocols, or deliver stimulation intermittently (not continuously) in one session. This approach rationally partitions the effective dose per continuous application to specific brain targets, thereby avoiding excessive stimulation intensity within a single session. This method aligns with neuroplasticity modulation mechanisms while reducing potential adverse reaction risks. It is noteworthy that cases involving exceptionally high single-pulse counts predominantly involve adult ASD patients (such as those with comorbid severe depression) or animal models. This phenomenon is closely linked to the specificity of research design. For adult patients, greater cortical maturity and tolerance allows them to withstand relatively higher single-pulse doses to pursue improvements in core symptoms (e.g., depressive mood and repetitive behaviours). By comparison, dosage settings for animal models require adjustment based on factors such as species-specific brain volume and cortical sensitivity. Despite the reduced number of pulses per session when compared to adult protocols, the total treatment duration is approximately equivalent to that of short-to-medium-term human interventions. This provides foundational safety and efficacy references for subsequent human clinical trial dose optimisation. This also underscores that rTMS dosage design must be fully tailored to the subject’s age, disease severity, and model type, rather than relying solely on pulse count to gauge the appropriateness of intervention intensity.

Current protocol selection and clinical rationales

In regard to the selection of stimulation modes, initial studies employed low-frequency rTMS as an intervention, which is predicated on the “minicolumnopathy” hypothesis and the core mechanism of E-I imbalance in ASD. Anatomical evidence indicates that minicolumns, the fundamental unit of information processing in the autistic brain, exhibit reduced size and altered internal structure. In particular, the number or function of gamma-aminobutyric acid (GABA)ergic neurons located in minicolumns may be abnormal, which could result in a weakening of inhibitory signaling between minicolumns and an increase in the ratio of cortical excitation to inhibition [47, 48]. Given that ASD is characterized by disrupted cortical excitability stemming from an elevated E-I ratio, inhibitory low-frequency rTMS was rationally selected as an early intervention strategy to restore the E-I balance. Consequently, rTMS modalities with inhibitory effects are employed for intervention purposes. A minority of studies employed high-frequency rTMS for intervention, primarily due to its capacity to enhance interactions within or between distinct brain regions. In contrast to the inhibitory mechanism of low-frequency rTMS, high-frequency rTMS is designed to target potential hypo-excitability in specific brain circuits of ASD individuals, thereby adjusting the E-I ratio through excitatory modulation. Contrary to the findings of previous studies, contemporary trends indicate a clinical preference for (or shift toward) excitatory iTBS as the intervention modality, marking a deviation from earlier research that was centred on inhibitory low-frequency rTMS. This shift in intervention protocols may be attributable to the divergent regulatory mechanisms underpinning the two approaches. The rationale behind iTBS is based on theories that emphasise excitatory long-term potentiation (LTP)-like effects and synaptic plasticity. Specifically, iTBS has been shown to induce LTP-like plasticity in the motor cortex via NMDA receptor modulation, which provides a direct electrophysiological basis for its regulatory role in synaptic plasticity and neural circuit function [49, 50]. Moreover, the only two studies of cTBS intervention in ASD were exploratory trials conducted by Ni et al. These studies sought to address a research gap by investigating the feasibility, tolerability, and safety of cTBS intervention protocols [51], as well as their effects on cerebral white matter macro-/microstructures and connectivity in patients with ASD [52]. It is worthy of note that in the selection of stimulation protocols, the preponderance of researchers tend to draw upon approaches that have been demonstrated to yield satisfactory outcomes in prior studies. For instance, Kang et al.‘s research confirmed that low-frequency rTMS significantly ameliorates irritability symptoms and repetitive behaviours in individuals with autism, while also enhancing event-related potential components associated with target stimuli [53]. Additionally, some studies have chosen to draw upon research findings concerning disorders exhibiting symptoms similar to ASD. For instance, Ameis’s study [54] was grounded in the premise that individuals with ASD and schizophrenia share comparable cognitive and functional impairments, with therapeutic medications exhibiting commonalities [55]. Furthermore, he incorporated prior research confirming that rTMS can ameliorate working memory deficits in schizophrenia patients [56], hence employing comparable rTMS parameters in his study. However, the extant research in this field remains markedly limited in scope. The majority of studies to date have focused on clinical intervention trials, with only a small number of basic research investigations utilising animal models. Furthermore, the results of these preliminary studies suggest that the intervention mechanisms may be associated with synaptic plasticity and neuroinflammation. Specifically, Xu et al. found that low-frequency rTMS (1 Hz) improved LTP deficits in the hippocampus of valproic acid-induced autism-like rats, normalising dendritic spine density alongside restored expression of synaptic proteins such as NR2B and PSD-95 [57]. Similarly, Afshari et al. confirmed that high-frequency rTMS (10 Hz) alleviated autistic-like behaviors in valproate-exposed rats by reducing neuroinflammatory markers, such as hippocampal tumor necrosis factor-α (TNF-α), and enhancing hippocampal synaptic plasticity through the upregulation of brain-derived neurotrophic factor (BDNF) and synaptophysin levels [58]. While accumulating evidence supports the efficacy of various TMS protocols in ASD, mechanistic understanding remains limited. Thus, the following section will elaborate on potential neurobiological mechanisms underlying TMS efficacy in ASD.

Factors contributing to inconsistent outcomes

From the perspective of overall efficacy, in studies with different levels of bias risk (refer to Tables 1 and 2), the regulatory effect of TMS on ASD shows clear stratification characteristics: in RCT studies with low bias risk, TMS (especially the 1 Hz/10Hz regimen for DLPFC) often exhibits a “slight but stable regulatory effect”, mainly improving repetitive behavior (such as reducing RBS-R scores by 0.3–0.5 standard deviations), and EEG indicators (such as θ wave power) show consistent changes; In non randomized studies with moderate bias risk, there is a divergence of conclusions regarding the improvement of social communication symptoms by TMS (approximately 60% of studies reported positive results, while 40% showed no significant difference), which may be directly related to sample heterogeneity and inconsistent stimulus parameters; As for small sample exploratory studies with high bias risk, although some reports have shown significant improvement in core symptoms, the reproducibility and reliability of these results remain questionable due to methodological flaws such as unblinding and insufficient sample size. The above regulatory effects are mainly reflected through the evaluation of relevant behavioral scales and changes in EEG indicators.

It is important to note that not all studies yielded positive outcomes, with some reporting negative conclusions. For instance, the high-bias, low-sample-size study conducted by Ni’s team did not support the efficacy of cTBS over sham stimulation in the left dorsolateral prefrontal cortex (DLPFC) [51]. In addition, this research group found iTBS to be ineffective in influencing macro- or micro-structural changes in brain white matter [59]. The causes of such negative outcomes are multifaceted, closely linked to the inherent high heterogeneity within ASD itself [54, 60], and are intrinsically linked to key elements of intervention design and methodological shortcomings in research methodologies (such as risks of bias). About heterogeneity, some children with autism exhibit hyperarousal to social information (excessive sensitivity to social stimuli leading to avoidance of social interaction), while others demonstrate hypoarousal (reduced social motivation and diminished interest). This distinction is supported by behavioural, eye-tracking, and neuroimaging studies [61]. The heterogeneity characteristics of ASD, such as age, severity of symptoms, and comorbidities, further limit the reliable evaluation of TMS efficacy, and fail to adapt stimulation regimens to specific arousal states and individual heterogeneity, which may directly impair treatment effectiveness. With regard to the design of interventions, considerable variations in pulse parameters are evident across studies. Specifically, conventional rTMS single-session pulse counts range from 150 to 6000, with total pulse counts for patterned rTMS spanning 600 to 38,400. It has been demonstrated that the outcomes of certain studies have been suboptimal, a phenomenon that can be attributed to the mismatch between the dosing and the characteristics of the target brain region excitability (for example, the application of low-dose inhibitory stimulation to areas that exhibit under-inhibition). At the cycle level, certain studies utilised only 1–5 brief treatment courses (e.g., specific iTBS studies with a single course), whereas the majority of favourable outcomes emerged from studies encompassing 12–20 extended courses. Short-term interventions are ineffective in inducing stable neuroplastic changes, thus failing to sustain long-term improvement in symptoms. At the level of study design, while most investigations incorporated sham stimulation groups, some early studies lacked placebo controls, and certain trials did not strictly implement double-blind protocols. This precludes ruling out interference from “natural symptom fluctuations” and placebo effects, thereby complicating the establishment of TMS-specific effects. The aforementioned design deficiencies, when considered collectively, have the potential to result in unfavourable outcomes. Consequently, greater attention should be paid to the heterogeneity of ASD, with differences in arousal states serving as a key basis for selecting excitatory/inhibitory TMS protocols. Furthermore, larger-scale and rigorously designed double-blind placebo-controlled randomized controlled trials will be needed in the future, and research will be conducted based on unified evaluation criteria. By further elucidating the subtypes of ASD and their neurobiological basis, while promoting dose standardization, rationalizing treatment cycles, improving control design, and standardizing research methodology to reduce the risk of bias, it is possible to develop precise treatment plans under the guidance of deep scientific theories.

Magnetic coils

The shape of the magnetic coil determines the pattern of the electric field. In the original TMS study, Barker and colleagues used circular coils, which have a high penetrating capacity, but the stimulating effect is not very focused, with a spatial selectivity of > 4 cm² [27], and are suitable for stimulating large and superficial motor areas, such as upper limb motor areas [25]. After conducting research, scholars designed the figure-eight coil. The coil consists of two adjacent wings with the same number of turns. The current in the two loops flows in opposite directions, resulting in the superposition of currents. This leads to direct stimulation effects on the superficial cortical areas located below the central segment, where neuronal fibers that are parallel to the central segment have the highest likelihood of being stimulated [62]. The figure-eight coil is a widely used type of coil in recent TMS-ASD studies (refer to Tables 1 and 2). This type of coil has the advantage of focusing the stimulus effect and generating the maximum current at the intersection of the two circular elements. However, it also has the disadvantage of limited penetration [27, 63]. To improve penetration, several coil models have been developed, including the Hesed coils. These coils have a flexible base that conforms to the curvature of the patient’s scalp, maximizing magnetic coupling at the desired location and direction [62]. In 2005, Zangen conducted a clinical study using Hesed coils for the first time. The study demonstrated that Hesed coils were effective in stimulating deeper regions of the brain at greater distances from the coil without inducing greater stimulation of superficial cortical areas [64]. The coil design combines the safety and convenience of non-invasive neuromodulation techniques with the depth of stimulation characteristic of invasive neuromodulation, greatly expanding the range of applications for TMS.

Stimulation targets and localisation techniques

In line with these anatomical and functional considerations, recent studies on TMS in ASD patients have focused on several important stimulation targets. The majority of these targets are directly associated with the neural mechanisms that are considered to underlie core ASD symptoms. Despite the confirmation provided by extant research that rTMS improves core ASD symptoms, the heterogeneity of intervention effects suggests that personalised target localisation based on individual brain functional differences is key to enhancing treatment precision. This conclusion is in alignment with the prevailing trends in the field of neuromodulation; the paradigm of precision medicine is driving a shift in TMS therapy from ‘standardisation’ towards ‘individualisation’. The accuracy of target localisation is the pivotal component in achieving this transition. The following section will elaborate on the core stimulation targets and research results in ASD treatment, combined with the application of TMS localization technology (relevant research data can be found in Tables 1 and 2).

Conventional approaches and limitations in target localisation

Conventional TMS target localization is based on the “standard” distance from the scalp to the stimulation site. The primary motor cortex (M1) is typically identified as the site that elicits the largest motor-evoked potential (MEP) in contralateral hand muscles. The dorsolateral premotor cortex and DLPFC are located approximately 2–3 cm and 5 cm anterior to M1, respectively [65]. Although rapid and convenient, this method is prone to inaccuracy, primarily due to inter-individual anatomical variability and operator-dependent inconsistencies [66]. It is primarily suitable for localizing the target brain region, such as M1, or brain regions with specific positional relationships with it. An alternative approach uses manufacturer-provided electrode caps (often based on the international 10–20 EEG system) with pre-marked functional regions to enable rapid target localization. Neuronavigated TMS based on individual T1-weighted Magnetic resonance imaging (MRI) has become widely adopted. This technique transforms standard target coordinates into Montreal Neurological Institute (MNI) space, registers them to the patient’s anatomy, and employs frameless stereotaxy for precise coil positioning [51, 54, 59, 67, 68]. Compared with electrode-cap methods, it offers significantly higher accuracy. However, anatomical location does not always correspond precisely to functional regions. To address this issue, structural images can be aligned with functional images, a method already used in clinical practice [69]. Advances in localisation techniques have enabled the identification of multiple stimulation targets that are highly correlated with core symptoms of ASD. These targets encompass key functional networks, such as cognitive regulation and social perception. The ensuing sections provide exhaustive elaboration on each core target.

DLPFC

The DLPFC is considered to be one of the most extensively studied targets in TMS therapy for ASD. The therapeutic value of the intervention lies in its ability to regulate multiple cognitive functions, including working memory, rule learning, planning ability, attention, and motivation. Impairments in these functions are characteristic of individuals diagnosed with ASD [70]. A substantial amount of clinical research has corroborated the therapeutic efficacy of targeting this region, with findings encompassing fundamental mechanism exploration and efficacy assessment [51, 53, 54, 71–75].

It is important to note that intervention effects on the DLPFC exhibit significant heterogeneity, a phenomenon that is directly linked to functional alterations in this brain region caused by central nervous system disorders. On the one hand, individual variations exist in the strength of connections between the DLPFC and other brain areas (such as distinct subregions of the subthalamic cingulate gyrus). The effects of TMS stimulation can propagate through anatomical connections to surrounding regions, thereby modulating specific neural circuit functions [76]. On the other hand, variations in target localisation methods also influence therapeutic outcomes. It is evident that traditional ‘standardised’ coordinate definitions struggle to match the individual specificity of brain function, whereas personalised localisation based on functional connectivity demonstrates superior potential.

Recent studies have revealed that individuals diagnosed with ASD exhibit significantly higher levels of peak functional connectivity between the right DLPFC and the nucleus accumbens than the general population. Furthermore, this connectivity strength exhibits a negative correlation with ASD symptom severity, suggesting that the nucleus accumbens may function as an effective “seed point” for guiding DLPFC localisation [77]. Specifically, the selection of the voxel within the right DLPFC exhibiting the strongest negative correlation with the nucleus accumbens as the stimulation target holds promise for more precise alleviation of core ASD symptoms. This hypothesis has been corroborated by subsequent studies; for instance, Cash et al.‘s review confirmed that highly effective TMS targets within the frontal cortex often exhibit stable functional connectivity with deep limbic regions such as the subthalamic cortex [78], further underscoring the clinical significance of personalised DLPFC localisation.

Posterior superior temporal sulcus (pSTS)

The pSTS has been identified as a key target for the regulation of social and perceptual functions in individuals diagnosed with ASD. Its core physiological functions are intrinsically linked to social cognition, language perception, and emotion recognition – domains where deficits constitute the core symptomatology of ASD [79]. As one of the recommended targets for TMS therapy, the therapeutic value of pSTS intervention has been validated by multiple clinical studies, covering efficacy assessments and optimisation of target localisation [59, 67, 68, 75, 80].

From a neural circuit perspective, the pSTS exhibits functional connectivity with the adjacent temporoparietal junction (TPJ), with both regions jointly participating in neural networks processing social information. However, in a manner analogous to the DLPFC, the efficacy of pSTS stimulation is contingent on precise localisation. The utilisation of conventional anatomical landmarks proves inadequate in accounting for the inherent functional variability amongst individuals. This underscores the necessity for the incorporation of functional imaging techniques into the identification of personalised targets. This requirement is closely aligned with the core characteristic of ASD brain functional heterogeneity.

Right inferior frontal gyrus (IFG) and right TPJ

Recent expert consensus explicitly recommends the right IFG and right TPJ as emerging targets for ASD-TMS treatment [21]. Despite the differences in functional emphasis, both models focus on core deficit domains of ASD. Furthermore, the TPJ, due to its proximity to the pSTS, forms a synergistic regulatory effect with this region and is therefore frequently discussed in conjunction.

The core therapeutic value of the right IFG lies in its ability to improve social deficits and communication impairments, while also constituting a key component of the theory of mind system (the neural mechanisms underpinning understanding others’ mental states) [81, 82]. The right TPJ is primarily associated with attention deficits, attentional shifting functions, and the regulation of theory of mind [83, 84]. This region has now become a key target in multicentre randomised controlled trials [85], with its intervention potential undergoing broader clinical validation.

Recent research indicates that the rationale for recommending these two targets stems not only from their functional associations but also aligns with the trajectory of personalised treatment development. In a manner analogous to the DLPFC, the efficacy of IFG and TPJ interventions is contingent on the strength of functional connectivity with deeper limbic regions. Consequently, peak functional connectivity-based localisation methods are equally applicable to these targets, offering prospects for further enhancing intervention precision.

Personalised transformation of localisation techniques: challenges and optimisation

A synthesis of extant research indicates that TMS localisation methods for ASD have shifted from “standardised” to “personalised” approaches. This transition, which is currently a research hotspot, is fundamentally grounded in the significant individual variation in human brain anatomy and function [86], and is primarily driven by localisation techniques guided by functional magnetic resonance imaging. This technique facilitates precise targeting based on individual functional connectivity patterns, such as DLPFC-striatal circuits or prefrontal-limbic system connections, rather than relying on universal anatomical landmarks [87]. For instance, the cortical partitioning method developed by Professor Liu’s team employs resting-state functional magnetic resonance imaging to map functional brain atlases at the individual level. The precision of the device has been validated through invasive cortical stimulation testing [88]. When applied to TMS treatment for post-stroke aphasia, this technique demonstrated outstanding efficacy in language function recovery [89], providing a technical reference for precise localisation in ASD. Furthermore, the selection of personalised targets can be refined to accommodate distinct ASD subtypes. For instance, targeting the medial prefrontal cortex-amygdala circuit may help modulate emotional and social information processing in individuals with social communication deficits [90], while regulating the DLPFC-striatal pathway could address repetitive behavioural symptoms, given that striatal circuit dysfunction is closely linked to stereotyped and repetitive behaviours in ASD [91]. In addition, closed-loop therapeutic approaches based on brain states (such as EEG-rTMS) have emerged as a significant avenue of research [92]. The aforementioned research pathways under discussion are predicated upon the identification of inter-individual variations in brain function. These findings provide a theoretical foundation for exploring the neural mechanisms underlying cognitive and behavioural changes. They also represent a crucial step towards achieving personalised precision medicine through neuromodulation.

Nevertheless, personalised diagnosis poses particular challenges in cases of ASD, especially in children. MRI scanning requires patients to tolerate high-decibel noise and to maintain head stillness for several tens of minutes, a requirement with which children diagnosed with ASD often struggle to comply. Consequently, sedation or anaesthesia using drugs such as propofol or dexmedetomidine is frequently employed in research settings to alleviate discomfort and optimise imaging quality [93]. Propofol remains the most frequently employed agent, administered either alone or in combination, while dexmedetomidine usage exhibits a marked upward trend. It is noteworthy that both drugs demonstrate a low incidence of adverse events [94].

It is imperative to acknowledge the potential for these medications to compromise the integrity of functional imaging results. For instance, propofol has been demonstrated to induce a comatose state, thereby reducing the amplitude of spontaneous low-frequency oscillations in functional MRI signals across multiple brain regions, including the prefrontal cortex, temporal pole, and hippocampus [95]. Consequently, when assessing the correlation between blood oxygen level-dependent signals in individuals with ASD and in healthy controls, the potential influence of sedatives must be accounted for in order to avoid misinterpretation of brain functional characteristics. The clinical evidence demonstrates the efficacy of personalised targeting, as evidenced by the significant superiority of personalised stimulation of key neural hubs, such as the amygdala-prefrontal cortex circuit, in comparison to standard protocols in terms of clinical efficacy, brain structural/functional reorganisation, and neural network regulation [96]. This provides a clear direction for TMS treatment in ASD.

Potential mechanisms for TMS intervention in ASDs

Ion channels

TMS relies on the principle of electromagnetic induction to generate electric fields in target tissues, how do TMS-induced electric fields convert physical stimulation into therapeutically relevant biological effects? Marino et al. proposed that magnetosensory evoked potentials elicited by magnetic stimulation arise from direct interaction between the induced electric field and neuronal ion channels. Specifically, the receptor potentials required to generate evoked potentials are triggered by direct interactions between the induced electric field and the ion channel. These interactions result in changes in the mean probability of the channel being in the open state [97], and that the strength of the induced field can alter the mean ion channel opening time [98]. We hypothesize that TMS directly modulates voltage-gated ion channels in neuronal membranes. These ion channels are widely expressed in many types of neurons throughout the brain as well as in non-neuronal tissues and are key modulators of neuronal excitability, making them effective targets for regulating neuronal function [99]. This finding is consistent with the fundamental logic established in previous research, which posits that rTMS exerts its influence on ion channel states and functions by modulating stimulation parameters. For instance, this study utilised cellular experiments and animal models to observe that rTMS can transiently open voltage-gated sodium channels, affect potassium channel activity, and also induce delayed alterations in intracellular calcium ion concentrations [100]. Subsequent investigators have shown that the effect of TMS on the excitability of neurons is related to the ion channel. Acute high-frequency rTMS at both 0.8 and 1.2 motor thresholds significantly activated voltage-gated sodium current, inhibited voltage-gated potassium current, and the delayed rectifier potassium current compared to controls. These effects were attributed to alterations in the dynamic properties of voltage-gated sodium and potassium channels. The above results suggest that ion channel modulation may be a potential intrinsic regulatory mechanism by which rTMS enhances neuronal excitability in dentate gyrus granule cells, with effects increasing with stimulus intensity [101]. The physical principle behind TMS is based on Faraday’s law, which induces electrical currents in neurons. An alternative explanatory hypothesis for the mechanism of the effect of magnetic stimulation on neurons was presented and justified by computational and numerical simulations in another study. The study demonstrated that transcranial static magnetic stimulation induces the Lorentz force, which generates friction between ions and the channel wall in membrane channels. This friction decreases channel conductance, and simulations using the Hodgkin-Huxley model found that even a slight reduction in conductance effectively inhibits action potentials and neuronal activity [102]. The study indicates that the Lorentz force acting on the ions flowing through the neuronal membrane channels could also be a candidate physical mechanism to reduce the excitability of the motor cortex through magnetic stimulation techniques. Whether other classes of ion channels undergo modulation comparable to that of voltage-gated ion channels under magnetic stimulation remains an open question in the field. Separately, Chu et al. discovered that transcranial magneto-acoustic stimulation (TMAS) can alleviate neuroinflammation, damage to synaptic plasticity, and abnormal neuronal oscillations in Alzheimer’s disease mouse models by activating microglial Piezo1, a mechanosensitive ion channel that converts relevant mechanical and electrical stimuli into biochemical signals that enhance microglial autophagy and promote phagocytosis and degradation of β-amyloid, as confirmed by the blockade of Piezo1 with the antagonist GsMTx-4, which prevented the beneficial effects of TMAS [103]. Interestingly, TMAS delivered a more robust intervention effect compared with ultrasound stimulation alone. This superiority may be attributable to the combined action of magnetic stimulation-induced electric fields on Piezo1, an ion channel with high electrical stimulation sensitivity. The dual modulation of magnetic and electrical signals thus exerts a potent superimposed effect through biological synergism. Based on the existing studies, we speculate that TMS can act directly on ion channels and convert physical stimulation effects into biological effects by mediating ion channels, providing a theoretical basis for TMS to intervene in other mechanisms of action, such as E-I balance, neural oscillations, and salient plasticity.

E-I imbalance

Normal functioning of neuronal circuits requires a balance between synaptic excitation and inhibition, maintained primarily by GABA in conjunction with glutamate, and failure to establish or maintain this balance may underlie the neural basis of neurological disorders such as schizophrenia and ASD [104]. The E-I imbalance hypothesis is recognized as a common underlying deficit in ASD patients and plays an important role in the pathophysiology of ASD [105]. This hypothesis suggests that the shifts in neuronal excitation and inhibition are controlled by the relative amount (likely resulting from elevated glutamatergic excitation and/or reduced GABAergic inhibition) [106] and activity of glutamatergic and GABAergic systems [107]. GABA is the primary inhibitory neurotransmitter in 20%-44% of human cortical neurons [108]. On the other hand, glutamate, which is a precursor of GABA, acts as the primary excitatory neurotransmitter in the CNS and is the most abundant free amino acid in the brain [109, 110]. Glutamate is at the crossroads of many physiological processes, including but not limited to learning, memory, cognition, and emotion [111]. Additionally, since glutamate is not broken down outside of the cell, the brain relies on glutamate transport performed by excitatory amino acid transporters as well as their ability to take up excess glutamate, preventing excitotoxicity from occurring thus maintaining proper neuronal function [112]. Studies have found that an imbalance between excitatory and inhibitory neurotransmission is associated with metabolic abnormalities [113], which can lead to increased noise and hyperexcitability in the cerebral cortex [107]. Various techniques have been used to examine the manifestations of this imbalance in ASD, and it has been found to consist mainly of both increased and decreased E-I ratio. Rubenstein and Merzenich proposed in their E-I imbalance model of ASD that some types of ASD may be caused by elevated E-I ratios in the sensory, memory, social, and emotional nervous systems [107]. Later, Yizhar and his team demonstrated, using optogenetic tools, that an elevated cellular E-I balance within the medial prefrontal cortex of mice induces severe impairments in cellular information processing. This impairment significantly affects social behaviors and conditioned reflexes, and triggers baseline (non-evoked) rhythmic high-frequency activity in the range of 30–80 Hz [114], and that behavioral deficits in ASD are associated with elevated high-frequency activity [115, 116]. Further studies have reported that social deficits resulting from an elevated cellular E-I balance can be partially alleviated by increasing inhibitory tone to restore balance [114]. Whereas, decreased E-I balance ratio is observed in Rett syndrome, a pervasive neurodevelopmental disorder associated with mental retardation and ASD behaviors [117]. Furthermore, individuals with ASD also exhibit a decreased E-I balance ratio. Magnetic resonance spectroscopy has been used to quantify the concentrations of the inhibitory neurotransmitter GABA and the excitatory glutamate-glutamine complex in the anterior cingulate cortex and DLPFC of both ASD patients and neurotypical controls. Specifically, elevated levels of GABA were detected in the left DLPFC of ASD patients [105]—a finding that directly supports the hypothesized reduction in E-I balance, as increased inhibitory signaling would shift the ratio toward suppression. These studies provide evidence supporting the hypothesis of an E-I imbalance in individuals with ASD.

E-I imbalances are responsible for the abnormalities in social, behavioral, emotional, cognitive, sensory, and motor control that are closely associated with ASD [113]. Scholars have confirmed abnormal E-I balance in the typical Rett syndrome patient group in ASD through paired-TMS [118]. However, the underlying mechanisms remain poorly understood, though they are primarily thought to rely on the following pathways. Briefly, an imbalance between E and I in the brain can affect synaptic plasticity and neural oscillations, which in turn can lead to alterations in learning and memory [119]. For example, the N-methyl-D-aspartate (NMDA) receptor, one of the glutamate receptors, is the main postsynaptic excitatory amino acid receptor in the CNS. Its activation elevates intracellular Ca²⁺ concentration, ultimately inducing LTP and long-term depression (LTD), which play a key role in learning and memory [120, 121]. In addition, α-amino-3-hydroxy-5-methyl-4-isoxazole propionic acid (AMPA) receptors, a type of glutamate, are abundantly expressed on larger dendritic spines in the head. They mediate rapid components of synaptic transmission and contribute to strong synaptic connectivity, making them a key determinant of dendritic spine morphology [122]. The AMPA receptor not only redistributes in response to changes in synaptic activity patterns, but the cyclic process of rapid entry and exit into and out of the postsynaptic membrane can also modulate synaptic transmission and plasticity [123]. On the other hand, E-I imbalance can also result in abnormal neural network oscillations. For instance, synchronized oscillations at beta/gamma band frequencies form functional networks that are primarily mediated by a continuous flow of changes in excitatory and inhibitory synapses [124], and the interconnections between these neurons determine the strength and duration of the oscillations and control local synchronization [125]. Research has demonstrated that neural network oscillations play a role in synchronizing neuronal firing in cortical networks and coordinating decentralized cortical communication for spatio-temporal brain connectivity [126], further confirming that they may be related to cognitive functions such as selective attention, short- and long-term memory, and multisensory integration [113, 127]. Furthermore, altered E-I balance has been linked to hyperexcitability and the development of epilepsy, a common complication of ASD. It is widely accepted that decreased inhibition and/or increased excitability are key factors contributing to the onset of epilepsy [128–130]. Several studies have shown that depolarizing GABA can trigger epileptic seizures, and sustained seizure activity can, in turn, lead to depolarizing GABA. Interestingly, altering the switching time from depolarizing to hyperpolarizing GABA may be the key to causing the E-I imbalance [130].

It has been confirmed in reports in the field of ASD that TMS may have therapeutic effects by restoring E-I balance. Ikeda et al. found that the application of rTMS at 20 Hz induced persistent changes in mRNA expression levels, including GABAergic and glutamatergic transporter proteins in the mouse brain, suggesting that rTMS may modulate neuronal activity and synaptic plasticity by regulating the rate of uptake of glutamate and GABA in the synaptic cleft [131]. However, there is currently a lack of systematic transcriptomic or epigenomic data to comprehensively elucidate the molecular basis of these regulatory effects. Future work could benefit from integrating transcriptomic and epigenomic profiling to delineate downstream gene networks and epigenetic marks modulated by TMS. Tan demonstrated in his report that low-frequency rTMS intervention improved behavioral symptoms associated with the ASD rat model. He used whole-cell membrane clamp electrophysiology experiments to find that low-frequency rTMS intervention successfully restored the amplitude of miniature inhibitory postsynaptic currents, rather than miniature excitatory postsynaptic currents. This was associated with an increase in the expression of reverse inhibitory synaptic receptors, specifically GABAA α1 receptor subunits and vesicular GABA transporter [132]. The regulation of E-I balance by TMS is supported not only by animal experiments but also by clinical trials, which provides strong evidence. Recently, a randomized, double-blind, sham-controlled clinical trial examining the effects of rTMS intervention on glutamate levels in patients with ASD found that the direction of change in glutamate levels is related to baseline levels, with low baseline levels increasing glutamate levels and high baseline levels decreasing glutamate levels [133]. A similar phenomenon was found in another study where researchers used an inhibitory rTMS (1 Hz) protocol on the primary motor cortex of healthy individuals. This protocol could not affect excitatory (glutamatergic) neurotransmitters, but it resulted in a tendency to increase and decrease GABA concentrations in the motor cortex on the side that was originally low and high compared to baseline, respectively [134]. The above phenomenon reflects that TMS can have adaptive action effects based on the direction of E-I imbalance in brain regions. Nevertheless, conclusions drawn from studies of healthy individuals should be cautiously extrapolated to the ASD population. In addition, Stagg et al. similarly found by using magnetic resonance spectroscopy that when cTBS stimulation was given in the M1 region of the brain, it was found to significantly inhibit synaptic transmission for reasons associated with an increase in the concentration of GABA. This enhanced inhibitory effect of GABAergic neurons contributed to the maintenance of the aftereffects of TBS, demonstrating that cTBS mediated the localized activity of the corticocortical pathway between inhibitory neurons in the cortex; further data suggested that cTBS activated cortical GABA-receptor-ergic interneuron populations and that the sustained increase in GABAergic activity may have been maintained by the induction of glutamic acid decarboxylase and an increase in GABA concentration in the cytoplasm of GABA-receptor-ergic interneurons [135]. The above results are consistent with previous reports of glutamatergic changes after TMS intervention [136, 137], suggesting that TMS may improve ASD-related behaviours by restoring E-I balance, providing guidance for subsequent optimisation of TMS protocols. However, some mechanistic evidence derives from non-ASD models (animal or healthy human subjects), and the specific regulatory pathways within ASD require further validation.

Synaptic plasticity

Synaptic plasticity refers to activity-dependent changes in the strength of synaptic connections between neurons, forming the cellular basis for learning and memory [138–140]. The theory of synaptic plasticity is based on Professor Donald Hebb’s conjecture that the strength of the connection between two cells will increase if one cell repeatedly or continuously stimulates another cell [141]. In the early 1970s, scientists discovered that high-frequency stimulation of perforant path fibers in the rabbit hippocampus resulted in an increase in granule cell excitability, a phenomenon that could last for several hours and was termed LTP [142]. Later, researchers found that low-frequency electrical stimulation of the hippocampal CA1 region could induce LTD [121, 143], and that most synapses that exhibit LTP also express the corresponding form of LTD [144]. Furthermore, another approach to inducing LTP and LTD involves another principle of synaptic plasticity—spike timing-dependent plasticity (STDP) [145]. This principle is based on experiments conducted by researchers on the associative stimulation of presynaptic and postsynaptic neurons. STDP dictates that synaptic strength increases (LTP) if presynaptic activation precedes postsynaptic firing, but decreases (LTD) if the order is reversed [146–148]. In a simplified model, LTP and LTD are mediated by glutamate acting on NMDA and AMPA receptors. Presynaptic glutamate release activates NMDA receptors, leading to Ca²⁺ influx. Rapid, high-magnitude Ca²⁺ elevations activate kinase pathways (e.g., Ca²⁺/calmodulin-dependent protein kinase II), promoting AMPA receptor insertion and phosphorylation for LTP. Slower, lower-magnitude Ca²⁺ rises activate phosphatase pathways, causing AMPA receptor endocytosis and LTD [149–152]. When the concentration of Ca²⁺ increases rapidly, it triggers the kinase pathway, leading to extracellular secretion and autophosphorylation of AMPA receptors, which corresponds to synaptic LTP. On the other hand, when the concentration of Ca²⁺ increases slowly, it triggers the calmodulin-dependent phosphatase pathways, which endocytose surface AMPA receptors, reducing receptor number and permeability, corresponding to LTD [150]. The mechanisms of LTP and LTD can influence synaptic strength over long periods of time, and they are the most widely studied candidate mechanisms for learning [153]. Synaptic plasticity, which gives the nervous system the fundamental ability to self-adapt functionally and structurally, is both important for maintaining mental health and represents a potential mechanism that could be targeted to achieve therapeutic effects.

Recent studies have shown that abnormal synaptic plasticity is associated with the onset and development of ASD [154, 155]. Genetic analysis of populations with ASD and related syndromes has revealed that most risk genes affect synaptic function and plasticity [138, 156, 157]. Related research has demonstrated that mutations in many ASD risk genes commonly affect long-term changes in synaptic efficacy and mediate the strength or number of synapses through neuronal activity and sensory input-induced pathways [158, 159]. Recent studies have further highlighted the central role of impaired synaptic plasticity in ASD-associated genetic variants, for instance, mutations in SHANK3—one of the most reliably linked ASD risk genes—disrupt dendritic spine morphology and impair LTP, thereby contributing critically to the pathophysiological progression of ASD [160, 161], and another notable example is the ASD risk gene NL3^R451C: in the CA1 region of the hippocampus, its mutations result in an approximately 1.5-fold increase in AMPA receptor-mediated excitatory synaptic transmission, with an even more pronounced enhancement of NMDA receptor-mediated transmission, which subsequently induces an approximately two-fold upregulation of NMDA receptors containing the NR2B subunit and a nearly two-fold augmentation of LTP [162]. In summary, the aberrant synaptic plasticity observed in ASD may result in impaired information transfer between neurons in the brain, which in turn may lead to social and emotional dysfunction.

TMS offers a key advantage in modulating synaptic plasticity. Its stimulatory effects can be maintained long after treatment by modulating stimulation parameters to induce LTP or LTD in stimulated neurons [163]. In general, iTBS activates cortical excitability and promotes LTP-like effects, whereas cTBS has the opposite effect [41]. TMS holds promise for treating CNS plasticity-related disorders via synaptic normalization, and emerging clinical trials further confirm that rTMS can enhance synaptic plasticity in individuals with ASD, potentially ameliorating synaptic deficits. For instance, Desarkar et al. demonstrated for the first time that rTMS stabilizes excessive LTD in individuals with ASD, specifically by ameliorating the exaggerated LTD-like synaptic plasticity that is prevalent in ASD and fragile X syndrome models. However, the study did not observe a stabilizing effect of rTMS on LTP, possibly due to a small sample size or the intervention’s more specific effects on LTD [73]. Animal models have shown that low-frequency rTMS treatment effectively alleviates autism-like symptoms induced by neonatal separation and restores the balance between E-I activities, as evidenced by increased inhibitory synaptic transmission and inhibitory synaptic receptor expression. These findings suggest that low-frequency rTMS may alleviate ASD-like behavioral symptoms induced by neonatal separation by modulating synaptic GABA transmission [132]. But this conclusion derives from animal models and should be interpreted with caution when extrapolating to human ASD patients. Direct evidence of TMS effects on synaptic plasticity in ASD remains limited, factors such as the frequency, duration, and strength of synaptic transmission can influence the initiation of synaptic plasticity. For example, frequent and persistent synaptic transmission between two neurons may trigger synaptic plasticity in LTP. Although there is a lack of studies exploring the effect of TMS on synaptic plasticity and the mechanism of action in individuals with ASD, the mechanism has been extensively explored in other studies. A recent study demonstrated that rTMS could exert a neuronal protective effect and ameliorate dysfunction by promoting synaptic ultrastructural remodeling and up-regulating protein levels and mRNA expression of synaptic plasticity-related proteins, such as BDNF, tropomyosin receptor kinase B, NMDA receptor 1, synaptophysin, and phosphorylated cAMP response element binding protein, which are closely related to the development of LTP, in the brains of traumatic brain injury rats (non ASD model) [164]. The above findings presented are similar to the regulatory mechanism of PAS. Specifically, PAS improves learning and memory in cerebral ischemic rats, corrects the ultrastructure of synapses in the CA1 region, enhances the LTP of synapses in the CA3 and CA1 regions of the hippocampus, as well as promotes the protein levels and mRNA expression of BDNF and NMDA receptor 1, and protects cognition after cerebral ischemia by mediating the synaptic plasticity pathway function [165]. Pharmacological studies were used to validate the effects of TMS intervention. A small amount of memantine was administered prior to iTBS and cTBS, which completely blocked their facilitatory and inhibitory effects. However, it had no effect on resting motor threshold and active motor threshold [166]. This suggests that the mechanism of TBS modulation of synaptic plasticity is dependent on the NMDA pathway. Furthermore, additional investigation is required to determine if TMS has therapeutic effects by correcting synaptic plasticity through alternative pathways, and the specificity of these mechanisms in ASD still requires further validation in human subjects or ASD-specific models.

Neural oscillations

Endogenous cerebral neural oscillatory activity is generated by the electrical activity of a population of neurons in aggregate and represents the synchronized activity of an ensemble of neurons [167, 168]. This activity manifests as fluctuations in extracellular voltage, which can be measured by electroencephalography or magnetoencephalography on the scalp and can also be detected by electrocorticography intracranially [168]. The firing of peripheral neuronal populations can be recorded by depth electrodes, which captures a slower brain rhythmic fluctuation signal known as local field potential. This signal is often used to study brain function and can be categorized into the following main types of activity based on frequency: delta (< 4 Hz), theta (4–8 Hz), alpha (8–12 Hz), beta (12–30 Hz), gamma (30–80 Hz), and high gamma [168, 169]. It should be noted that the precise range of frequencies mentioned above varies among studies [170, 171]. Gamma oscillations are high-frequency electrical signals that can temporarily insulate excitation from subsequent inhibition in neural networks. They focus neuronal firing to specific phases of the oscillation cycle, provide the basis for a variety of functionally relevant synchronized activities [172], and improve the effectiveness, precision, and selectivity of communication between multiple regions [173]. Gamma oscillations have received significant attention as studies have shown that they can mediate a range of basic neural functions, including perceptual grouping, visual and perceptual awareness, sensory-motor integration, attention-dependent stimulus selection, and Neural Synchrony [174, 175]. Researchers have utilized cross-frequency coupling to identify gamma sub-bands, which include slow gamma (30–50 Hz), mid-frequency gamma (50–90 Hz), and fast gamma (90–140 Hz), and these sub-bands can coexist or occur separately [176]. Interestingly, the discrete 40 Hz point belonging to the slow gamma is significant, as McDermott highlights in his thesis that this frequency has become a point of interest in neurophysiological studies [177] due to an external stimulation paradigm that evokes gamma oscillations in the brain and may activate the cerebellum by increasing regional cerebral blood flow [178]. It is integral to cortical arousal and the processing of sensory and other information [179], as well as being linked to bottom-up-driven gestalt perception and cognitive functioning, such as selective attention, learning, and memory [180]. Therefore, the gamma frequency band at 40 Hz is a research-valuable gamma frequency band for use in targeted intervention studies and modulation.

Recent studies on the pathophysiology of ASD have identified several reliable physiological variants in the ASD population, including the prevalence of abnormal neural oscillatory patterns in the gamma frequency band in patients with ASD [181]. An et al. demonstrated that the phase of motion-associated gamma oscillations from the contralateral primary motor cortex of patients with ASD has a lower peak frequency and reduced power when compared to children with TD, and that these oscillations are associated with ASD symptom severity. The results tentatively confirm that indices of motor-induced gamma oscillations and behavioral performance represent potentially adequate biomarkers of ASD [182]. In addition to being associated with motor stimulus induction, patients with ASD also exhibit disturbed gamma oscillation patterns during visual [183, 184], auditory [185–187] and perceptual [188, 189] stimulation, particularly in the 40 Hz frequency band. Lovelace discovered enhanced resting-state gamma power in a mouse model of Fragile X syndrome, a common co-morbidity of ASD, in his animal studies. Additionally, he found reduced gamma band inter-trial coherence in the frontal cortex and auditory of the mouse in response to acoustic stimulation from 1 to 100 Hz. These findings are consistent with the characteristic manifestations of neural oscillatory deficits in patients suffering from the same disease and have important clinical implications [190]. Furthermore, studies have shown reduced coherence in other frequency bands, such as beta [191] and theta [192] oscillations in individuals with ASD, but for the time being, there is a lack of relevant studies that could provide sufficient reference significance for the study of ASD, and so the present focus will be mainly on the advances in research related to gamma oscillations.

Gamma oscillations are thought to result from the synchronized activity of a group of parvalbumin (PV) interneurons defined by a fast-spiking phenotype and the expression of the calcium-binding protein PV [193, 194]. These interneurons are a family of GABAergic inhibitory neurons found throughout the cerebral cortex [167], with basket cells being the most abundant and serving as typical PV interneurons [195]. PV cells are important for generating gamma oscillations, and impaired gamma oscillations may be a physiological biomarker of abnormal PV neuron function. In vivo optogenetic experiments have shown that PV interneuron activity responds to changes in cortical network oscillations. In particular, activation of PV cells selectively amplifies gamma oscillations [196], while inhibition of PV interneurons suppresses gamma oscillations in vivo, whereas driving these interneurons is sufficient to generate emergent gamma frequency rhythms [193]. Alterations in PV interneurons have also been found to be strongly associated with ASD and various other neurological disorders. Studies have reported that in brain samples from human patients with ASD [47] and in classical ASD animal models (such as valproic acid induction, neuroligin 3 R451C knockin, Cntnap2 mutant, Chromosomal 16p11.2 deletion, and Sarm1 knockout) [197–201], the number of PV-expressing interneurons, related gene expression, and protein levels were down-regulated. In addition, cortical hyperreactivity is observed in Shank3 knockout (leading to Phelan-McDermid syndrome with a high prevalence of ASD), Ube3 knockout (leading to Angelman syndrome, a neurodevelopmental disorder associated with ASD), and Fmr1 knockout (leading to Fragile X syndrome), and this hyperreactivity may be related to dysfunction of PV interneurons [202–205]. PV^−/− mice lacking PV expression exhibit core behavioral symptoms of ASD (e.g., social deficits and repetitive behaviors) and associated comorbidities (e.g., increased seizure susceptibility), as well as changes in brain morphology similar to structural changes reported in human patients with ASD (increased cortical volume and hypoplastic cerebellum), and also exhibit the classic pathological mechanism of ASD, i.e., a pre- and postsynaptic E-I imbalance [206], which echoes the previously mentioned E-I imbalance affecting gamma oscillations.

Is there a connection between the well-known hypotheses of pathomechanisms in ASD, such as E-I imbalance, gamma oscillations, and synaptic plasticity? To this end, the minicolumns theory will be employed to provide a detailed explanation. Minicolumns are comprised of pyramidal cells in the radial layers II - VI of the cerebral cortex and axon-dendritic interneurons. The pyramidal cells possess a high number of dendritic branches, which enable them to receive vast amounts of input information from other neurons, and their axons can transmit nerve impulses to other brain regions or different parts of the same brain region, occupying a crucial position in the process of information transmission and integration. Interneurons mainly form local connections with surrounding neurons. They modulate the activities of principal neurons, such as pyramidal cells, by releasing inhibitory neurotransmitters, including GABA, thereby achieving precise regulation of neural information transmission. Based on their immunoreactivity to three calcium-binding proteins (calbindin, calretinin, and PV), inhibitory interneurons can be categorised into various subgroups. These interneurons engage in dynamic interactions with pyramidal cells, contributing to the regulation of information processing within the circuits of the cerebral cortex. Calbindin and calretinin interneurons, which are immunoreactive, primarily function in intricolumnar communication, while PV interneurons are involved in transcolumnar signal transduction [207]. From an anatomical perspective, a reduction in the number of GABAergic neurons in minicolumns within the brains of individuals with autism would result in the weakening of inhibitory signals among minicolumns. Consequently, this might disrupt the excitatory/inhibitory balance in the cerebral cortex of autistic subjects and ultimately affect gamma oscillations [47, 48]. This is due to the fact that PV basket cells, a critical type of GABAergic neuron, are intimately linked with the generation of gamma oscillations. These cells play a pivotal role in the generation of gamma oscillations through the inhibitory effect mediated by GABA receptors. For a more thorough examination of this topic, readers are directed to the scholarly articles published by Professor György Buzsáki. In these articles, Professor Buzsáki elucidates the generation mechanisms of gamma oscillations from multiple viewpoints by employing classic neuronal models, such as the I-I model and the E-I model [176]. Alterations in gamma oscillations can also impact synaptic plasticity, given the temporal coincidence of these oscillations with the critical time window of STDP. Within this time window, the firing time relationship between neurons plays a crucial role in the induction of synaptic plasticity. According to the STDP rule, when the presynaptic neuron fires within specific time periods either before or after the postsynaptic neuron fires, it can respectively lead to an enhancement (for example, when the presynaptic spike precedes the postsynaptic spike by approximately 15 milliseconds, it results in LTP) or a reduction (for example, when the presynaptic spike lags behind the postsynaptic spike by approximately 6 milliseconds, it results in LTD) of synaptic strength. The underlying mechanism involves the facilitation of NMDA receptor opening by presynaptic glutamate release and the removal of the Mg²⁺ block of NMDA receptors by the backpropagation of postsynaptic spikes [208]. The temporal precision of spike-timing relationships is facilitated by gamma oscillations, which establish an accurate temporal framework. This temporal framework enables neurons to interact within the appropriate time window, thereby promoting or inhibiting the generation of synaptic plasticity.

The aforementioned conjecture is illustrated by the content in Fig. 2. Nevertheless, it is imperative to acknowledge that the underlying reality is considerably more intricate than the unidirectional regulatory relationship depicted in the figure. It is crucial to underscore the intricate and reciprocal relationship among these three components. The objective of this paper is to underscore the close interconnections among these elements within the framework of ASD. Furthermore, the internal logic among these three elements is yet to be thoroughly investigated.

Fig. 2.

The potential relationship between three hypotheses within the cortical microcolumn theory of ASD. (a) and (b) respectively demonstrate the disparities in microstructure between normal brains and autistic brains. A comparison of autistic brains with normal brains reveals an increased number of minicolumns, narrower widths, and a reduced number of interneurons involved in intricolumnar and transcolumnar signal transduction. (c, d, and e) represent the three hypotheses of E-I imbalance, gamma oscillation, and synaptic plasticity, respectively. E-I ratio, Excitatory-Inhibitory Ratio; LTP, Long-Term Potentiation; LTD, Long-Term Depression; NMDAR, N-Methyl-D-aspartate Receptor; AMPAR, α-Amino-3-hydroxy-5-Methyl-4-Isoxazole Propionic Acid Receptor; CaMKII, Ca2+/Calmodulin-Dependent Protein Kinases II

Studies have demonstrated that TMS can correct abnormal gamma oscillations in various disorders, such as Alzheimer’s disease [209, 210], schizophrenia [211], depression [212], and Parkinson’s [213], as well as related brain functions, including memory [214] and cognition [215]. For instance, TMS can normalize excessive gamma oscillations and improve cognitive function in patients with schizophrenia by acting on the DLPFC [216]. In another study focusing on healthy subjects, no significant changes were observed in the frequency range except for gamma band changes after TMS intervention, suggesting that TMS has a selective modulatory effect on gamma oscillations in frontal regions of the brain [217], which may be related to the fact that TMS can modulate GABAergic inhibitory neurons. Furthermore, the application of a TMS brain stimulation protocol that is based on modulating the gamma frequency band (40 Hz) induced healthy subjects’ inhibitory physiological after-effects that were well-tolerated [218]. The effects of TMS on gamma oscillations have received significant attention from researchers in the field of ASD. The researchers used EEG to detect brain network characteristics and found that children with ASD had significantly lower node degree, clustering coefficient, global efficiency, and local efficiency in all frequency bands compared to children with TD, indicating that the degree of neural correlation and the ability to integrate information between different brain regions are weakened in children with ASD. After a period of rTMS intervention stimulation, children with ASD showed improved social behavior, decreased connectivity from O1 to T7 and P7 to Fp1 in the alpha band of the EEG, and decreased effective connectivity from Pz to T8 in the gamma band, which is generally consistent with the phenomenon that effective connectivity from the posterior to the anterior is lower in children with TD [219]. It is important to note that this finding, which may appear to contradict previous observations that ‘TMS modulates only the gamma band’, is actually due to a fundamental difference in the core conditions of the two types of study: In previous studies, the N-back task was utilised as a specific cognitive task, and healthy subjects demonstrated no significant abnormalities in their brain networks. TMS required only targeted modulation of the gamma band to meet task demands, thus exhibiting ‘selective modulation’ characteristics. Conversely, individuals diagnosed with ASD have been shown to exhibit fundamental imbalances across the entire spectrum of brain networks (including baseline abnormalities in the gamma band). The objective of rTMS intervention is to rectify the dysfunction of the entire network and align it with normal patterns, as opposed to selectively modulating a single frequency band. Consequently, it is imperative to influence connectivity across multiple bands, such as alpha and gamma, to enhance cross-regional information integration capabilities synergistically. This phenomenon corresponds precisely with the fundamental characteristic of individuals diagnosed with ASD: The phenomenon of diminished efficiency has been observed across the entire spectrum of brain networks. In a recent clinical trial, a new metric (ringing decay) of gamma oscillations was used to evaluate the efficacy of TMS in the field of ASD. The study found a significant difference in the higher amplitude of event-related gamma oscillations in patients with ASD compared to the TD group. Following TMS intervention, the time required to reach the peak amplitude of gamma oscillations decreased significantly, while the time required for ringing decay increased and normalized. Additionally, ringing decay could be utilized to detect the impedance provided by inhibitory neurons to gamma oscillations [220]. What is the mechanism of action of TMS-mediated gamma oscillations? Previous studies have suggested that PV interneurons, which generate gamma oscillations, may be the key factor. In his report, Benali noted that different TMS modalities have varying modulatory effects on cortical excitability due to differences in the regulation of the activity of inhibitory cell classes. Specifically, iTBS may influence the inhibitory control of pyramidal cell output activity, whereas cTBS inhibits the activity of interneurons expressing calbindin D-28k, while the activity of another class of interneurons expressing the major calcium-binding protein, calretinin, remains unaffected. Importantly, iTBS intervention increased spontaneous neuronal firing activity and gamma power [221], a finding that contradicts the previous notion that PV interneurons promote gamma oscillations and suggests that the relationship between PV cells and gamma oscillations is not straightforward and may be influenced by other factors. It is essential to emphasize that the millisecond-scale rhythmicity of gamma oscillations can strictly constrain the firing timing of both presynaptic and postsynaptic neurons. Consequently, this phenomenon determines the temporal windows for enhancement or suppression of STDP through NMDAR-mediated current dynamics [208, 222]. When TMS enhances gamma oscillation power, phase synchronisation within neuronal ensembles markedly increases [208], concentrating presynaptic and postsynaptic firing events more precisely within the effective STDP time window [222]. This facilitates more accurate synaptic weight allocation [223], while the amplification of gamma oscillations itself directly modulates STDP expression efficiency [224]—this mechanism may constitute the cellular basis for TMS improving information integration deficits in ASD patients, though direct evidence specific to ASD remains lacking at present. It is noteworthy that, based on the results of our preliminary search of the relevant literature, it has been demonstrated that the ameliorative effect of TBS protocols modulated to mimic endogenous theta rhythms in the brain on ASD has been observed in some studies (refer to Tables 1 and Table 2). A study indicates that TBS, designed to mimic endogenous theta-gamma coupling, has shown promise in modulating cortical excitability and plasticity [225], and may interact with gamma oscillations within a nested hierarchical framework [226]. This θ-γ coupling mechanism is regarded as a core mode of brain information processing, in which θ rhythms act as a ‘carrier’ integrating global network activity, while γ rhythms are responsible for the fine-grained encoding of information within local neuronal clusters [227]. Dysregulation of their coordination may constitute a significant pathological basis for brain dysfunction in individuals diagnosed with ASD [226, 228]. Research into theta rhythms within the ASD field remains relatively scarce, a situation potentially attributable in part to early studies focusing more intensely on frequently demonstrated abnormal frequency bands such as gamma oscillations [229, 230]. Furthermore, the function of theta rhythms is frequently reflected through their phase-amplitude coupling (PAC) with higher-frequency oscillations, such as gamma waves. For instance, during natural speech processing tasks, children with ASD typically exhibit weakened or absent theta-gamma coupling, accompanied by abnormal beta-gamma coupling [227]. The high dependence of theta rhythms on the coupling context (in particular, specific cognitive tasks) poses a significant challenge in isolating the independent role of theta rhythms during non-task or resting states, thereby increasing the complexity of research analysis. In view of these findings, the future development of closed-loop TMS-EEG systems is of paramount importance. By capturing the phase, power, and spatiotemporal coupling characteristics of individual γ (and θ) oscillations in real-time via EEG, it becomes possible to identify ASD subtypes exhibiting specific neural oscillatory abnormalities (such as weakened θ-γ PAC or enhanced β-γ PAC). This facilitates the precise modulation of TMS stimulation parameters through adaptive adjustment [231]. This ‘monitor-modulate’ closed-loop framework shows great promise in enhancing intervention precision, and thus provides critical technological support for developing efficient, personalised ASD treatment strategies. In conclusion, it can be posited that protocols modulating gamma-band neural oscillatory activity may represent a viable alternative intervention for autism spectrum disorders.

Neuroinflammation

Neuroinflammation is an inflammatory response that aims to protect and maintain the normal structure and function of the brain. It is characterized by the infiltration of the CNS parenchyma by blood-borne lymphocytes and monocyte-derived macrophages, which leads to intense activation of glial cells [232]. Microglia can polarize into distinct activation states in response to environmental cues or specific stimuli. These states include classical activation, alternative activation, and acquired deactivation [233]. Microglia in the classical activation state are defined as M1 microglia. They are activated in response to inflammation and injury and induce pro-inflammatory cytokines such as TNF-α, interleukin-1β (IL-1β), and IL-6, as well as superoxide, reactive oxygen species (ROS), and nitric oxide production, leading to an inflammatory response and neuronal damage [234, 235]. M2 microglia refer to microglia in the alternative activation and acquired deactivation states. They play a crucial role in the late inflammatory and tissue repair phases by producing anti-inflammatory cytokines (such as IL-4, IL-10, and transforming growth factor-β) and neurotrophic factors, which help to reduce the inflammatory response and promote tissue repair [236, 237]. Research has demonstrated that microglia activation is a dynamic process that occurs along a continuum of M1 and M2 phenotypes [232]. Maintaining a balance between M1- and M2-type microglia is crucial for the normal functioning of the CNS. Like microglia, astrocytes also respond to CNS injury by undergoing morphological, molecular, and functional changes, resulting in the conversion to reactive astrocytes that generate an immune response. Reactive astrocytes can be categorized into two polarized states: a neurotoxic or pro-inflammatory phenotype (A1) and a neuroprotective or anti-inflammatory phenotype (A2) [238]. Activated microglia induce the formation of A1-reactive astrocytes through the secretion of IL-1α, TNF-α, and complement component 1, q subcomponent. Although type A1 astrocytes lose many canonical astrocyte functions—such as supporting neuronal survival and growth, maintaining synaptic function, and phagocytosing synaptic and myelin debris—they acquire potent neurotoxicity. Specifically, these cells are capable of rapidly killing newly born immature neurons and mature oligodendrocytes by releasing a number of pro-inflammatory factors and neurotoxins (e.g., complement protein C3, D-serine, nitric oxide, and TNF-α) [239, 240]. Furthermore, A1-reactive astrocytes significantly upregulate classical complement cascade genes, which have been demonstrated to be detrimental to synapses [240]. Conversely, A2 astrocytes are protective, upregulating neurotrophic or anti-inflammatory genes, and promoting neuronal survival and growth [239]. Additionally, A1/A2 astrocytes can communicate bidirectionally with microglia and other cells through both extracellular and intracellular signaling pathways to achieve mutual regulation [241].

A mounting body of evidence underscores the pivotal role of neuroinflammatory dysregulation in the pathophysiology of ASD [2, 242–244], marked by persistent activation of glial cells within the central nervous system [245, 246]. Vargas et al. systematically validated active neuroinflammatory processes in autopsied ASD brain tissue via immunohistochemistry, cytokine protein arrays, and ELISA. These processes manifested as marked activation of microglia and astrocytes within the cerebral cortex, white matter, and cerebellum, accompanied by progressive loss of Purkinje cells [247]. At the cytokine level, studies of central samples (brain tissue and cerebrospinal fluid) demonstrated high consistency: Macrophage chemotactic protein-1 (MCP-1) demonstrated significant elevation in both brain tissue (frontal cortex, anterior cingulate cortex, cerebellum) and cerebrospinal fluid (CSF) of individuals diagnosed with ASD. This finding suggests that MCP-1 is the most consistently elevated chemokine. In addition, IL-6 has been found to be markedly upregulated in CSF and the anterior cingulate cortex. TNF-α, IFN-γ, IL-8, and GM-CSF have also been consistently reported to be elevated in multiple brain regions (e.g., frontal cortex) [247, 248]. It is important to note that there is currently no conclusive evidence for elevated IL-1β protein levels in brain tissue. Li et al. utilised high-sensitivity multiplex flow cytometry to analyse frozen frontal cortex samples, observing a tendency towards elevated IL-1β levels, though failing to attain statistical significance (p = 0.11) [248]. Conversely, Tsilioni et al. noted a substantial upregulation of IL-18 (a constituent of the IL-1 family, alongside IL-1β) gene expression in the amygdala and DLPFC of children diagnosed with ASD. Furthermore, the results of the present study demonstrate that neuropeptide Y stimulates IL-1β production in human-derived microglia, thus suggesting the potential involvement of the IL-1 family in inflammation. Nevertheless, the elevation of IL-1β protein itself within the central nervous system remains controversial [249]. The histopathological evidence substantiated these molecular alterations. Morgan et al. utilised stereotaxic quantitative analysis to reveal significantly increased grey matter microglial density (p = 0.002) within the DLPFC of subjects diagnosed with ASD, alongside markedly enlarged mean microglial volume (p = 0.013) in white matter. Morphologically, these cells exhibited a classic activated phenotype, characterised by swollen cell bodies, shortened and thickened processes, and increased filopodia [250]. Li et al. also detected elevated levels of TNF-α, IL-6, GM-CSF, IFN-γ, and IL-8 proteins directly in the frontal cortex, alongside a significantly increased Th1/Th2 ratio (IFN-γ/IL-10), suggesting adaptive immune activation with a Th1 bias within the brain [248]. Zantomio et al. emphasised in their review that the mGluR5 signalling pathway downregulates microglial activation, and its reduced expression in the DLPFC of ASD patients may constitute a critical interface between synaptic dysfunction and neuroinflammation [251]. It is important to acknowledge that the aforementioned neuroinflammation-related findings are primarily based on ex vivo studies, such as post-mortem histology and central cytokine detection. Conversely, neuroimaging studies reflecting the state of the living brain – particularly those targeting the imaging marker for glial activation, the transporter protein TSPO – have yielded conflicting evidence that diverges from post-mortem findings. For example, Zürcher et al. used [¹¹C]PBR28 Magnetic Resonance Imaging-Positron Emission Tomography (PET-MR) to scan young adult male patients with ASD, revealing significantly lower TSPO expression in several brain regions (including the bilateral insular cortex, posterior cingulate cortex, superior temporal gyrus, and parahippocampal cortex) than in the control group. This suggests potential neuroimmune or mitochondrial dysfunction in these areas rather than atypical glial activation [252]. A preliminary study of female ASD patients observed elevated TSPO binding in the periventricular grey matter of the midbrain and caudate nucleus, suggesting that gender may be a key factor in differences in the neuroinflammatory phenotype [253]. Furthermore, a systematic review revealed that the three existing PET studies on TSPO expression in ASD patients produced contradictory findings: two reported decreased expression and one increased expression. This inconsistency may be due to sample heterogeneity (e.g., differences in age, sex, and clinical phenotype), tracer selection, or variations in analytical methods [254]. Indeed, the divergence in research conclusions is fundamentally linked to methodological limitations in assessing central nervous system inflammation in ASD. The evaluation of such inflammation necessitates the employment of multiple complementary approaches, each of which possesses distinct advantages and limitations. CSF cytokine assays have been shown to provide a direct reflection of central immune activity (e.g., elevated MCP-1, IL-6), yet these assays are invasive and capture only transient states within a narrow time window [247, 248]. similarly, peripheral blood cytokine measurements, while readily accessible, exhibit poor correlation with central levels due to blood-brain barrier selectivity and systemic confounding factors [254, 255]. PET using TSPO ligands, such as [¹¹C]PBR28, enables in vivo visualization of glial activation; however, results exhibit heterogeneity due to TSPO genetic polymorphisms, tracer kinetic variations, and participant differences [252–254]. Post-mortem microglial transcriptomics and immunohistochemistry (e.g., Iba1, GFAP, S100β) provide high-resolution cellular evidence of chronic activation [247, 250], yet remain confined to terminal pathology and fail to reflect dynamic developmental trajectories. Serum GFAP and S100β have been explored as peripheral surrogates for astrocytic activation, though their specificity for central processes remains contentious [256]. Collectively, these methodological constraints underscore the challenge of establishing a unified neuroinflammatory phenotype in ASD. Furthermore, although post-mortem histology and central cytokine studies provide compelling evidence for neuroinflammation, the reproducibility of cytokine research is compromised by factors such as sample heterogeneity (e.g., clinical phenotype, comorbidities, age), post-mortem interval, and detection methods (protein array vs. multiplex flow cytometry), with particularly high inconsistency in peripheral blood findings [247–251]. Consequently, the establishment of a single, universal ‘ASD inflammatory biomarker’ remains unattainable at this time. However, the recurrent detection of MCP-1 and IL-6 in central samples, coupled with sustained glial cell activation, collectively points to neuroinflammation as a pivotal component in the pathophysiology of ASD.

Glial cells may indirectly contribute to the development of ASD disease by participating in pathways such as maintaining E-I homeostasis, in addition to inducing inflammatory responses. Astrocytes are primarily responsible for maintaining homeostasis of E-I processes in the brain [257]. Astrocytes play a critical role in maintaining a balance between glutamate release and uptake when excess extracellular glutamate causes neurotoxicity by controlling expression of glutamate uptake transporters and a Ca²⁺-dependent exocytotic mechanism that inhibits glutamate excitotoxicity and modulates neuronal excitability (see Fig. 3a). These results suggest that astrocytes are essential for promoting E-I homeostasis [258]. This study confirms that astrocyte-specific glutamate transporter protein GLT1 knockout mice exhibit pathological ASD-related repetitive behaviors, such as excessive self-grooming and repetitive head twitching, and intervention with the NMDA receptor antagonist memantine drug ameliorated pathological repetitive behaviors in this mouse model [259]. These results suggest that astrocytes play a key role in promoting E-I homeostasis. On the other hand, microglia can also be involved in glutamate signaling through the X_c⁻system (see Fig. 3b). specifically, the X_c⁻ transporter protein in the X_c⁻ system is responsible for expelling glutamate out of the cell while translocating an equal amount of cysteine/cystine into the cell, and microglia are stimulated to secrete ROS to activate the TLR4 signaling pathway, which leads to an increased X_c⁻ expression which consequently promotes glutamate efflux [260], and the E-I imbalance caused by glutamate excess is an important influence on the development of ASD. Notably, microglia also regulate E-I homeostasis through pruning synapses. Studies have shown that depletion of microglia during growth and development can lead to long-term defects in inhibitory and excitatory synaptic connectivity [261]. Unlike previous studies [262] that have primarily focused on excitatory synaptic studies, Favuzzi and colleagues found that microglia selectively prune inhibitory synapses, but not excitatory synapses, and that disruption of this process can lead to permanent defects in inhibitory connectivity [261]. Additionally, glial cells modulate another key pathogenic factor in ASD that has a significant impact on synaptic plasticity. Glial cells have been shown to play a critical role in maintaining brain homeostasis under both physiological and pathological conditions by modulating neuronal activity and synaptic plasticity through changes in synaptic coverage, expression of neurotransmitter receptors, and release of neuroactive substances, with broad perisynaptic distribution enabling them to perform these functions effectively (see Fig. 3c) [263]. In their early experiments, Roumier et al. used gene editing technology to conduct animal experiments. They found that mice with defects in the transmembrane peptide KARAP/DAP12, which is expressed only in microglia, exhibited altered synaptic function and plasticity, including enhanced hippocampal LTP and a significant reduction in synaptic expression of the BDNF receptor tyrosine kinase receptor B [264]. Furthermore, long-term injection of lipopolysaccharide into the fourth ventricle of rats resulted in chronic neuroinflammation caused by microglial activation, which significantly attenuated LTP in the dentate gyrus, ultimately leading to impaired spatial memory [265]. These findings emphasize that glial cells, primarily microglia, can impede synaptic plasticity through an inflammatory response, thereby hindering normal neuronal communication. Beyond the aforementioned mechanisms, the microbiota-gut-brain axis (MGBA) plays a pivotal role in ASD-associated neuroinflammation by regulating central nervous system function through immune, metabolic, and neural pathways [266]. For instance, the metabolites of gut microbiota, such as short-chain fatty acids (SCFAs), have been demonstrated to modulate microglial activation and synaptic pruning in a concentration-dependent manner. Concurrently, the microbiota exerts influence over the neuroinflammatory microenvironment through the Th17/Treg balance and immune cell migration, while remotely regulating central immune and neural functions by transmitting signals—including tryptophan derivatives—via the vagus nerve or blood-brain barrier [267, 268]. It is worth noting that while no definitive conclusions have yet been reached, existing evidence suggests neuroinflammation may serve as an early driver of neurodevelopmental disruption in some ASD cases, rather than merely a secondary/concomitant feature. For instance, prospective cohort studies have identified transcriptional alterations in autoimmune-associated genes prior to symptom onset in high-risk infants, with pro-inflammatory pathways (such as those linked to systemic lupus erythematosus gene sets) showing enrichment trends [269]; Maternal immune activation models further demonstrate that prenatal inflammation can induce ASD-like phenotypes in offspring, accompanied by elevated pro-inflammatory cytokines (e.g., TNF-α, IL-6) in brain regions including the hippocampus and cerebellum [270]. However, neuroinflammation may also arise from primary synaptic or metabolic abnormalities, necessitating further longitudinal studies (e.g., integrating infant PET glial imaging with multidimensional biomarker tracking) to clarify its causal temporal sequence [271]. In summary, neuroinflammation can serve as both an evaluative indicator for the diagnosis of ASD and a potential target for therapeutic strategies. However, it is important to note that existing studies have primarily confirmed the presence of neurological immune dysfunction in individuals with ASD. The relationship between neuroinflammation and ASD remains unclear, and whether neuroinflammation is a contributing factor or a consequence of ASD requires further investigation.

Fig. 3.

The Mechanisms of Action of Neuroglial Cells in Glutamate Signaling and Synaptic Plasticity. (a) Astrocytes participate in the uptake and release of glutamate within the synaptic cleft. The glutamate uptake process reveals that two types of glutamate transporters, EAAT-1 and EAAT-2, located on the membrane of astrocytes, facilitate the uptake of glutamate from the synaptic cleft into the astrocytes. VGLUT1 and VGLUT2, expressed by astrocytes, regulate the transport of glutamate from the cytoplasm into vesicles. The activation of GPCRs leads to the generation of IP3. IP3 activates the endoplasmic reticulum, resulting in the release of Ca2+. The elevation of Ca2+ concentration is sensed by synaptotagmin 4, 7, or 11, which triggers the fusion of vesicles with the cell membrane and consequently leads to the release of glutamate within the vesicles to the extracellular space. (b) Microglia participate in glutamate signaling through the Xc- system. Activation of microglia results in the release of ROS and the subsequent activation of NF-κB via the MyD88 pathway. The activated NF-κB subsequently translocates into the nucleus, where it binds to specific binding sites in the promoter region of the Xc- gene, facilitating Xc- gene transcription, increasing the expression of Xc-, and influencing the transport of cystine and cysteine. (c) Glial cells influence synaptic plasticity by altering the degree of synaptic coverage around synapses and releasing neuroactive substances. EAAT: Excitatory Amino Acid Transporters; VGLUT: Vesicular Glutamate Transporter; IP3: Inositol 1,4,5-Trisphosphate; GPCR: G-Protein-Coupled Receptor; NMDRR: N-Methyl-D-Aspartate Receptor; ROS: Reactive Oxygen Species; TRL4: Toll-Like Receptor 4; Xc-: Cystine/Glutamate Antiporter; MyD88: Myeloid Differentiation Primary Response Gene 88; NF - κB: Nuclear Factor - kappa B

In a recently published study, the authors discovered that the hippocampus of rats with an ASD model induced by prenatal valproic acid exposure exhibited elevated inflammatory factors and over-activation of microglia. Furthermore, the intervention using rTMS was observed to improve the autism-like abnormal behaviors in the ASD rat model. The underlying mechanism may be attributed to the significant reduction in neuroinflammation and restoration of synaptic plasticity by down-regulating NF-κB activation in microglia, thereby exerting neuroprotective effects [57]. Similarly, other studies have demonstrated that TMS modulates glial cell polarization and further ameliorates the inflammatory microenvironment, thereby achieving suppression of CNS inflammation. A report combining in vitro and ex vivo experiments confirmed that rTMS stimulation intervention at 10 Hz significantly inhibited neurotoxic polarization of astrocytes after oxygen-glucose deprivation/reoxygenation and cerebral ischemia/reperfusion injury, attenuating neuronal injury, and promoting synaptic plasticity to exert a neuronal protective effect [272]. In addition to regulating astrocytes, TMS may also play a role in balancing the polarized state of microglia. The research team from Fudan University demonstrated that iTBS can inhibit pro-inflammatory M1-type activation and promote anti-inflammatory M2-type activation in the peri-infarct region, but not in the core region. This is achieved by inhibiting the TLR4/NFκB/NLRP3 signaling pathway, mediating the balance of microglial cell M1/M2 phenotype, and thus attenuating the motor deficits and pyroptosis caused by cerebral ischemia/reperfusion injury. Subsequent authors utilized a CSF1R inhibitor to deplete microglia, which nullified the ameliorative effects of iTBS on motor function [273]. Another study found that 25 Hz rTMS intervention ameliorated neuroinflammation, enhanced synaptic plasticity, and inhibited neuronal apoptosis in 3xTg Alzheimer’s disease mice by inhibiting microglial activation and activating the PI3K/Akt/GLT-1 pathway, including increasing PI3K/Akt activity and GLT-1 expression, which is the major transporter protein for removing excess glutamate from the synaptic cleft in rodents. Additionally, rTMS decreased amyloid beta 1–42 levels in hippocampal brain regions, improved oxidative stress and glucose metabolism, cognitive functions, and produced various types of neuronal protective effects. However, these effects were not observed when the PI3K/Akt inhibitor LY294002 was used [274]. On the other hand, TMS modulation can cause changes in cytokine levels, either directly or indirectly. For example, Cha et al. conducted a clinical trial that demonstrated the efficacy of high-frequency rTMS intervention in improving cognitive function and reducing mRNA levels of pro-inflammatory cytokines (IL-1β, IL-6, TNF-α, and transforming growth factor-β) in blood samples from stroke patients, and found a correlation between the reduction in IL-6 levels and scores on the complex figure copy test and auditory verbal learning test [275]. Similarly, animal experiments have shown that iTBS intervention significantly reduces the content of pro-inflammatory cytokines, such as IL-1β, IFN-γ, TNF-α, and IL-17 A, while increasing the level of the anti-inflammatory cytokine IL-10 in brain tissues of cerebral ischemic mice. This phenomenon is associated with the promotion of microglial cells from M1 to M2 phenotype [273]. It is important to acknowledge that the majority of current studies have focused on a single type of glial cell. However, it is well established that different glial cells can interact with each other. Consequently, it is not yet clear which type of cell is the most appropriate for intervention by TMS. Furthermore, there is a lack of sufficient experimentation and validation in the field of autism. Of particular significance is the current absence of direct evidence demonstrating that TMS can modulate neuroinflammatory markers in individuals with ASD. The existing mechanistic data supporting the anti-inflammatory effects of TMS are almost exclusively derived from animal models of stroke, cerebral ischaemia, or Alzheimer’s disease (refer to Table 3). Consequently, it is not possible to extrapolate these findings directly to the ASD population. In summary, it is evident that neuroinflammation represents a potential therapeutic target for ASD diagnosis and treatment. Although TMS has been demonstrated to suppress central neuroinflammation and exert neuroprotective effects, its anti-inflammatory efficacy in ASD remains hypothetical and requires validation through direct experimental evidence in ASD patients or specific ASD animal models.

Table 3.

Summary of TMS regulated neuroinflammation related research

Species/Model	TMS protocol parameters	Inflammatory Markers Altered	Behavioral/Functional Outcomes	Implicated Signaling Pathways	Reference
SD rat Valproic acid-induced autism model	1 Hz rTMS, 900 pulses; 1 time/day, 2 weeks	Reduce TNF-α, IL-1β, IL-6; Inhibit microglial overactivation; Increase IL-10	Relieve autism-like/anxiety-like behaviors; Improve cognition; Ameliorate hippocampal synaptic plasticity	NF-κB signaling pathway	[57]
SD rat MCAO model & OGD/R primary astrocyte models	1/5/10 Hz rTMS (10 Hz optimal), 600 pulses, 1 time/day, 1 week	Reduce TNF-α, IL-1β, IL-12, IL-23, C3, iNOS; Increase IL-10, IL-1ra, Arg1, S100A10;	Reduce infarction volume, neuronal apoptosis; Improve neurological function; Enhance spatial learning/memory; Promote synaptic plasticity;	NF-κB/STAT3 signaling pathway	[272]
C57BL/6 J mouse MCAO/r model	iTBS (10 × 50 Hz bursts, 3 pulses/burst, 20 repeats at 5 Hz intervals), 2 times/day, 1 week	Reduce IL-1β, IL-17 A, TNF-α, IFN-γ, CD86, iNOS; Increase IL-10, CD206, Arg1;	Reduce cerebral infarction volume, inhibit neuronal pyroptosis; Improve motor function and gait, enhance spatial learning/memory; Promote microglial M2 polarization;	TLR4/NFκB/NLRP3 signaling pathway	[273]
3xTg-AD mouse model	25 Hz rTMS, 1000 pulses, 60% max output, 1 time/day, 3 weeks;	Reduce IL-6, IL-1β, TNF-α, ROS, MDA; Increase SOD, GSH	Improve cognitive function, brain glucose metabolism; Enhance synaptic plasticity; Reduce Aβ1–42 levels, neuronal apoptosis;	PI3K/Akt/GLT-1 signaling pathway	[274]
Post-stroke cognitive impairment patient	20 Hz rTMS, 2000 pulses, 1 time/day, 5 days/week, 2 weeks;	Reduce the expression of IL-1β, IL-6, TNF-α, TGF-β, mRNA	Improved cognitive and motor function, significantly enhanced activation of cognitive related brain regions	Anti-inflammatory and brain network regulation pathway	[275]

MCAO: Middle Cerebral Artery Occlusion; OGD/R: Oxygen-Glucose Deprivation/Reperfusion; AD: Alzheimer’s Disease; TNF-α: Tumor Necrosis Factor-α; IL-1β: Interleukin-1β; IL-6: Interleukin-6; iNOS: Inducible Nitric Oxide Synthase; ROS: Reactive Oxygen Species; MDA: Malondialdehyde; SOD: Superoxide Dismutase; GSH: Glutathione; NF-κB: Nuclear Factor-kappa B; STAT3: Signal Transducer and Activator of Transcription 3; TLR4: Toll-like Receptor 4; NLRP3: NOD-like Receptor Pyrin Domain-containing 3; IFN-γ: Interferon-γ; CD86: Cluster of Differentiation 86; rTMS: Repetitive Transcranial Magnetic Stimulation; iTBS: Intermittent Theta-Burst Stimulation; PI3K: Phosphatidylinositol 3-Kinase; Akt: Protein Kinase B; GLT-1: Glutamate Transporter 1; C3: Complement 3; IL-1ra: Interleukin-1 Receptor Antagonist; Arg1: Arginase 1; S100A10: S100 Calcium Binding Protein A10; Aβ1–42: Amyloid β 1–42; SD: Sprague-Dawley

Gut microbiome

Trillions of microorganisms, such as bacteria, viruses, fungi, and other life forms, inhabit the human body. Different classes of microbes are present in different organs, and those in the gut are of particular interest in biomedical research [276]. The gut microbiome is a diverse group of microorganisms that live in symbiosis with the host and interact with it. They play a role in various host functions, such as nutrient absorption, colonization resistance, immune function modulation, and intestinal barrier maintenance [277, 278]. Previous research has demonstrated that gut microbiome has a significant impact on the physiological functions of the host, both directly and indirectly, through self-produced or modified metabolites [279], the nervous system [280, 281], immunomodulation [282], hypothalamic-pituitary-adrenal (HPA) axis [283]. Therefore, it is vital to maintain a balance of gut microbiome for the host’s health.

In recent decades, studies have confirmed interactions between the gut microbiome and the brain in individuals with autism or other neuropsychiatric disorders. ASD has also been recognized as a brain-gut-microbiome axis disorder [284, 285]. The MGBA theory explains the communication between the gut microbiome and the CNS through various pathways, including immune-related, neural, endocrine, and metabolic signaling pathways [286]. Clinical and animal studies have demonstrated that MGBA facilitates bidirectional communication between the gut and the brain, contributes to brain homeostasis, and helps regulate cognitive and emotional functions [281, 287, 288], and that disorders of the gut microbiome can affect neurological function and behavior through MGBA [289]. It is noteworthy that, despite repeated reports from observational studies indicating a significant association between ASD and gut microbiota abnormalities, there remains no conclusive evidence demonstrating that microbiome dysregulation constitutes a direct aetiological factor in the core neuropathology of ASD [290–292]. The relationship between the two is more likely to reflect complex bidirectional interactions and shared genetic or environmental drivers, such as dietary preferences, antibiotic use, or abnormal gastrointestinal motility [290, 291], rather than a simple causal chain [292]. Epidemiological data indicate that approximately 40% of individuals with ASD exhibit pronounced gastrointestinal symptoms, including altered bowel habits, abdominal pain, and gastroesophageal reflux [293, 294]. Moreover, symptom severity frequently correlates positively with the prevalence of gastrointestinal issues [293]. This comorbidity has prompted researchers to investigate the structural characteristics of the gut microbiome in ASD patients. For instance, Li et al. utilised 16 S rRNA gene sequencing to compare faecal samples from children with ASD and healthy controls, revealing significant differences in microbial composition between the two groups. The ASD cohort exhibited enrichment in specific bacterial families such as Alcaligenaceae, Enterobacteriaceae, and Clostridium [295]. This finding has been validated in multiple independent studies, with elevated Clostridium abundance being particularly consistent [296, 297]. Interestingly, Clostridioides, a subspecies of Clostridium, is one of the most frequently detected dysbiotic bacteria in patients with ASD [298] due to its production of potentially toxic metabolites such as 4-ethylphenyl sulfate and p-Cresol sulfate, which are thought to enter the bloodstream and cross the blood-brain barrier to influence processes such as neuroinflammation and microglial phagocytosis [268]. In addition, Clostridium tetani has been found to release transporting tetanus neurotoxin and subsequently translocate it to the CNS to disrupt neurotransmitter release, thereby inducing a wide variety of behavioral deficits in ASD [299]. Beyond bacterial toxins, other microbially derived molecules also participate in the pathological processes of ASD. For instance, SCFAs (such as propionic acid, butyric acid, and acetic acid) are products of gut microbiota fermentation of dietary fibre, undigested starch, and amino acids. At physiological concentrations, they exert anti-inflammatory effects, maintain intestinal barrier integrity, and regulate immunity. However, dysbiosis in ASD patients leads to abnormal elevation of SCFAs like propionic acid. Clinical studies confirm significantly increased propionic acid concentrations in faeces from children with ASD, correlated with symptom severity. Animal studies demonstrate that intraventricular injection of propionic acid induces ASD-like behaviours, triggering glial activation, mitochondrial dysfunction, and oxidative stress [300]. It may also interfere with neurodevelopmental gene expression by inhibiting histone deacetylases or activating neuroinflammatory pathways via free fatty acid receptors (e.g., GPR41/FFAR3) [301]. Another pivotal mechanism involves immune-mediated neuroinflammation. Gut dysbiosis disrupts the intestinal epithelial barrier (‘leaky gut’), allowing bacterial lipopolysaccharides and other substances to enter the circulation. This activates the peripheral immune system and releases inflammatory mediators such as IL-6, TNF-α, and IL-17a [302]. These factors can activate central microglia and astrocytes via the compromised blood-brain barrier or vagal signalling. Post-mortem brain tissue from ASD patients and animal models consistently shows reactive glial cell proliferation and morphological alterations, with their abnormal activation exacerbating neuroinflammation [245, 303]. This ultimately leads to abnormal synaptic pruning, neurotransmitter imbalances, and disrupted neural circuit function, which constitutes a core neurobiological feature of ASD [303]. Furthermore, inflammatory mediators can further increase blood-brain barrier permeability, facilitating peripheral immune cell infiltration into the central nervous system and forming a ‘peripheral inflammation – central neuroinflammation’ cascade reaction [245]. In summary, existing evidence strongly supports the notion that the gut microbiota plays a regulatory or promoting role in the pathogenesis of ASD, and its intervention may become an important adjunct to comprehensive treatment strategies for ASD.

Although there is limited research on TMS interventions with gut microbiome, researchers have tentatively demonstrated that TMS may exert its therapeutic effects by correcting disturbed gut microbial compositions or modulating specific bacterial species. The research team from the University of Milan utilized deep TMS (dTMS), a type of rTMS protocol that delivers a magnetic field through a special H-shaped coil wrapped in a helmet to stimulate deeper brain regions, to investigate the effects on the gut flora of obese patients. The results suggest that high-frequency dTMS protocols can effectively modulate the composition of the gut microbiome of obese subjects, reverse obesity-associated microbiome changes, and promote representative bacterial species with anti-inflammatory properties, such as Faecalibacterium [304]. In a separate study, it was discovered that low intensity rTMS intervention increased the abundance of the anti-inflammatory Roseburia spp. This increase was significantly correlated with behavioral data from forced swimming experiments and MRI results. Additionally, the KEGG functional annotation of the rTMS-intervention group showed that apoptotic pathway abundance was the only indicator of a decrease. These findings suggest that rTMS intervention may have anti-inflammatory and protective effects on the gut microbiome, which are associated with its therapeutic effects [305]. A recent study showed that intervention with 15 Hz rTMS attenuated depressive-like behavior in a model of depression mice and modulated the abundance of Cyanobacteria, Proteobacteria, and Actinobacteriota phylum, which are associated with neurotransmitters and gut inflammation, as well as the levels of polyunsaturated fatty acids in plasma and brain tissue [306]. It should be emphasized that the above studies did not focus on the ASD population or ASD specific models, and their conclusions need to be cautiously extrapolated to ASD scenarios.

Then, what are the potential possible mechanisms underlying the intervention of TMS on gut microbiome? As illustrated in Fig. 4, it is widely acknowledged that TMS can ameliorate neuroinflammation by modulating neuroimmune signalling. Importantly, TMS further influences the gut microbiome through bidirectional communication along the MGBA, with the vagus nerve serving as a key pathway. Studies confirm that TMS can also modulate central noradrenergic system activity, altering central noradrenaline release levels. This central change then impacts the gut microenvironment via descending sympathetic pathways—moderately regulated central noradrenaline levels reduce overgrowth of pathogenic gut bacteria, creating stable conditions for colonisation by beneficial microbiota such as short-chain fatty acid-producing bacteria [307]. Concurrently, TMS modulates the central adenosine signalling system: as a key neuro-immune modulator, adenosine’s A₁R and A_2AR receptors are widely expressed in the basal ganglia (e.g., caudate nucleus) and brainstem. iTBS restores A₁R/A_2AR equilibrium, thereby inhibiting neuroinflammation mediated by excessive adenosine-A_2AR signalling in basal ganglia regions [308]. Through bidirectional regulation via the MGBA, TMS-induced central neurochemical alterations may further influence intestinal mucosal immunity and metabolite levels (e.g., SCFAs), thereby indirectly modulating gut microbiota composition and function [307, 308]. Previous studies have demonstrated that noradrenaline levels decrease following five weeks of dTMS treatment, with this change significantly correlated to alterations in the abundance of Bacteroides, Eubacterium, and Parasutterella [304]. This phenomenon is mediated through a pathway involving both the dopaminergic reward system and HPA axis regulation: dTMS first activates the dopaminergic reward system (including the striatum, ventral tegmental area, and nucleus accumbens), while simultaneously triggering systemic regulatory responses via the HPA axis. These dual effects collectively lead to a reduction in local intestinal noradrenaline levels, thereby inhibiting the proliferation of harmful gut pathogens and diminishing their virulence [309], ultimately exerting beneficial effects on the composition of the gut microbiota. However, this mechanism has yet to be validated in ASD-related research.

Fig. 4.

TMS restores the disordered gut microbiome by influencing neuroinflammation. Given the mechanism by which gut microbiome can communicate with and have an impact on glial cells in the brain through pathways such as immune mediation, metabolites, and neural regulation, TMS acts on the polarization of glial cells to ameliorate neuroinflammation, thereby achieving the regulation of gut microbiome. TMS, Transcranial Magnetic Stimulation

It is noteworthy that the aforementioned studies indicate TMS intervention may lead to alterations in the abundance of specific bacteria rather than affecting the overall diversity of the bacterial community. This suggests that TMS’s regulatory effect does not broadly influence the entire gut microbial system but instead specifically targets particular bacterial species. However, existing gut microbiome research suffers from numerous limitations that severely undermine the reliability and persuasiveness of its evidence: most studies are small-scale exploratory investigations, making it difficult to rule out interference from individual variations and reducing the generalisability of conclusions; considerable heterogeneity exists across studies regarding TMS stimulation frequency, intensity, and target sites, coupled with a lack of systematic investigation into parameter-effect relationships, preventing the identification of optimal parameter combinations for regulation; Intervention cycles typically span 4–8 weeks with short follow-up periods, hindering assessment of long-term effects and post-treatment recovery to baseline microbiome composition. Methodological flaws further undermine credibility, including insufficient sequencing depth, inadequate control of confounding factors (diet/lifestyle/concurrent medication), and absence of negative controls or randomised designs. More critically, these limitations, compounded by the unclear associations between specific bacterial relative abundance changes and TMS parameters, target sites, and disease types, collectively result in the intrinsic mechanisms of neuro-endocrine-immune-microbiome cross-regulation remaining unexplained. This also implies that the current evidence regarding the association between TMS and the gut microbiome remains exploratory in nature, insufficient to draw definitive clinical conclusions, and there is as yet no direct evidence confirming the existence of this association in ASD. Consequently, future research must employ large-sample, multicentre, randomised controlled designs, standardise TMS parameter settings, extend follow-up periods, and optimise microbiome detection methodologies, with a primary focus on exploring the aforementioned associations and intrinsic regulatory mechanisms. Concurrently, given the scientific plausibility of the hypothesis that TMS improves ASD and other central nervous system disorders alongside associated gastrointestinal dysfunction by modulating bidirectional gut-brain axis communication, future research should incorporate investigations into this application. This not only holds significant scientific value but also represents a core pathway for enhancing the credibility of research evidence in this field and advancing clinical translation.

Limitation

This study provides a preliminary overview of the application of TMS research protocols in the field of ASD, covering common stimulation patterns, coil selection, and target planning and localisation. However, given that TMS effects are influenced by multiple factors and involve interdisciplinary knowledge systems, this paper does not delve deeply into its potential mechanisms or parameter optimisation, nor does it systematically assess the risk of bias and methodological quality of the included studies. Significant heterogeneity exists across studies in terms of sample size, blinding design, and control group configuration. This may compromise the assessment of TMS’s true efficacy and its underlying mechanisms. Future research should systematically explore these aspects using standardised tools such as the Cochrane risk of bias assessment tool to deepen our comprehensive understanding of factors influencing TMS efficacy. It is noteworthy that most current studies employ a single frequency/intensity protocol, and the dose-response and time-dependency relationships remain insufficiently unexplored. In particular, it remains unclear whether the modulation of E-I balance or ion-channel states exhibits linear, threshold-dependent, or non-monotonic responses to TMS parameters, including frequency, intensity, number of pulses per train, inter-train interval, and total session duration. Moreover, the pathophysiological mechanisms of ASD exhibit considerable heterogeneity, encompassing multidimensional factors including genetics, epigenetics, neuroscience (such as neurobiology and brain networks), immunology, and environmental influences. Given the unclear effects of TMS on non-neurobiological factors, this study focuses on several classic neurobiological hypotheses of ASD. Although limited in scope, this approach facilitates deeper analysis of their intrinsic relationships, providing a more refined theoretical perspective on TMS intervention mechanisms. It is worth noting that this review has not yet fully explored personalised TMS strategies in ASD. Future investigations should determine how inter-individual variability—such as specific genetic backgrounds, baseline neurophysiological profiles (e.g., E-I ratio measured by TMS-EEG), and synaptic plasticity status—influences TMS responsiveness. Personalized protocols may maximize efficacy and minimize adverse effects. In terms of research classification, this paper categorises all TMS protocols under a single umbrella to preliminarily explore their regulatory mechanisms in ASD. It must be emphasised that this work represents an exploratory phase; subsequent research will undertake detailed categorisation and systematic comparison of different intervention protocols to consolidate the theoretical and practical foundations for TMS’s clinical application in ASD.

Conclusion

A synthesis of extant research findings demonstrates the potential value of TMS as an intervention for ASD. However, the current application of TMS is beset by multiple uncertainties, necessitating objective consideration. On the one hand, studies have indicated suboptimal efficacy outcomes, which are closely linked to variations in TMS stimulation protocols and the highly complex heterogeneity of ASD, and studies with differing levels of bias risk report divergent findings regarding efficacy. On the other hand, existing research has predominantly focused on clinical efficacy observations, with markedly insufficient exploration of the fundamental mechanisms underlying TMS intervention for ASD. The classical mechanisms currently proposed have largely been derived from studies of other neurological disorders and do not specifically align with core pathological pathways in ASD. Consequently, these fundamental mechanisms offer limited reference value for optimising clinical protocols, hindering the formation of a closed-loop guidance system linking mechanism, protocol, and efficacy.

Consequently, future research should prioritise the following directions: Firstly, efforts must be made to address the existing gaps in fundamental mechanism studies. To this end, the present study proposes a shift in research focus from generalised classical mechanisms to a more specific investigation of the influence of TMS on core pathological pathways in ASD. The clarification of its specific action principles will provide scientific grounds for the optimisation of stimulation protocols, personalisation of target planning, and the achievement of precise localisation. Secondly, it is necessary to conduct rigorously designed large-scale, multicenter randomized controlled trials, strictly control the risk of bias, systematically evaluate the efficacy and safety of TMS in different subtypes, age groups, and symptom severity populations of ASD, clarify the dose-response relationship and duration of efficacy, and gradually establish precise treatment strategies for specific symptoms. It is only through such systematic exploration that the full potential of TMS as an intervention can be realised, thus offering evidence-based alternatives for clinical interventions in children diagnosed with ASD.

Abbreviations

ASD

Autism Spectrum Disorder

TMS

Transcranial Magnetic Stimulation

CNS

Central nervous system

E-I

Excitatory-inhibitory

sTMS

Single-pulse Transcranial Magnetic Stimulation

M1

Primary motor cortex

pTMS

Paired-pulse Transcranial Magnetic Stimulation

rTMS

Repetitive Transcranial Magnetic Stimulation

TBS

Theta burst stimulation

EEG

Electroencephalogram

iTBS

Intermittent theta burst stimulation

cTBS

Continuous theta burst stimulation

GABA

Gamma-aminobutyric acid

DLPFC

Dorsolateral prefrontal cortex

MRI

Magnetic Resonance Imaging

pSTS

Posterior superior temporal sulcus

IFG

Inferior frontal gyrus

TPJ

Temporoparietal junction

TMAS

Transcranial magneto-acoustic stimulation

NMDA

N-methyl-D-aspartate

LTP

Long-term potentiation

LT

Long-term depression

AMPA

α-amino-3-hydroxy-5-methyl-4-isoxazole propionic acid

STDP

Spike timing-dependent plasticity

BDNF

Brain-derived neurotrophic factor

PV

Parvalbumin

PAC

Phase-amplitude coupling

TNF-α

Tumor necrosis factor-α

IL-1β

Interleukin-1β

MCP-1

Macrophage chemotactic protein-1

CSF

Cerebrospinal fluid

ROS

Reactive oxygen species

MGBA

Microbiome-gut-brain axis

HPA

Hypothalamic-pituitary-adrenal

Author contributions

Writing – original draft and drawing: Xingxing Liao; writing – review and editing: Hui Li, Kaiyue Han and Junzi Long; formatting adjustments: Ying Liu, Zhiqing Tang and Jiarou Chen; supervision and conceptualization: Hesheng Liu; funding acquisition and conceptualization: Hao Zhang.

Funding

This study was supported by the Program of the China Disabled Persons’ Federation [grant number CDPF2023KF00001] and the Sub-project of the Program of the Ministry of Science and Technology of the People’s Republic of China [grant number 2022HZ-06-01].

Data availability

Not applicable.

Declarations

Ethics approval and consent to participate

Not applicable.

Consent for publication

Not applicable.

Competing interests

The authors declare that they have no competing interests.