Epigenetic profiles of response to transcranial magnetic stimulation in treatment resistant depression

Abstract

Background

Transcranial magnetic stimulation (TMS) is an established treatment for major depressive disorder (MDD), yet response rates remain suboptimal and biomarkers predictive of treatment outcomes are currently lacking. Recently, DNA methylation (DNAm) has shown promise as an epigenetic predictor of antidepressant and electroconvulsive therapy treatment outcomes but no study to our knowledge has characterized DNAm profiles of treatment outcomes in the context of TMS. Here, we present the first genome-scale DNAm analysis of TMS outcomes in patients with treatment-resistant depression (TRD).

Methods

Peripheral blood samples from 60 TRD patients were collected prior to a standard 36-session TMS course. DNAm was profiled using the Illumina EPIC array and filtered to retain only the top 5% most variable probes for subsequent analysis in relation to three treatment outcomes in an analytic sample of 46 patients: treatment response (> 50% PHQ-9 reduction), symptom trajectory (ΔPHQ-9), and remission (PHQ-9 < 5).

Results

Methylated CpG Set Enrichment Analysis (mCSEA) identified 67, 23, and 163 differentially methylated regions (DMRs) associated with treatment response, symptom trajectory, and remission, respectively (FDR < 0.05). Sixteen DMRs were common across all outcomes, implicating genes involved in neurodevelopment (HOXA4, HOXA5), immune signaling (RUNX1, OTUD5), and synaptic function (EFNB1, RAP2C). Targeted analysis of 84 CpGs in the BDNF promoter revealed 10 nominally significant CpGs that differentiated responders from non-responders. Several DMRs showed strong blood–brain methylation concordance (r > 0.5), supporting their potential relevance to central nervous system mechanisms.

Conclusions

Despite the limited sample size, these findings suggest distinct epigenetic signatures prior to treatment that are associated with TMS outcomes, supporting the potential utility of DNAm as a biomarker for response stratification in TRD.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12920-025-02240-2.

Background

Major depressive disorder (MDD) is one of the most prevalent, debilitating, and costly psychiatric disorders affecting ~ 290 million people worldwide [1]. MDD presents with significant symptom heterogeneity characterized by extended periods of a depressed mood, anhedonia, and suicidal ideation, among a range of other symptoms [2]. MDD is frequently recurrent [3, 4] and associated with a significant disease burden imposing substantial direct (e.g., healthcare utilization) and indirect (e.g., lost productivity) costs on individuals, employers, and society at large [5, 6]. Current MDD treatments are significantly limited: only approximately one-third of patients respond to initial antidepressant treatments [7–9], while another 30–50% of patients insufficiently benefit from even multiple antidepressant trials [10, 11] –a condition commonly defined as treatment-resistant depression (TRD) [12, 13].

In treatment-resistant populations, neuromodulation treatments such as electroconvulsive therapy (ECT), vagus nerve stimulation, and transcranial magnetic stimulation (TMS) are effective alternatives to molecular-acting therapies [14–17]. Among these treatment modalities, TMS is among the least invasive option providing relief of depressive symptoms often without significant side effects [18]. TMS treatment involves stimulating the brain via a magnetic field resulting in acute depolarization and altered electrical activity in the underlying neurons. When applied repetitively, TMS alters the excitability of the surrounding region for a sustained period outlasting the TMS exposure and is thought to exert its therapeutic effects through altered cortical excitability [19]. The most common TMS treatment course involves a total of 36 sessions [20] with daily treatments over a number of weeks, applied in an outpatient setting, requiring a substantial time commitment from patients. Although multiple clinical trials [21, 22] and meta-analyses [23, 24] demonstrate the efficacy of TMS in reducing depressive symptoms, only about 40% of patients with TRD respond to TMS [25], highlighting the need to identify characteristics that distinguish those who do vs. do not respond to this treatment modality.

One such characteristic may be differences in epigenetic profiles. As stable but modifiable mediators of environmental effects on the genome, epigenetic mechanisms regulate gene expression and translation via chemical modifications independent of the underlying DNA sequence. As such, the investigation of epigenetic mechanisms in MDD treatment outcomes can provide insights into the interaction of genetic and environmental factors. DNA methylation (DNAm), i.e., the addition of a methyl group to cytosine bases in CpG dinucleotide sites [26], is already widely utilized as a biomarker in clinical evaluation in oncology [27], cardiovascular disease [28] and neurodevelopmental disorders [29]. More recently, DNAm has been increasingly investigated in psychiatric disorders and has emerged as a potential predictive molecular biomarker in MDD treatment response [30]. Such emerging studies profile peripheral DNAm in patients before treatment, or longitudinally, to link differential methylation to treatment outcomes [31–33]. The assessment of pre-treatment DNAm can thereby be utilized to characterize epigenomic variation associated with treatment outcomes and in turn, inform the potential development of tools with predictive utility [30]. The majority of epigenetic studies of MDD treatment outcomes to date focus on candidate genes implicated in the context of MDD, with only a limited number of emerging genome-scale efforts [30]. A recent review of studies characterizing DNAm profiles associated with MDD identified only one locus— Brain-derived neurotrophic factor (BDNF)—that showed some evidence for replication across studies [30]. BDNF is a neurotrophic growth factor that is involved in neuronal growth, maintenance, and plasticity- key mechanisms identified to be dysregulated in MDD [34]. Ample evidence documents reduced peripheral BDNF levels in acutely [35] and chronically [36] depressed patients- findings that may underly differences in DNAm [37]. In relation to treatment outcomes, BDNF promoter methylation has been associated with differential antidepressant treatment and ECT outcomes but not in TMS [30].

Collectively, the vast majority of existing studies investigate DNAm patterns in relation to pharmacological treatment outcomes for MDD [32, 38, 39]. To date, there is a relative dearth of studies investigating DNAm patterns in relation to other treatment modalities such as psychotherapy and brain stimulation therapies [30].

To our knowledge, this is the first study to characterize genome-scale DNAm in the context of TMS treatment outcomes to identify possible predictors of TMS treatment response. We compared pre-treatment DNAm profiles in relation to three treatment outcomes following completion of at least 36 sessions of TMS treatment in patients with TRD. In addition, we performed a hypothesis-driven analysis of BDNF promoter methylation to examine its association with treatment outcomes following TMS in patients with TRD.

Methods

This study was approved by the institutional review board of the University of South Florida. Data privacy was maintained in compliance with the Health Insurance Portability and Accountability Act of 1996 (HIPAA). All study participants provided written, informed consent after a detailed explanation of the study procedures.

Participant demographics and assessment of MDD

A total of 60 patients over the age of 18 diagnosed with MDD according to the DSM-V were recruited at two recruitment sites, the TMS of South Tampa Clinic and the Neurotherapies Clinic at the University of South Florida in Tampa, Florida, between 2020 and 2023. Patients were assessed for symptoms of MDD with the Patient-Health Questionaire-9 (PHQ-9) [40] at baseline, 4.5 week into the TMS treatment and at the 9 week end of treatment. Demographic characteristics, current and past psychiatric diagnoses, medication history, and current use of medication in addition to alcohol, smoking, and marijuana use were assessed via questionnaires administered by study staff at baseline. In addition, medication prescriptions during the course of the TMS treatment were derived from the patients’ electronic health records. Medications were classified according to the World Health Organization Anatomical Therapeutic Chemical (ATC) classification [41], an internationally recognized standard for drug classification utilized in other epigenetic studies [42]. Meeting the criteria of TRD for inclusion in our study was defined as having two or more failed antidepressants trials. Participants who had undergone TMS treatment within the past three months were excluded from our study. Patients with a history of ECT or any other neuromodulation treatment were not specifically excluded from the study. However, to our knowledge, none of the enrolled participants had previously undergone ECT or any other neuromodulation treatment. A minimum of 36 completed TMS session was required to be included in the analysis, leading to the exclusion of five study participants. Additionally, study participants (n = 9) with PHQ-9 scores below at least moderately severe MDD symptoms (PHQ-9 ≥ 15) were excluded from our primary analysis consistent with depression measure exclusion cutoffs applied in previous treatment-outcome DNAm studies [32, 38], leaving a total analytic sample of 46 study participants. An overview of our study design is presented in Fig. 1.

Fig. 1.

Overview of Study Design. TRD patients undergoing TMS treatment (n = 60) were assessed for depressive symptoms from the pretreatment baseline to the end of treatment after 9 weeks of treatment. Blood spots were collected for genome-scale DNAm analyses at the baseline measure. From the 60 patients initially enrolled, five were excluded for completing fewer than 36 TMS sessions and nine were excluded for baseline PHQ-9 scores below the threshold for at least moderately severe depression (PHQ-9 ≥ 15), yielding a final analytic cohort of 46 participants

Transcranial magnetic stimulation treatment

TMS treatments at both participant recruitment sites were delivered with the NeuroStar^® TMS system (Malvern, PA, USA) operated by site personnel. Briefly, the treatment is administered by placing the participants in a recliner, and ear plugs are inserted to minimize possible hearing impairment from the repetitive machine noise. The TMS treatment parameters are captured automatically including information on the session date, treatment location, motor thresholds (MT), number of administered pulses per treatment, pulse duration, intertrain interval, and number of treatment sessions. Resting MT threshold is determined at the first treatment session via single pulses stimulation of the participant’s motor cortex area corresponding to the abductor pollicis brevis. The Beam F3 method and 5 cm rule was used for coil placement over the dorsolateral prefrontal cortex (DLPFC) target. Patients treated with high frequency left DLPFC receive 10 Hz stimulation at 120% of the estimated motor threshold as 4 s of stimulation with 11 s intertrain interval for a total of 3000 pulses (N = 12; 26.1%) and patients treated with low frequency right DLPFC stimulation (N = 5; 10.9%) receive 1 Hz stimulation at 110% of the estimated motor threshold as 1 s of stimulation with no interval for a total of 3000 pulses. Patients receiving both high frequency left, and low frequency right were characterized as sequential bilateral (N = 12; 26.1%). The remaining participants received a “combination” of protocols i.e. switching from left or right DLPFC to sequential bilateral (N = 17; 37.0%). Standard of care treatment consists of 36 sessions of left DLPFC rTMS at 10 Hz. Protocols were characterized accordingly when 80% or more treatment sessions met the criteria for a specific protocol [43]. Treatments are delivered one session per day Monday through Friday. The various TMS protocols used in this study, including high-frequency left-sided and low-frequency right-sided stimulation, have demonstrated therapeutic equivalency in the treatment of depression, as supported by prior literature [44]. Patients are treated five days a week for up to nine weeks. Supplementary Table S1 provides additional summary information on TMS parameter characteristics.

Blood sampling and DNAm profiling

Blood samples were obtained from the participants’ fingertips using a sterile single-use lancet (BD Microtainer) at their last consultation appointment before the start of their TMS treatment. Drops of blood were applied to filter paper (Whatman 903 Protein Saver Card), dried overnight at room temperature in a biosafety cabinet, and subsequently stored at −30 °C for long-term storage as in our prior work [45]. DNA was extracted from ten 3 mm blood punches derived from the filter paper, per sample, using the QIAamp DNA Micro Kit (Qiagen, Hilden, Germany) and stored at −30 °C. Samples with ≥250 ng of genomic DNA were bisulfite converted with the EZ-96 DNAm Kit (Zymo Research, Irvine, CA, USA). DNAm of bisulfite-converted DNA of the participants’ samples was profiled using the Illumina Infinium MethylationEPIC Beadchip (Illumina, CA, USA) at the molecular genomics core of the Moffitt Cancer Center (Tampa, Florida) using the Illumina iScan System (Illumina, CA, USA).

Data preprocessing and quality control

Raw DNAm data were preprocessed and quality controlled in in R version 4.0.5 (https://cran.r-project.org/) using standardized consortium-developed pipelines [46]. Briefly, samples exhibiting probe detection call rates below 90% and average signal intensities either less than 50% of the overall sample mean or below 2000 arbitrary units (AU) were excluded from further analysis. Probes with detection p-values > than 0.01 were deemed low quality and subsequently treated as missing data. Probes missing in more than 10% of samples, as well as probes known to be cross-hybridizing [47], were also removed. Checking for potentially sex-discordant samples as well as the annotation of probes affected by genetic variation, was performed with the R package minfi [48] and ewastools [49], respectively. Data normalization was performed using the single-sample Noob (ssNoob) method implemented in the minfi R package. Quality control removed a total of 1,203 CpGs due to low signal intensity or missing values along with 44,260, cross reactive probes, resulting in 820,628 probes for subsequent analysis. To correct for batch effects associated with chip position, ComBat adjustment was performed using a Bayesian framework within the SVA R package [50] while preserving variation attributable to treatment outcomes, age and sex. The genomic inflation factor (λ) was calculated using the same statistical model and covariates as our main linear model outlined below. Due to our λ = 1.04 being close enough to 1.0 we did not apply any additional methods to control for genomic inflation. After quality control all 46 samples remained for subsequent analysis. A detailed graphical overview of our bioinformatics workflow is available in the supplementary materials (Supplementary Fig. F1).

Estimation of covariates

Blood cell-type composition—including proportions of CD8 + T cells, CD4 + T cells, natural killer (NK) cells, B cells, monocytes, and neutrophils—was estimated using the IDOL algorithm implemented in the Epidish [51] employing reference data optimized for the EPIC array [52]. We calculated smoking scores from DNAm data based on the weights obtained from 39 CpGs located at 27 loci [53] to account for smoking status in our analysis. Ancestry principal components (PCs) were generated from DNAm implementing the method described by Barfield et al. [54] but not included as a covariate due to the lack of diversity in our participant population. Prescription-derived antidepressant use data was extracted from participants’ electronic health records and an antidepressant use variable was generated and incorporated as a covariate in the models. Antidepressants use variable was considered among patients having a prescription for selective serotonin reuptake inhibitors (SSRIs; ATC N06AB), non-selective monoamine reuptake inhibitors (e.g. TCAs; ATC N06AA), or other antidepressants (ATC N06AX) as described by others [42] to account for the impact of antidepressant use on baseline DNAm [42, 55].

Statistical analysis and differentially methylated regions (DMRs)

Given our modest sample size, we focused our analysis on the top 5% variable probes (CpGs) across the EPIC array in the final analytic sample of 46 patients, following quality control. DMRs associated with our treatment outcomes of interest were identified using Methylated CpGs Set Enrichment Analysis (mCSEA) [56] and focusing on promoter regions. The mCSEA algorithm is designed to identify DMRs that have a small but consistent change in methylation related to complex phenotypes and is implemented in R Bioconductor. Methylation sites were categorized as occurring in promoters if the UCSC_RefGene_Group column in the IlluminaHumanMethylationEPICanno.ilm10b2.hg19 data package [57] contained the terms: TSS1500, TSS200, 5ʹ untranslated region (UTR), or 1stExon. First, linear or logistic regression models [58] were applied to estimate the association between baseline DNA methylation at each CpG probe and treatment outcomes, adjusting for covariates including age, sex, antidepressant use, baseline PHQ-9 scores, blood cell composition (CD8T, CD4T, NK, B cell, and monocyte proportions), and smoking status. These models yielded either t-statistics (for continuous outcomes) or z-scores (for binary outcomes), which were used to rank probes for downstream enrichment analysis. Baseline depression severity (PHQ-9 score) was included as a covariate in all models due to its biological and clinical relevance to treatment outcomes [59]. In addition, controlling for baseline depression severity offered overall improvement of the linear models determined by Bayesian Information Criterion (BIC) method [60]. Following the linear model, the resulting statistics were used to create a user-supplied ranking list, which ranked probes by the strength and direction of their association with our treatment outcome variables of interest. Subsequently, the ranked probes are utilized in an enrichment analysis identifying differentially methylated promoters from which overrepresented regions that have minimum of five (minCpGs = 5) differentially methylated CpGs are detected as a DMR. The default threshold of five CpGs in mCSEA was used, as this probe density provides sufficient coverage to support robust and interpretable enrichment testing. Given the dense coverage of the Illumina EPIC array and our QC’d dataset, this threshold balances sensitivity and reliability by reducing noise from sparsely annotated regions, while retaining biologically meaningful signals. This approach is consistent with prior studies using mCSEA and similar region-based methylation analyses [61].

Our main treatment outcomes of interest in our DMR analysis included: (1) treatment response (responders vs. non-responders) defined as a >50% decrease in PHQ-9 scores from the baseline measure (T0) to the end of treatment (T2), (2) symptom trajectory assessing whether methylation levels vary with the degree of symptom improvement (capturing the absolute change in PHQ-9 score (ΔPHQ-9) from baseline to end of treatment, and (3) remission status (remitters vs. non-remitters) defined as PHQ-9 score of < 5 following completion of the TMS treatment [62]. Results of the DMR analysis were considered significant at false discovery rate (FDR) adjusted p-value of < 0.05.

Functional enrichment of genes associated with methylation regions

To assess the functional correlation between significant methylation regions and nearby genes across the genome, we conducted a Gene set enrichment analyses (GSEA) performed using the gometh() function from the missMethyl [63] (v1.41.0) package. We first defined a background set of CpGs comprising the top 5% most variable CpG sites across samples that were included in our DMR analysis. We included CpG sites identified as leading probes for genes implicated in the mCSEA results, retaining only those gene sets that were significant at a false discovery rate (FDR) of q < 0.05.

Targeted analysis of methylation differences in TMS outcomes at candidate CpG sites

Baseline differential methylation at targeted CpG sites was further explored through a hypothesis-driven analysis of candidate genes previously implicated in treatment outcomes. Due to converging evidence from prior candidate gene studies [30] linking BDNF promoter methylation to differential MDD treatment outcomes we focused our targeted analysis to probes annotating to this region. Differential methylation of CpG sites annotated to the BDNF promoter was analyzed in relation to our three treatment outcomes: treatment response, symptoms trajectory, and remission. Probes mapping to the BDNF promoter were identified by filtering the Illumina EPIC array annotation data, obtained using the IlluminaHumanMethylationEPICanno.ilm10b4.hg19 R package [57], for probes with ‘BDNF’ in the UCSC_RefGene_Name field and ‘TSS1500’ or ‘TSS200’ in the UCSC_RefGene_Group field. These categories denote probes located within 1500 bp and 200 bp upstream of the transcription start site, respectively, and are widely used to define promoter regions in array-based epigenetic studies [64, 65]. Only those promoter probes present in the methylation dataset were retained for analysis. A design matrix was constructed to model treatment outcomes while adjusting for relevant covariates outlined in our main models. Methylation β-values for the BDNF promoter probes were converted to a numeric matrix and fit to the design matrix using the limma package [58]. Empirical Bayes moderation was applied to stabilize variance estimates. Differential methylation associated with treatment outcomes at the end of treatment was assessed, and p-values were adjusted for multiple comparisons using the Benjamini–Hochberg false discovery rate (FDR) correction [66].

Concordance of DNAm signatures between blood and brain

The BECon tool [67] was used to assess the correlation between DNAm levels in blood and brain among DMRs overlapping the three treatment outcomes of interest. This approach allowed us to prioritize DMRs containing probes with evidence of cross-tissue concordance, thereby increasing confidence in their potential relevance to central nervous system processes underlying TRD and treatment outcomes.

X-linked DMR evaluation and X-Chromosome inactivation status filtering

Due to the unbalanced distribution of males and females in our cohort as well as the limited sample size prohibiting robust sex-stratified analyses, we adopted a filtering strategy to further minimize sex-related confounding. Specifically, we annotated all X-linked DMRs using the X-Chromosome Inactivation (XCI) status classifications from Tukiainen et al. (2017) [68], which provide tissue-informed annotations of genes as subject to or escaping XCI. Genes annotated as subject to XCI in blood were flagged and excluded from downstream interpretation while escape genes were retained and interpreted with caution, as differences in methylation at these loci may reflect sex-dosage imbalance rather than epigenetic differences related to treatment outcomes.

Results

Study participants

Descriptive statistics for demographic and clinical characteristics are shown in Table 1. The average depression measures (PHQ-9 scores) over the course of the study from the baseline measure to the end of treatment among responders and non-responders is shown in Supplemental Fig. 2. All enrolled participants completed at least 36 TMS sessions. Patients continued their prescribed antidepressant medication throughout the treatment. For summary statistics of the prescribed medication classes during enrollment period (see Supplemental Table S1).

Table 1.

Participant characteristics

	Total (N = 46)	Responders (N = 33)	Non-responders (N = 13)	Remitters (N = 18)	Nonremitters (N = 28)	p-value Responders/Non-responders (group comparison)	p-value Remitters/Nonremitters (group comparison)
Age (SD)	44.76 (18.29)	43.73 (18.40)	47.38 (18.49)	44.17 (17.65)	45.14 (19.0)	0.547	0.826a
Female (%)	31 (67.4%)	21 (63.6%)	10 (76.9%)	12 (66.7%)	19 (67.9%)	0.606	1.000b
Male (%)	15 (32.6%)	12 (36.4%)	3 (23.1%)	6 (32.3%)	9 (32.1%)
White (%)	40 (87%)	29 (87.9%)	11 (84.6%)	28 (100%)	22 (78.6%)	1.000b	0.097b
Hispanic or Latino* (%)	26 (56.5%)	20 (60.6%)	6 (46.1%)	10 (55.5%)	16 (57.1%)	0.6760b	0.801b
African American (%)	4 (8.90%)	3 (9.4%)	1 (7.7%)	0 (0%)	4 (14.8%)	1.0b	0.240b
More than one Race (%)	2 (4.4%)	1 (3.1%)	1 (7.7%)	0 (0%)	2 (7.4%)	1.0b	0.658b
PHQ-9 scores
At baseline (SD)	22.28 (2.54)	21.76 (2.69)	23.62 (1.45)	21.56 (2.06)	22.75 (2.73)	0.023a	< 0.001a
4.5 weeks into treatment (SD)	11.89 (6.62)	9.58 (5.84)	17.77 (4.62)	7.00 (4.09)	15.04 (6.03)	< 0.001a	< 0.001a
9 weeks into treatment (SD)	(8.93) (7.16)	5.24 (4.36)	18.31 (3.01)	2.44 (1.29)	13.11 (6.19)	< 0.001a	< 0.001a
∆PHQ-9	13.35 (6.89)	16.52 (5.06)	5.31 (3.47)	19.11 (2.56)	9.64 (6.21)	< 0.001a	0.036a
Antidepressants
Yes	28 (60.9%)	21 (63.6%)	7 (53.8%)	13 (72.2%)	15 (53.6%)	0.782.b	0.339b
No	18 (39.10%	12 (36.36%)	6 (46.15%)	5 (27.8%)	12 (46.43%)
Smokers
Yes	6 (13.00%)	4 (12.1%)	2 (15.4%)	2 (11.1%)	4 (14.3%)	1.0b	1.0b
No	47 (87.00%)	29 (87.9%	11 (84.62%)	16 (88.9%)	24 (85.71%)

Continuous variables are presented as mean (standard deviation) and categorical variables presented as n (%)

PHQ-9 Patient Heath Questionnaire 9, SD Standard deviation, a = t test, b = χ² test; ∆PHQ-9 continuous measure of absolute change in depression measure from baseline to completion of treatment. *White and African American participants include participants that identified as Hispanic or Latino

Differentially methylated region analyses

(i) Treatment response

To assess the relationship between baseline DNAm and binary treatment response, we analyzed the top 5% most variable CpG sites using logistic regression. A total of 67 DMRs were identified after multiple testing correction (adjusted p < 0.05) (see Supplemental Table S2). Of these, 21 DMRs (31.3%) were hypomethylated and 46 DMRs (68.7%) were hypermethylated in treatment responders (Fig. 2A).

Fig. 2.

Hypermethylated vs. Hypomethylated DMRs Across Treatment Outcomes. The bar plots visualize the number of DMRs in TMS Outcomes. Hypermethylated DMRs: Represented by the salmon-colored bars, these indicate genomic regions with significantly higher methylation levels in the indicated treatment outcome. Hypomethylated DMRs: Represented by the light, blue-colored bars, these indicate genomic regions with significantly lower methylation levels in the indicated treatment outcome. Plot (A) Hypermethylated and hypomethylated DMRs in treatment responders. Plot (B) Hypermethylated and hypomethylated DMRs that are significantly associated with symptom trajectory measured by the degree of symptom improvement (ΔPHQ-9). Plot (C) Hypermethylated and hypomethylated DMRs in treatment responders in remitters vs. non-remitters

(ii) Symptom trajectory (ΔPHQ-9)

To characterize DMRs associated with the symptom trajectory and degree of symptom improvement we examined the association between methylation levels and the absolute change in PHQ-9 scores from baseline to the end of treatment (ΔPHQ-9). A total of 23 significant DMRs were identified after multiple testing correction (padj < 0.05), suggesting that baseline methylation at these loci predicted the magnitude of change in depression severity (see Supplemental Table S3). Of the 23 DMRs identified, 9 (39.13%) were hypomethylated, while 14 (60.87%) were hypermethylated (Fig. 2B).

(iii) Remission status

To evaluate the association between baseline DNAm and remission status, we compared DMRs between individuals who achieved remission (PHQ-9 score of < 5) and those who did not. A total of 163 significant DMRs were identified after multiple testing correction (padj < 0.05) (see Supplemental Table S4). Of the 163 DMRs identified, 128 DMRs (78.5%) were hypomethylated, indicating lower methylation levels in individuals who achieved remission compared to non-remitters. In contrast, 35 DMRs (21.5%) were hypermethylated (Fig. 2C).

Converging DMRs among TMS treatment outcomes

There is a total of 16 DMRs overlapping across three models overlap in DMRs identified across three models— treatment response, symptom trajectory, and remission and all 16 of them show the same direction of effect in each model (see Supplemental Fig. 2 & Table 2).

Table 2.

Differentially methylated regions associated with treatment outcomes

DMR	Chromosome	Treatment Response	Symptom Trajectory	Remission
HOXA4	chr7	↑ (FDR = 0.00198)	↑ (FDR = 6.62e-05)	↑ (FDR = 0.00106)
DDX43	chrX	↑ (FDR = 0.000596)	↑ (FDR = 8.3e-05)	↑ (FDR = 7.4e-05)
TNNT1	chr19	↑ (FDR = 0.000596)	↑ (FDR = 8.03e-06)	↑ (FDR = 0.0169)
TMEM72	chr12	↑ (FDR = 0.000596)	↑ (FDR = 1.18e-05)	↑ (FDR = 0.0142)
ATP6V1C1	chr8	↑ (FDR = 0.000596)	↑ (FDR = 1.18e-05)	↑ (FDR = 0.000122)
MAGEE2	chrX	↑ (FDR = 0.0165)	↑ (FDR = 0.026)	↑ (FDR = 0.0395)
S100A1	chr1	↑ (FDR = 0.0255)	↑ (FDR = 0.000238)	↑ (FDR = 0.00106)
S100A13	chr1	↑ (FDR = 0.0411)	↑ (FDR = 0.00106)	↑ (FDR = 0.00243)
LOC144571	chr12	↑ (FDR = 0.0431)	↑ (FDR = 0.00606)	↑ (FDR = 2.95e-05)
GSTM5	chr1	↑ (FDR = 0.0246)	↑ (FDR = 0.0004)	↑ (FDR = 0.000503)
LDHC	chr11	↑ (FDR = 0.0389)	↑ (FDR = 0.0145)	↑ (FDR = 0.0285)
RUNX1	chr21	↑ (FDR = 0.0255)	↑ (FDR = 0.0377)	↑ (FDR = 0.0327)
MIR886	chr5	↑ (FDR = 0.000596)	↑ (FDR = 7.9e-08)	↑ (FDR = 0.00092)
TMEM232	chr5	↑ (FDR = 0.000816)	↑ (FDR = 1.33e-06)	↑ (FDR = 7.54e-05)
HOXA5	chr7	↑ (FDR = 0.00193)	↑ (FDR = 5.61e-08)	↑ (FDR = 3.6e-05)
RUFY1	chr5	↑ (FDR = 0.000654)	↑ (FDR = 0.0111)	↑ (FDR = 0.00877)

DMRs associated with TMS treatment outcomes across three models, with chromosomal location and direction of effect. This table summarizes 16 DMRs that were consistently identified across three outcome models: treatment response, treatment trajectory (ΔPHQ-9), and remission status. For each DMR, the corresponding gene annotation and chromosomal location are listed. Directionality of methylation is indicated by arrows: ↑ denotes hypermethylation, based on the sign of the log₂ fold change (log2err). Statistical significance is represented by the Benjamini–Hochberg false discovery rate (FDR)–adjusted p-value from each model. Only DMRs reaching FDR significance (padj < 0.05) in all three models are included, reflecting regions with robust and convergent methylation signals associated with treatment outcomes

Functional enrichment of genes associated with methylation regions

Gene Ontology (GO) enrichment analysis for Biological Processes (BP) as well as KEGG Pathway Enrichment Analysis was performed to assess the functional relevance of DMRs associated with binary treatment response, but no identified terms reached statistical significance after multiple testing correction (FDR < 0.05; Supplemental Table S10 -S15).

Targeted analysis of methylation differences in TMS outcomes in BDNF CpG sites

We performed a targeted analysis of DNAm at 84 CpG sites located within the promoter region of the BDNF gene. Linear models were fitted using the limma package to compare methylation levels among treatment outcomes, adjusting for relevant covariates. Due to the hypothesis-driven nature of this analysis and the small number of tests, we report results based on nominal p-values (p < 0.05) without correction for multiple comparisons. In our targeted analysis of differential BDNF methylation between treatment responders and non-responders we identified 10 CpG probes that were nominally significant (p < 0.05) (Supplemental Table S5). The most significant probe, cg10558494, showed a hypomethylation in responders compared to non-responders, with a nominal p-value of 0.0012 and a log fold change (logFC) of −0.0156. Similarly, cg25962210 and cg24249411 demonstrated nominal significance, with p-values of 0.0098 and 0.0105, respectively. Both probes also exhibited lower methylation levels in responders relative to non-responders. Other nominally significant probes, including cg06684850 and cg24650785, followed a similar trend, further supporting the observation of hypomethylation in responders. The methylations status of the 84 assessed BNDF probes in relation to treatment response is illustrated in Fig. 3.

Fig. 3.

Volcano Plot of Differentially Methylated Probes in Responders vs. Non-responders. The volcano plot displays the nominally significantly differentially methylated probes identified in BDNF promoter regions using limma analysis. The x-axis represents the log2 fold change (logFC) in methylation levels between responders and non-responders. Positive logFC values indicate higher methylation (hypermethylation) in responders, while negative logFC values indicate lower methylation (hypomethylation) in responders. The y-axis shows the -log10(p-value), highlighting the significance of the differences. Probes with p < 0.05 are considered significant and are color-coded: 🔴 Red points: Hypermethylated probes in responders, 🔵 Blue points: Hypomethylated probes in responders, and ⚪ Grey points: Non-significant probes. Labels indicate the names of the most significant probes. Dashed vertical lines represent the threshold for no change in methylation (logFC = 0), while the dotted horizontal line marks the significance threshold (p = 0.05)

Our symptom trajectory model identified 7 CpG probes that were nominally significant (p < 0.05), indicating an association between methylation at these sites and changes in depression severity over time (Supplemental Table S6). The most significant probe, cg10558494, exhibited a negative log fold change (logFC) of −0.000784 and a nominal p-value of 0.0082, suggesting a trend toward hypomethylation at this site in individuals with greater reductions in depression scores. Similarly, cg25962210 and cg24249411 demonstrated nominal significance, with p-values of 0.0127 and 0.0168, respectively, and negative logFC values, indicating hypomethylation in association with improved treatment outcomes. Interestingly, cg22973087 and cg02947993 showed positive logFC values (0.000499 and 0.000461, respectively). The remaining two nominally significant probes, cg05575117 and cg14036474, exhibited trends consistent with the observed pattern of hypomethylation, highlighting the potential role of epigenetic regulation at these sites in modulating treatment outcomes.

Five probes including cg10558494, cg25962210, cg24249411, cg24650785, cg02947993 overlapped between the treatment outcomes and symptom trajectory model showing concordant directions of effect.

In our analysis of BDNF probes in remitters vs. non-remitter we did not identify any significant probes (Supplemental Table S7).

Concordance of DNAm signatures between blood and brain

To evaluate the potential relevance of peripheral methylation findings to brain tissue, we assessed blood–brain DNAm concordance for CpG probes within differentially methylated regions (DMRs) that overlapped across all three outcome models (treatment response, symptom trajectory, and remission; n = 16 DMRs). For each DMR, we identified CpG probes that appeared in the leading-edge subset of all three models, yielding a set of 129 probes. Shared probes were queried using a publicly available reference database BECon tool [67], which provide probe-level estimates of methylation concordance between peripheral blood and multiple brain regions providing correlations for 90 out of the 129 shared probes. (see full list in Supplemental Table S8) We identified a total of four DMRs with probes showing blood-brain correlation R >0.5 (see Table 3). Our identified differential methylation of BNDF promoter probes did not show any methylation concordance (Supplemental Table S9).

Table 3.

DMRs with CpG probes showing strong blood–brain methylation concordance (r > 0.5)

DMR	Correlated Probes (r > 0.5 & p < 0.05)	Probe IDs with Correlation (r)
DDX43	1	cg12045875 (r = 0.53)
MAGEE2	1	cg05152459 (r = 0.53)
LDHC	3	cg07093428 (r = 0.62); cg11821245 (r = 0.62); cg14323483 (r = 0.53)
MIR886	7	cg00124993 (r = 0.72); cg06536614 (r = 0.69); cg08658272 (r = 0.66); cg18894933 (r = 0.63); cg12028246 (r = 0.61); cg02926764 (r = 0.59); cg20701887 (r = 0.58)

DMRs with CpG probes showing strong blood– Brain- methylation concordance (r > 0.5). This table summarizes DMRs identified across all three outcome models (treatment response, symptom trajectory, and remission) that contain one or more CpG probes exhibiting strong cross-tissue concordance between peripheral blood and brain. Blood–brain correlation values (r) represent mean methylation concordance across multiple cortical regions (Brodmann Area 10, 20 and 7) and were derived using reference datasets from the BECon Tool. Only probes with a correlation coefficient > 0.5 and p < 0.05 are listed. These findings highlight a subset of DMRs with peripheral methylation signatures that may reflect central nervous system epigenetic regulation relevant to TMS Outcomes

X-linked DMR evaluation and XCI status filtering

Among the six X-linked DMRs in our treatment response model MAOA, SAT1, EFNB1 were classified as escaping X-inactivation and therefore excluded from interpretation in our discussion. In our remission status analysis were six X-linked DMRs of which only EFNB1 was excluded. The symptoms trajectory model did not include any X-linked DMRs.

Discussion

The aim of this study was to characterize blood-derived DNAm-based epigenetic differences in TMS treatment outcomes in patients with TRD. We analyzed baseline DNA methylation (DNAm) in relation to three clinical outcomes derived from PHQ-9 depression measures: (1) treatment response, (2) symptom trajectory, and (3) remission status. Further we conducted a hypothesis driven analysis investigating BDNF promoter methylation in relation to TMS treatment outcomes. To our knowledge, this represents the first genome-scale investigation of DMRs associated with distinct TMS outcomes.

Treatment response

In the treatment response model comparing responders and non-responders to TMS treatment, we identified a total 67 significant DMRs that occur in genes belonging to pathways previously implicated in MDD [34] and TRD [69] including DMRs involved with the regulation of monoamines [70] and inflammation [71, 72]. Among the 21 DMRs hypomethylated in responders, our top identified DMR was BCL-6 corepressor (BCOR), a regulator of early embryogenesis previously implicated in a spectrum of pathologies, including oncogenesis [73], neurodevelopmental disorders [74], and cutaneous syndromes [75]. Additionally, hypermethylation of BCOR has also been observed in mothers that were pregnant at the time of the Rwandan genocide as well as in their offspring [76].

Further, we identified significant hypomethylation of oxytocin/neurophysin I prepropeptide (OXT) in TMS responders. Hypomethylation of OXT has previously been associated with MDD where individuals with MDD showed significantly lower DNAm in the OXT gene compared to healthy controls [77]. A recent case control study examining changes in oxytocin levels and oxytocin receptor gene (OXTR) gene methylation during inpatient psychotherapy among women with MDD, identified higher exon 2 OXTR methylation in patients achieving remission, suggesting that DNAm at specific OXTR sites may be linked to treatment outcomes [78]. Oxytocin can exert strong inhibitory effects on the Hypothalamic–Pituitary–Adrenal axis, a pathway known to be dysregulated in TRD [79], and MDD more broadly, by decreasing the secretion of corticotropin-releasing hormone from neurons in the paraventricular nucleus, inhibiting adrenocorticotropic hormone secretion from the anterior pituitary, and potentially decreasing cortisol release directly at the adrenal glands [80]. Further among our top DMRs among treatment responders, we identified significant hypomethylation of Major Facilitator Superfamily Domain Containing 6 Like (MFSD6L). Hypomethylation of this DMR was also previously identified in PTSD cases compared to trauma-exposed controls, suggesting reduced expression of MFSD6L in PTSD [81].

Symptom trajectory (ΔPHQ-9)

In our symptom trajectory model, we identified a total of 23 significant DMR associated with the degree of symptom improvement. Many of these overlapped with the treatment response results, including Homeobox A4 (HOXA4), S100 calcium binding proteins A1/A13 (S100A1/S100A13), troponin T1, slow skeletal type (TNNT1), DEAD-box protein DDX43 (DDX43), V-type proton ATPase subunit C 1 (ATP6V1C1), and Transmembrane protein 72 (TMEM72​). We also observed new significant DMRs in this model. For instance, Runt-related transcription factor 1 (RUNX1) showed higher methylation correlated with better improvement​. RUNX1 encodes a hematopoietic transcription factor that regulates immune and inflammatory gene expression– a pathway that’s increasingly implicated in depression and treatment outcomes [71, 82] and treatment resistance [83, 84].

Remission

The remission model yielded the largest number of significant DMRs (n = 163), converging on previously identified loci observed in the other models as well as novel candidates that may be specific to remission from MDD. Many of the identified DMRs are involved with key domains previously implicated in MDD including neurotransmission, neurodevelopment, immune regulation, metabolism, and glial support. Among DMRs unique to remission, additional DMRs related to immune and inflammatory pathways emerged. For example, NF-kappa-B essential modulator (IKBKG), a central regulator of the NF-κB inflammatory pathway frequently implicated in stress-related inflammation and depression [85] was hypomethylated in remitters, potentially implying enhanced activity in this pathway among remitters. NF-κB signaling has been repeatedly implicated in depression etiology [86] as well as antidepressant response [87] and might show a unique methylation pattern in patients achieving remission following TMS.

Converging DMRs among TMS treatment outcomes

Collectively, we identified 16 DMRs overlapping across three models of antidepressant outcomes—binary response, symptom trajectory, and remission. These DMRs were enriched in genes involved in neurodevelopment such as HOXA4, Homeobox A5 (HOXA5), RUNX1), immune signaling e.g. OTU deubiquitinase 5 (OTUD5), Genotypic glucose-6-phosphate dehydrogenase (G6PD), S100A1), and synaptic function including Ephrin B1 (EFNB1), Ras-related protein Rap-2c (RAP2C) and Myelin-associated Oligodendrocyte Basic Protein (MOBP). Importantly these 16 identified DMRs exhibit a consistent direction of effect across models, warranting further evaluation as baseline correlates of TMS treatment outcomes. Among the overlapping differentially methylated regions (DMRs) identified in our study, probes of four DMRs—DDX43, MAGE Family Member E2 (MAGEE2), Lactate dehydrogenase C (LDHC), and MIR886—demonstrated nominally significant blood–brain DNA methylation correlations (R > 0.5), suggesting potential relevance to central nervous system processes.

Divergent patterns among outcomes

While most overlapping DMRs showed concordant directions of methylation across the binary vs. symptom trajectory (17) and symptom trajectory vs. remission (21) comparisons—and some in the binary vs. remission (19)—a subset of DMRs (15), including BCOR and EFNB1, exhibited divergent patterns between the binary and remission models. These differences in direction of effect may be rooted in the distinct clinical constructs underlying treatment response and remission. Response is commonly defined as a ≥ 50% reduction in depressive symptom severity, reflecting meaningful clinical improvement but often leaving residual symptoms [88]. In contrast, remission represents a more stringent outcome, indicating near-complete symptom resolution that might be reflected biologically. These clinical differences may explain the partial overlap in methylation signatures at baseline, as individuals meeting response criteria may not have achieved full remission. Biologically, this divergence may suggest that distinct epigenetic profiles may underlie symptom reduction versus full recovery. Overall, these findings underscore the need for carefully designed longitudinal studies that could disentangle the epigenetic correlates of symptom improvement versus remission.

Targeted analysis of BDNF

We conducted a hypothesis driven analysis to investigate BNDF promoter methylation due to its previous association with antidepressant and ECT treatment outcomes in previous candidate gene studies [30]. Five of our assessed 84 BDNF promoter CpG probes showed nominally significant associations in both treatment outcomes and symptom trajectory models (p < 0.05 in both) and shared a consistent direction of effect: cg10558494, cg25962210, cg24249411, cg24650785, cg0294799. All but one of those overlapping probes (cg02947993) showed hypomethylation in treatment responders or greater symptom improvement. This overlap in findings may suggest that hypomethylation at these BDNF promoter CpG sites may be predictive of increased TMS treatment success. Interestingly, the opposite direction of effect --hypomethylation of BNDF promoter CpGs–has previously been associated with treatment non-response to antidepressant treatment [89, 90]. However, differences in scope of assessed CpGs complicate a direct comparison of findings. Nevertheless, our study suggests that methylation differences of the BNDF promoter do exist among responders and non-responders to TMS.

Limitations

The findings of our study need to be interpreted in the context of several limitations. Our modest sample size of 46 participants necessitated analytic choices aimed to improve interpretability and robustness such as restricting analyses to the top 5% most variable probes. While this approach enhances signal detection in smaller samples, it also narrows the scope of CpGs evaluated, potentially limiting the generalizability of our findings. Despite identifying significant DMRs in TMS treatment outcomes, our results need to be explored in additional, larger samples. Secondly, our analysis was limited to baseline DNAm profiles, captured prior to the initiation of TMS treatment. While informative for identifying predictive biomarkers of treatment response, this cross-sectional design precludes insight into how methylation patterns may dynamically change in response to neuromodulation. Therefore, longitudinal profiling of DNAm across multiple timepoints—e.g., pre-, mid-, and post-treatment—would allow for the identification of potentially TMS-induced epigenetic changes. Additionally, the absence of a sham control group limits the ability to attribute symptom changes solely to TMS rather than nonspecific effects. However, this limitation does not apply to the DNA methylation analyses, which were based exclusively on baseline samples collected prior to treatment initiation.

Another limitation concerns the use of the PHQ-9 as our primary symptom measure. The PHQ-9 captures depressive symptoms over only the past two-week period and may underrepresent chronic, low-grade symptoms or residual functional impairment, particularly in individuals with partial remission. In such cases, low PHQ-9 scores may not fully reflect ongoing clinical burden, potentially affecting treatment response classification. Because many patients with TRD were referred to our sites specifically for TMS, comprehensive medical records predating study enrollment were not consistently available. As a result, we were unable to rigorously examine previously implicated predictors in TRD outcomes such as duration of the current depressive episode, number of prior episodes, history of ECT or other brain stimulation treatments or adequacy of prior pharmacological trials. Additionally, our TMS treatment protocol is not confined to a single FDA-approved protocol (e.g., 10 Hz left DLPFC), our cohort also included patients treated with alternatives such as low-frequency right DLPFC and sequential bilateral stimulation. Although this diversity reflects real-world clinical practice and preserved our already modest sample size, it constrains strict comparability with trials restricted to one stimulation protocol. Lastly, while DNAm was assessed prior to the initiation of TMS treatment—it is important to acknowledge that most participants were already receiving pharmacotherapy at that time, often involving multiple antidepressants. Therefore, baseline methylation profiles may partly reflect pharmacological influences (although we controlled for medication use in our analyses). Future studies with medication stratified cohorts could help clarify the specific contributions of antidepressants to baseline epigenetic variation among TMS treatment outcomes.

Conclusions

In conclusion, our analysis is the first study to our knowledge that characterizes blood-derived DNAm profiles in relation to TMS treatment outcomes in patients with TRD. Our study identified distinct DNAm-based epigenetic signatures among TMS treatment outcomes consistently involving genes in pathways of inflammation and immune functioning. Future studies leveraging larger sample sizes and employing longitudinal designs will be critical to the development of DNAm biomarkers in the context of TMS. Integrating findings from epigenic studies with those of other predictive modalities—such as neuroimaging, genetic, or clinical markers—may further validate the predictive value of the epigenetic candidates toward actionable biomarkers of TMS response.

Abbreviations

TMS

Transcranial Magnetic Stimulation

MDD

Major Depressive Disorder

TRD

Treatment–Resistant Depression

DNAm

DNA Methylation

DMR

Differentially Methylated Region

PHQ-9

Patient Health Questionnaire–9

mCSEA

Methylated CpG Set Enrichment Analysis

FDR

False Discovery Rate

CNS

Central Nervous System

GO

Gene Ontology

KEGG

Kyoto Encyclopedia of Genes and Genomes

BDNF

Brain–Derived Neurotrophic Factor

EPIC array

Illumina Infinium MethylationEPIC BeadChip

QC

Quality Control

Authors’ contributions

Conceptualization: M.U., G.C, A.L.J.; Funding Acquisition: M.U., G.C., A.L.J., G.D.; Project Administration: J.D, Z.G.; Clinical: K.P., G.C., A.L.J.; Investigation: J.D.; Z.G.; Data Curation: J.D., Z.G.; Methodology: J.D., M.U., K.P., G.C; Formal Analysis: J.D., M. M. H. S.; Visualization: J.D; Writing – Original Draft Preparation: J.D.; Writing – Review & Editing: M.U, K.P; Resources: M.U., K.P., G.C.; Supervision: M. U.; All authors reviewed the manuscript.

Funding

This study was supported by the University of South Florida College of Public Health internal award and USF Microbiome Institute Research award. J. Dahrendorff was supported by the College of Public Health Doctoral Fellowship at the University of South Florida.

Data availability

The raw and quality-controlled DNA methylation data, together with the associated phenotype data is publicly available in the Gene Expression Omnibus (GEO) repository under accession number GSE300009. All analysis code presented in this manuscript can be found on GitHub https://github.com/uddin-research-group-at-usf/TRD-TMS-Analysis/tree/main.

Declarations

Ethics approval and consent to participate

This study was approved by the institutional review board of the University of South Florida. The study was conducted in accordance with the ethical standards of the institutional and/or national research committee and with the 1964 Declaration of Helsinki and its later amendments or comparable ethical standards. Additionally, data privacy was maintained in compliance with the Health Insurance Portability and Accountability Act of 1996 (HIPAA). All study participants provided written, informed consent after a detailed explanation of the study procedures.

Consent for publication

All participants provided written informed consent, which included permission for their de-identified data to be published in scientific journals.

Competing interests

Kenneth Pages received honoraria from Neuronetics Inc. and serves as a consultant on their Medical Advisory Board.