Acute TMS/fMRI response explains offline TMS network effects – An interleaved TMS-fMRI study

BACKGROUND
Transcranial magnetic stimulation (TMS) is an FDA-approved therapeutic option for treatment resistant depression. However, exact mechanisms-of-action are not fully understood and individual responses are variable. Moreover, although previously suggested, the exact network effects underlying TMS' efficacy are poorly understood as of today. Although, it is supposed that DLPFC stimulation indirectly modulates the sgACC, recent evidence is sparse.


METHODS
Here, we used concurrent interleaved TMS-fMRI and state-of-the-science purpose-designed MRI head coils to delineate networks and downstream regions activated by DLPFC-TMS.


RESULTS
We show that regions of increased acute BOLD signal activation during TMS resemble a resting-state brain network previously shown to be modulated by offline TMS. There was a topographical overlap in wide spread cortical and sub-cortical areas within this specific RSN#17 derived from the 1000 functional connectomes project.


CONCLUSION
These data imply a causal relation between DLPFC-TMS and activation of the ACC and a broader network that has been implicated in MDD. In the broader context of our recent work, these data implicates a direct relation between initial changes in BOLD activity mediated by connectivity to the DLPFC target site, and later consolidation of connectivity between these regions. These insights advance our understanding of the mechanistic targets of DLPFC-TMS and may provide novel opportunities to characterize and optimize TMS therapy in other neurological and psychiatric disorders.


Introduction
Transcranial magnetic stimulation (TMS) is one of the most valuable tools in neurocognitive research as well as treatment for a variety of psychiatric disorders, especially major depressive disorder (MDD). TMS over the dorsolateral prefrontal cortex (DLPFC) is a promising FDA-approved therapy for treatment-resistant individuals with MDD (Connolly et al., 2012). In depression treatment the optimal stimulation targeting approach has been shown to evaluate patients' individual ACC connectivity based on an initial resting state fMRI session suggesting network engagement due to TMS (Cash et al., 2020;Cash et al., 2019;Cole et al., 2021;Cole et al., 2020a;Fox et al., 2013;Ge et al., 2020;Weigand et al., 2018;Williams et al., 2018b). However, as of today the exact mechanisms of action that underly success in MDD treatment effects remain widely unclear.
Proposed mechanisms of action are: 1. Direct induction of plasticity in the target area (left DLPFC), 2. Remote regulatory effects in functionally interconnected brain areas (e.g. sgACC), 3. Interaction between these effects and auditory (A1, Goetz et al. (2015)) and somatosensory (S1, Adair et al. (2020)) induced brain plasticity due to TMS clicks and stimulation of cranial nerves.
However, in order to evaluate the acute effects of TMS it has to be combined with another imaging modality. In this respect, concurrent TMS/fMRI has been introduced twenty years ago and is a promising tool to map network effects that are caused by the stimulation itself.
A small number of previous studies has investigated the influence of TMS over left DLPFC on immediate BOLD-response (Bergmann et al., 2021;Dowdle et al., 2018) in healthy participants. In this respect, Hanlon et al. (2013) showed that acute TMS leads to activity increase in left middle frontal gyrus, right superior frontal gyrus, left superior temporal sulcus and left dorsal caudate nucleus. In another study by Vink et al. (2018), increased ACC activity was observed in 4 out of 10 participants during TMS over left DLPFC, while acute stimulation led to increased activity in left DLPFC and VLPFC in the majority of study participants. Dowdle et al., (2018) found an activity increase in the left and right middle frontal gyrus, the bilateral insula, thalamus, superior temporal cortices and anterior cingulate cortex. In a very recent study by Oathes et al. (2021) it was shown that TMS over the frontal cortex resulted in a negative response of the subgenual ACC performing an ROI analysis.
Taken together, there is only sparse evidence for acute TMS effects due to left DLPFC stimulation in healthy organisms with heterogeneous outcomes. These heterogeneous outcomes might be due to different methodological shortcoming associated with the combination of TMS/fMRI, such as limited field of view, low signal-to-noise ratio and the choice of sham control. This sparse evidence however suggests a functional involvement of the targeted left DLPFC and the ACC that is suggested to be influenced by TMS stimulation in MDD treatment. It remains, however an open question, which functional network is stimulated during TMS over left DLPFC.
In an offline approach, we previously scanned , individuals before and after high frequency rTMS and used an unbiased data-driven approach to identify DLPFC-TMS responsive networks according to canonical resting state networks from the 1000 Functional Connectomes Project (Biswal et al., 2010b). We demonstrated that TMS over the DLPFC changed connectivity within only a single RSN, namely RSN#17, which comprises the stimulated DLPFC as well as dorsal cingulate cortex, posterior dorso-medial prefrontal cortex, inferior parietal lobule, inferior frontal cortex and posterior temporal lobes. Nonetheless, while these data imply RSN#17 involvement in the DLPFC-TMS mechanism-of-action , it remains unclear whether this modulation is a direct consequence or an indirect, compensatory effect of TMS.
Thus, we opted to collect data during concurrent TMS/fMRI in order to evaluate if TMS related network effects are a direct consequence of stimulation.
In this study, we attempted to overcome some of the shortcomings of previous concurrent interleaved TMS/fMRI studies over left DLPFC by employing a stateof-the-art purpose-designed ultra-thin multi-channel receive MR coil array  to delineate networks and downstream regions activated by DLPFC-TMS. Our coil array enhances fMRI sensitivity and facilitates investigation of the acute neuronal effects of TMS in online TMS/fMRI experiments (Navarro de Lara et al., 2018). This approach allowed us to acquire whole-brain EPI data during TMS over left DLPFC.
Our primary study aim was to evaluate robust stimulation effects in remote brain areas and at the stimulation site. We were furthermore interested, if these effects are similar to a network, which we identified to show offline connectivity modulations in the sgACC in a previous study . Third, we were interested, if acute stimulation results in sgACC modulations as a proposed mechanism in MDD treatment.
Overall, we aimed at integrating recent considerations of TMS connectivity effects into the framework of acute stimulation modulations.
In order to investigate lasting TMS effects (Study 2), we re-analysed study data published in Tik et al. 2017 comprising 60 healthy right-handed subjects (31 female, age: 25.01 ± 4.6 years) who underwent two sessions of rTMS (sham vs. real) and pre-and post-stimulation resting-state fMRI at a 3T Tim Trio MRscanner (Siemens, Erlangen, Germany).

Ethics
The studies were approved by the local Ethics Committee of the Medical University of Vienna and are in accordance with the Declaration of Helsinki. All subjects gave informed written consent to their participation.

Image Acquisition
Functional images were acquired using an EPI sequence for Study 2 (Offline TMS/fMRI) with TR/TE = 1800/38 ms, matrix = 128 × 128, 23 axial slices parallel to the AC-PC-plane, voxel size 1.5 x 1.5 x 3 mm 3 and the 32-channel manufacturer's default head coil (Siemens, Erlangen, Germany). Resting state fMRI was assessed with eyes open and cross-fixation for 6 mins. The scanner bed was carefully prepared for maximum comfort and minimised motion of the subject. In this respect, the subject's head and neck was put on top of upholstery including memory foam cushions (NoMoCo Pillow, CA, USA).
For Study 1 (online TMS), EPI data were acquired with TR/TE=1000/38 ms, 36 slices, 3 x 3 x 3 mm 3 , MB-factor=4 using two in-house built ultra-thin 7-channel receive coil arrays. An acquisition pause of 320 ms within the TR allowed for artifact free imaging during TMS. Voxel size was optimised in order to allow for whole-brain imaging within a TR of 1000 ms and accounting for an acquisition gap of 320 ms.

Data Preprocessing
Before entering the preprocessing pipeline, each data set was checked for data quality issues by visual inspection.
Further resting-state data processing included regressing cerebrospinal fluid and white matter signals, FFT-based band-pass filtering and motion scrubbing (see Tik et al., 2017 for details).

Neuronavigation
In order to track the M1 region for resting motor threshold (rMT) definition, as well as our target region in the DLPFC, we used the Brainsight software suite (Brainsight, Rogue Research, Canada) in combination with a Polaris Spectra infrared camera (NDI, Waterloo, Canada). For online positioning and tracking we 3D-printed coil and subject trackers ( Figure S2). Online motion was observed throughout the course of the scanning session.

DLPFC-Target
We opted to target the coordinate -42, 28, 21 (MNI) for DLPFC stimulation. This coordinate has been shown to result in significant lasting modulatory network effects in the ACC in a previous study combining offline TMS and fMRI . It is at the center of BA46 as suggested in Fox et al. (2012), however shifted towards the sulcus in order to increase targeting accuracy .

TMS
All stimulation experiments were performed using a MagProX100 stimulator with an MRi-B91 MR-compatible TMS coil (Magventure, Farum, Denmark). For all stimulation sessions, biphasic pulses of approximately 280 μs duration were used. For both studies, the motor threshold (MT) was defined for each participant as the di/dt value displayed on the stimulator that produced an MEP response between 0.05-0.1 mV within a time frame between 15 and 35 ms after pulses in the first dorsal interosseous muscle in five out of ten trials during 50% of maximum voluntary contraction. The MEP was recorded using three EMGelectrodes in a belly-tendon application on the right hand. The coil was positioned in a 45° angle in relation to the surface of the skull. Thresholds were determined outside of the MR environment. For Study 1 (online TMS/fMRI) a 7channel surface RF-coil was mounted to the TMS coil during MT determination to account for the increased distance introduced by the RF-coil. During sham stimulation an empty coil housing was placed between RF and TMS coil.

Online TMS/fMRI (Study 1)
For TMS/fMRI our custom built concurrent TMS/fMRI setup came to use. We employed two 7-channel MR receive coil arrays: one coil array mounted underneath the TMS-coil; the other coil array mounted on the contralateral hemisphere to enable whole-brain coverage. The TMS/fMRI coil array was positioned over the left DLPFC using scanner-adapted metal free neuronavigation tracker (see Figure S2). 20 triplets of 10 Hz rTMS at different intensities (80%, 90%, 100% and 110% of rMT) were applied every 30 s in random order, i.e. 5 triplets per condition. Including an ITI of 30 s the TMS/fMRI run lasted for 10 mins and 38 s. To avoid stimulation-based artefacts there was an EPI acquisition pause of 320 ms during stimulation (Navarro di Lara, 2015; see Figure 1 and Figure S1).

Offline rTMS (Study 2)
Subjects were stimulated over left DLPFC at 90% of their rMT, at a frequency of 10 Hz and with 1200 pulses in total. Including ITI of 20 s the stimulation lasted for 10 mins. For sham TMS, the vertex was stimulated in another session. Order of real and sham sessions was counterbalanced over subjects. Resting-state fMRI runs had a duration of 6 mins and were performed 10 mins pre and 15 mins post stimulation. Subjects were stimulated outside the magnet lying on the scanner bed at the scanner gantry.

Statistical analyses.
All group statistical analyses were performed using SPM12 for TMS/fMRI data and resting-state connectivity.
For single-subject (first level) analysis of online TMS/fMRI (Study 1), linear regression was performed at each voxel using generalized least squares with a global approximate AR(1) autocorrelation model and discrete cosine transform basis drift fitting (128 s cutoff), as implemented in SPM12. The design matrix comprised four regressors (one per stimulation intensity). Additionally, realignment parameters were added into the model as regressors of no interest.
Resulting maps from single-subject GLM analyses were used for group analysis as implemented in SPM12, that is, linear regression was performed at each voxel, using generalized least squares with a global repeated measures correlation model. The second-level flexible factorial design matrix for the TMS/fMRI experiment (Study 1) comprised one factor representing different stimulation amplitudes (80%, 90%, 100%, 110% of the individual rMT) of 10 Hz TMS triplets over the left DLPFC, for first-level contrasts against implicit baseline.
For resting-state fMRI data (Study 2), we restricted our analysis to RSN#17 of the fcon1000 project results (Biswal et al., 2010a), as this particular RSN has previously been shown to be specifically affected by DLPFC TMS .
The fcon1000 activation map of RSN#17 was thresholded at p<0.05 whole-brain FWE corrected and used as seed-mask. The corresponding mean time-courses were correlated to the entire brain data and transformed into Fisher's z maps.
For the offline TMS experiment (Study 2), we entered the functional connectivity maps into a flexible factorial design in SPM12 with factors stimulation condition (verum/sham) and run (pre/post1/post2). The resulting contrasts for RSN#17 were modelled for positive as well as negative factor weights (p<0.05, whole brain FWE-corrected), and for positive weights in the acute TMS condition (p<0.05, cluster level FWE-corrected). The second run after rTMS (post2) is not considered in this manuscript but remains part of the statistical model. The results of the second run are reported in a previous study .

Online TMS acutely leads to increased target network engagement
We first analysed TMS-related brain activation from concurrent TMS/fMRI over all four TMS intensities (80%, 90%, 100% and 110% of rMT) together. Figure 2 (upper panel) shows the resulting activation maps overlaid to a surfacerendered brain. For comparison, the lower panel in Figure 2 shows the restingstate network, i.e. RSN#17 that exhibits prolonged changes of functional connectivity after a single session of rTMS as obtained in our previous study . Dice indices between online and offline effects revealed an average of 0.41 (SD=0.09) for real and 0.15 (SD=0.11) for sham as compared to offline.
For comparison, t-values of RSN#17 (pre stimulation) connectivity are reported next to acute TMS/fMRI induced activation changes (Table 1) (Table 1).  Figure 4, where the functional connectivity map from (Fox et al., 2012a) for the sgACC seed calculated in the Human Connectome dataset (top), which is putatively involved in TMS mechanism-of-action, is compared with RSN#17 before stimulation (right).

Involvement of the DLPFC/ACC axis
To further strengthen this observation we additionally calculated the sgACC seed correlation map (left) using our data (n=60).

ACC target engagement during TMS
In order to investigate the effect of stimulation dose, we examined the acute effect of TMS intensity on brain activation patterns by contrasting maps from supra-(110% rMT stimulation intensity) and sub-threshold (80% rMT) TMS.
The resulting maps are shown in Figure 5 (upper panel). We find statistically significant clusters (p<0.001, uncorrected) in the sgACC (peak: -2, 12, -12 mm [MNI], Tacute = 3.60) and midbrain periaqueductal gray (peak: -2, -32, -10 mm [MNI], Tacute = 3.59), indicating a dose-dependent increase in sgACC engagement as TMS intensity is increased. Activations did however not survive FWE correction. We have moreover calculated dose response plots at the peak voxels of these areas showing a linear increase in midbrain and a less unequivocal pattern at the sgACC (see Figures S3-S5). The same effect could not be observed during sham TMS/fMRI (see Figures S6-S8). All other comparisons between conditions did not reveal statistically significant differences between stimulation intensities.

Offline TMS leads to decreased anti-correlation in the sgACC
We have calculated correlations and anti-correlation patterns in the TMSresponsive network RSN#17 in one sample (n=60). Importantly, comparing anticorrelation patterns within RSN#17 before and after a single rTMS session also shows connectivity changes in the sgACC region ( Figure 5, lower panel). Within the targeted network RSN#17, sgACC anti-correlation evident in the pre-TMS map (yellow circle) changes to non-significant positive correlation after rTMS. A direct comparison of RSN#17 before and after rTMS shows distinct changes in the sgACC (yellow arrow). This highlights the role of RSN#17 as the responsive target network for sgACC modulation.

TMS-responsive network
Here, we showed that stimulation over left DLPFC leads to elevated activity in a network resembling the TMS-responsive RSN#17 , including bilateral DLPFC, ACC, IPL and caudate nucleus. Importantly, the stimulation target in the DLPFC is included in this network. Thus, this is the first study providing evidence for increased BOLD response at the DLPFC stimulation target site (compare Rafiei and Rahnev (2021)).
The RSN#17 is similar to a network introduced by Fox et al. (2012a) to provide an efficient target for depression treatment. We were able to show that the RSN#17 changes connectivity in the sgACC due to TMS over left DLPFC. The DLPFC is part of a meso-cortico-limbic reward circuit that includes the ventral tegmental area (VTA) and nucleus accumbens (NAcc) (Feil et al., 2010). The ACC was shown to be crucially involved in mood disorders (Drevets et al., 1997) and has manifold associative fibres to limbic structures, as the amygdala and hippocampus, as well as the NAcc (Paus, 2001;Wakana et al., 2004). The IPL is a brain area involved in a variety of cognitive integration processes, such as memory, language and social behaviour and turned out to be dysregulated in depression (Müller et al., 2013). The caudate nucleus has recently been shown to be associated with social anhedonia in depressive individuals (Enneking et al., 2019).

Stability of a TMS-responsive Network
In depression treatment, there is consensus among practitioners for the left DLPFC to be the optimal cortical target for effective symptom reduction. While TMS over left DLPFC is FDA-approved for depression treatment, it still remains a matter of debate, which sub-field of the left DLPFC is the best target for depression treatment. Modern approaches suggest to define the optimal target either based on fMRI tasks (Luber et al., 2017) or resting-state connectivity (Fox et al., 2012a). The most efficient target was proposed to be the area of the left DLPFC with the highest anti-correlation to the sgACC (Baeken et al., 2015;Cash et al., 2019;Fox et al., 2013a;Opitz et al., 2017;Williams et al., 2018). Current approaches select a seed that is the most anti-correlated to the sgACC using hierarchical clustering (Cole et al., 2020b;Williams et al., 2018a). However, signal-to-noise-ratio in the sgACC is low and it might thus not be the most reliable target. We propose here a whole TMS-responsive network that reveals sgACC correlation changes by reproducing the network via seed-based correlation analysis of the mean time-course within network masks from the 1000 connectomes' RSN#17. This network was shown to be highly selective for TMS over DLPFC application compared to 19 other most common resting-state network. The RSN#17 might therefore provide a robust target network for the evaluation of depression treatment response and definition of the DLPFC stimulation target.
Beyond that the targeting of other limbic structures has been proposed in depression treatment (Oathes et al., 2021;Xu et al., 2021), taking into account different biotypes of depression (Drysdale et al., 2017). Therefore, targeting of other structures might be taken into account deciding on the optimal TMS treatment.

Concurrent TMS/fMRI and target engagement
In this study, we provided new insight into the specific mechanism-of-action of rTMS delivered to the left DLPFC using a novel setup for concurrent TMS/fMRI.
We showed that acute activation changes during online TMS/fMRI co-localize with the longer-lasting changes in functional connectivity evoked by offline TMS.
Specifically, both acute and longer lasting changes co-localize to the previously suggested RSN#17 . Concurrent TMS/fMRI based on the specific high-sensitivity setup used in this study, is the key to understanding the hitherto unknown immediate neural response to TMS. This consistency of acute and longer lasting effects strongly suggests that changes in resting-state functional connectivity after rTMS are not delayed compensatory effects but direct consequences of the TMS stimulation applied. Therefore, repetitive TMS of one region within RSN#17 leads to repeated co-activation of the whole network resulting in a modulation of network-intrinsic synchrony.
Concurrent TMS/fMRI conceptually allows identification of TMS-induced changes in brain activity and connectivity at the brain circuit level (Bergmann et al., 2016) but due to technical limitations most former concurrent TMS/fMRI studies targeted the motor system. There, the observed distal cortical and subcortical BOLD changes provided early evidence for causal connectivity between these regions and the stimulated motor cortex (Bestmann et al., 2003;Bestmann et al., 2004;Bharath et al., 2015;Watanabe et al., 2014).
In a recent review, (Beynel et al., 2020) (Liston et al., 2014) it was shown that hyper-connectivity between sgACC and components of the default-mode network decreased after five weeks of TMS in MDD patients. In the same way, (Xiao et al., 2019) found a decrease between sgACC/DMN connectivity after TMS over left DLPFC treatment. Also, (Philip et al., 2018) showed that MDD symptom reduction leads to reduced connectivity between sgACC and DMN, DLPFC and insula after three weeks of TMS over left DLPFC. (Jing et al., 2020) found that stimulation responders had stronger sgACC/DLPFC connectivity before treatment. In another study, (Baeken et al., 2015) demonstrated that left DLPFC stimulation treatment response was marked by sgACC decrease of regional glucose metabolism. Moreover, (Dichter et al., 2015) concluded that TMS-response was predicted by sgACC connectivity. While these studies focused on correlates of treatment success in terms of longer-lasting changes in pre-selected networks, we recently employed an unbiased approach that did not require a pre-selection of possible RSNs or seed regions . We found that rTMS over DLPFC causes transient changes to only one specific resting state network, namely RSN#17 and concurrent TMS/fMRI activations overlap with this responsive network.
Some work implicates the ACC as an intermediate relay between the DLPFC and sgACC (Posner and DiGirolamo, 1998), which may itself play an important role in cognitive and affective aspects of depression (Barker et al., 2018;Pizzagalli et al., 2018;Rolls et al., 2019). The putative indirect connection between DLPFC and sgACC via the ACC has become an important proxy for depression treatment and there are hints to some extent for the sgACC being modified by TMS over left DLPFC. The subgenual anterior cingulate cortex (sgACC) has in general been implicated in clinical outcomes across a wide range of antidepressant treatments (Mayberg, 2006). Consequently, this region has been targeted using deep brain stimulation (Drobisz and Damborska, 2019;Merkl et al., 2018), and has demonstrated increasing clinical efficacy with recent refinements in connectivity-based targeting strategies (Holtzheimer et al., 2012;Johansen-Berg et al., 2007;Kennedy et al., 2011;Lozano et al., 2008;Mayberg, 2006;Mayberg et al., 2005). shown that auditory stimulation can lead to changes in cortical plasticity and excitability (Clapp et al., 2005;Mears and Spencer, 2012;Pellegrino et al., 2022;Schuler et al., 2022;Zaehle et al., 2007). Also, cranial nerve stimulation is suggested to induce plasticity in neural circuits (Adair et al., 2020). Since there might be a non-linear relationship between peripheral and transcranial effects it remains to be evaluated, if the "pure" transcranial effect can be disentangled form peripheral effects.

Limitations
Here we aimed at unravelling TMS network effects during left DLPFC stimulation in order to reveal the direct mechanisms of action of stimulation. This could shed new insights on the mechanisms of TMS treatment of MDD. However, we do not know whether the effects observed in a healthy sample translate directly to a clinical population. This will be the subject of a follow-up study.
Furthermore, our between-subjects design lacks statistical power to statistically compare effects of real and sham stimulation. To investigate sham effects in more detail, future studies based on within-subject comparisons are required. As of today, there is no optimal sham condition for TMS/fMRI. While we herein increased the distance between coil and scalp to mimic sensory effects without having transcranial effects, stimulation of peripheral nerves might, however, not have been the same during real and sham stimulation.

Conclusions
With this study, we were able to shed light on target engagement of TMS over left DLPFC, as commonly used in depression treatment. We could show that concurrent TMS/fMRI over left DLPFC activates a network, which has been shown to be modulated after rTMS. Based on our observation we suggest the DLPFC hub of the RSN#17  as a stable target. Figure 1. ONLINE and OFFLINE TMS/fMRI setups. Study 1, left: For online, concurrent TMS/fMRI we used our dedicated coil setup and adapted EPI sequence for imaging of acute TMS effects. EPI (TR=1) recording was conducted continuously intermitted by triplets 10 Hz TMS every 30 seconds (320ms acquisition pause). Intensities were applied in doses of 80%, 90%, 100% and 110% of the subjects' motor threshold (randomized order, 20 TMS triplets). Study 2, right: Using an unbiased analysis approach we included data from 60 subjects pre and post 10 Hz rTMS of the left DLPFC (MNI: -42, 28, 21).   . Immediate online and lasting offline TMS-effects. Immediate response to 80% and 110% of the motor threshold as well as the contrast image (110%>80%) is shown (upper panel). TMS with higher doses leads to increased activity in the sgACC (yellow arrow, p<0.001 uncorrected). The TMS-responsive network shows a decrease in anti-correlation of the sgACC (lower panel, yellow circle, p<0.001 uncorrected) after rTMS. A comparison between pre-and post-TMS connectivity reveals an increase in connectivity of the sgACC to the RSN#17 (yellow arrow). This is in line with findings by (Fox et al., 2012a;Fox et al., 2013).