The concept of modularity (i.e., system segregation) has a long history in the development of our understanding of how the brain processes information. More recently, with advances in neuroimaging techniques, this notion has been supported by the finding that the brain has a modular functional organization consisting of distributed, minimally overlapping large-scale networks (e.g., Bullmore and Sporns, 2009; Power et al., 2011; Thomas Yeo et al., 2011. Note that those large-scale networks are also referred to as “resting-state networks,” “modules,” “communities,” etc. Here we will refer to them as “modules.” Furthermore, although the brain’s anatomical organization has also been found to be modular, here we focus on the brain’s functional organization). Although the specification of the individual functional modules is still an active research topic, there is general agreement that primary sensory-motor processing areas (e.g., sensorimotor cortex and primary visual cortex) form modules that carry out specialized, domain-specific functions (Biswal et al., 1995; Bertolero et al., 2015; Thomas Yeo et al., 2015), and that so-called “association cortex” is composed of multiple modules (e.g., default-mode network, frontoparietal network, etc.) that are involved in high-level functions such as executive control, task switching, and so on (Dosenbach et al., 2007; Cole et al., 2013).

One key advantage of a modular organization is its efficiency in processing complex information by maintaining an optimal balance between flexibility and wiring costs (Bullmore and Sporns, 2012; Sporns, 2013). Specifically, a modular organization consists of segregated modules (i.e., groups of tightly inter-connected nodes) with sparse long-distance connection between modules. Within such a topographical scheme, segregated sub-systems/modules can perform tasks with maximal proficiency and automaticity while minimizing wiring costs. Costly global interactions, which are critical to global coordination and flexibility, are held to a minimum. Empirical evidence supporting these organizational principles has been obtained from functional neuroimaging studies with healthy participants. For instance, higher functional segregation, or modularity, has been linked to greater processing proficiency, e.g., when the task demand is lower (Kitzbichler et al., 2011; Braun et al., 2015) and after a novel task had been learned (Bassett et al., 2011, 2015; Mohr et al., 2016). Additionally, reduced modularity and associated behavioral deficits have been reported in neurological conditions, including stroke (Siegel et al., 2018), Alzheimer’s disease (Brier et al., 2012, 2014) as well as age-related decline (Chan et al., 2014; Geerligs et al., 2015). In sum, a modular neural organization supports efficient information processing, and pathological changes to the organization can have significant behavioral consequences.

Although the concept of modularity is typically applied to the global topology of the brain, it emerges from, and can be evaluated at, the regional level. For instance, reduced global modularity can be caused by decreased within-module connectivity, increased between-module connectivity, or both. Moreover, different diseases may selectively target specific networks (e.g., default-mode network in Alzheimer’s disease; Seeley et al., 2009). Therefore, regional characteristics may be particularly important for understanding network changes in disease and for developing effective NIBS treatments.

The goal of the current study was to understand tDCS-induced changes affecting the functional segregation/integration of the stimulation site – the LIFG – in a cohort of individuals with PPA receiving concurrent language therapy. When combining tDCS and behavioral therapy, one important issue is the pairing between the task and the simulation location. Specifically, it has been argued that, in order for tDCS to maximally affect neuronal processes, the stimulated region should be actively recruited by the behavioral task that participants engage in during the stimulation (Stagg and Nitsche, 2011; Bikson and Rahman, 2013). Therefore, in the current study, we aimed to target the anterior/ventral part of the LIFG [i.e., triangularis, or BA45, referred to as LIFG-triangularis (LIFG-tri) hereinafter], as this region plays a critical role in lexical semantic knowledge retrieval (Bookheimer, 2002; Devlin et al., 2003; Binder et al., 2009), a key cognitive process targeted in the language therapy (written naming) used in this study. Previous studies that also applied anodal tDCS to the LIFG paired with a semantic task showed improved performance along with neural changes in the anterior/ventral LIFG in healthy participants as well as individuals with mild cognitive impairment (MCI) (Holland et al., 2011; Meinzer et al., 2012, 2013, 2015).

In terms of data analysis, the approach was as follows: First, in order to evaluate system segregation, we defined a reference modular organization on the basis of the healthy control (HC) data which corresponded to the division of the whole-brain functional connectome into a set of non-overlapping sub-systems/modules. Second, on the basis of those reference modules, we calculated the level of global connectivity diversity for the LIFG-tri using the network measure participation coefficient (or PC, Guimerà and Amaral, 2005). Specifically, PC quantifies the extent to which a node’s connections are evenly distributed across all the modules such that if all of a node’s connections are within one module, the node’s PC value equals zero, whereas if the connections are evenly distributed across all the modules, the node’s PC value approaches one. In other words, nodes with high PC values are highly integrated with other modules (thus are often called as the “connector hubs”), while nodes with low PC are rather isolated from other modules. Third, as PC only provides a single summary statistic, we also examined the underlying distributions of connections for the LIFG-tri, focusing on within-module and between-module connectivity, i.e., LIFG-tri connections within its own module and connections between the LIFG-tri and other modules. We hypothesized that the within/between-module property is relevant not only because it is closely related to the PC measure, but also because the balance of within and between-module connectivity is a key feature of global modular organization and system segregation. Fourth, in order to evaluate the degree to which the tDCS effects were regionally specific, we also examined the FC profiles of several control ROIs using the same network measures. We examined the right-hemisphere homolog of the LIFG-tri given the strong structural and functional connectivity between homologs. The other reason that the IFG triangularis (RIFG-tri) region was selected as a control region is the long-standing interest in understanding the role of the right-hemisphere homolog subsequent to left-hemisphere damage. Findings in this regard have been mixed with some researchers finding that the right-hemisphere plays a compensatory role to support language functions subsequent to left-hemisphere stroke, while others have found that it plays a maladaptive role (Saur et al., 2006; Turkeltaub, 2015). The second control ROI was the right precuneus. We expected this to be a “neutral” ROI, as this region is relatively spared in PPA and considered to belong to a different functional network than the LIFG-tri (the default-mode network). Thus, the right precuneus would allow us to evaluate if tDCS-induced FC changes occur brain wide or are more spatially specific. Finally, we also examined the other two LIFG subdivisions, the orbitalis and the opercularis, given that they were within the stimulation range.

In sum, in order to further our understanding of the neural changes associated with tDCS effects, this study examined the global FC profile of the stimulation target LIFG-tri before and after a tDCS-augmented naming therapy, with comparisons to a sham-stimulation group as well as to age-matched HCs. Moreover, we examined four “control” regions (RIFG-tri, right precuneus, and the orbitalis and opercularis subdivisions of the LIFG) to investigate whether tDCS-induced FC changes were specific to the targeted LIFG-tri.

Resting-state fMRI data were collected from 32 participants with PPA (16 females, mean age = 67, SD = 6.73) before and after a tDCS-augmented language therapy as part of a clinical trial (NCT02606422). Half of the participants (n = 16) received anodal tDCS over the LIFG and half received sham (n = 16), and both treatments were coupled with behavioral language intervention (Figure 1A).¹ This cohort included the three PPA variants (non-fluent, logopenic, and semantic) and, for each variant, similar numbers of participants received tDCS or sham (10 non-fluent variant: 4 received tDCS, 14 logopenic variant: 8 received tDCS, 8 semantic variant: 4 received tDCS). All participants had a history of progressive language deficits without other etiology (e.g., stroke, tumors, etc.) or primary memory deficits. Differential diagnosis was based on three types of evidence: neuropsychological and language testing, MRI, and clinical assessment, according to criteria in Gorno-Tempini et al. (2011) (see also Neophytou et al., 2019; Themistocleous et al., 2020). Demographic and clinical information for the participants is reported in Table 1. The participants were randomly assigned to the tDCS and the Sham groups, after which it was determined that the two groups did not differ with regard to overall clinical dementia rating for the revised Fronto-Temporal Lobar Degenerations Clinical Dementia Rating (FTLD-CDR, Knopman et al., 2008) and the language sub-component. The overall dementia scores were independently calculated by three raters based on information collected from the participant and family, language and cognitive testing, and questionnaires. The raters then convened to discuss and produce a consensus score. In addition, as in this study we were targeting the LIFG-tri, we also examined whether the tDCS and the Sham group differed in other language functions that have been closely related to the LIFG-tri (i.e., semantic fluency, naming, syntactic processing). As shown in Supplementary Figure 5, we found no difference between the two groups in those assessments.

MRI data were also collected from a group of age-matched HCs (n = 19, 15 females, mean age = 65, SD = 7.81, see Table 1). All control participants were right-handed and native English speakers. The study was approved by the Johns Hopkins Hospital and Johns Hopkins University Institutional Review Board. All participants provided informed consent.

The design and tDCS methods have been reported in our previous publication (Tsapkini et al., 2018) and hence here we summarize the key information relevant to the current study. The behavioral effects of this clinical trial have been described in Tsapkini et al. (2018) in which data from a total 36 PPA participants were analyzed. Eleven participants in this previous study were not included in the current one because the MRI data were not collected due to health issues (pacemakers, claustrophobia, and other medical conditions), and data from seven additional participants who were subsequently evaluated was included in the current participant group. For a comprehensive report of the behavioral results, see Tsapkini et al. (2018).

The tDCS methods have been previously reported in detail (Ficek et al., 2018; Tsapkini et al., 2018) and are summarized here (Figure 1B). Anodal tDCS was delivered by a Soterix 1 × 1 CT device and was applied to the left frontal lobe corresponding to the F7 electrode using the EEG 10–20 electrode position system (Homan et al., 1987). The reference electrode, the cathode, was placed on the right cheek given that extracephalic placement of the reference electrode has been shown to improve current density and delivery (Russell et al., 2017). Current was delivered at 2 mA intensity (estimated current density 0.08 mA/cm²) for 20 min in the tDCS condition. Non-metallic, conductive rubber electrodes covered with saline-soaked 5 cm-by-5 cm sponges were used to minimize the possibility of chemical reactions at the skin/electrode interface. For both tDCS and sham interventions, the electrical current was ramped up at stimulation onset, eliciting a transient (typically 30 s) tingling sensation. In the sham condition, after ramping up, current intensity was decreased to 0 mA. Both active tDCS and sham conditions lasted for 20 min. Behavioral therapy started at the same time as the simulation (for both conditions) and continued for another 20–25 min. Both the therapist and participant were blind to the stimulation condition. Participants were asked to report their general pain level once or twice during each session with the Wong-Baker FACES Pain Rating Scale.²

All MRI data were collected using a Phillips 3T scanner at the F.M. Kirby Research Center for Functional Brain Imaging (Baltimore, MD, United States). The scanning protocol of each PPA participant included one session of resting-state and multiple structural scanning protocols, including a T1-weighted structural image included in this analysis (see Ficek et al., 2018 for further details). For all but 10 PPA participants, the resting-state fMRI (rs-fMRI) scan lasted 8.75 min (210 data-points), and for the remaining 10 participants the scan lasted 6.5 min (156 data-points). The acquisition parameters of the rs-fMRI were as follows: TR = 2500 ms, TE = 30 ms, FOV = 240 mm × 141 mm × 240 mm (ap, fh, rl), flip angle = 75°, voxel dimension = 3 mm × 3 mm × 3 mm, data matrix = 80 × 80 × 47. The T1-weighted structural MRI acquisition parameters were as follows: TR = 8.1 ms, TE = 3.7 ms, FOV = 224 mm × 160 mm × 180 mm (ap, fh, rl), flip angle = 8°, voxel dimension = 1 mm × 1 mm × 1 mm, data matrix = 224 × 224 × 160.

For the HCs, seven underwent the same scanning protocol as described just above for the PPA participants. The other 12 were scanned with a slightly different protocol as they were recruited for a different experiment: Two 7-min runs of rs-fMRI (175 time-points) were carried out consecutively, the acquisition parameters were as follows: TR = 2400 ms, TE = 20 ms, FOV = 206 mm × 123 mm × 220 mm (ap, fh, rl), flip angle = 90°, voxel dimension = 1.7 mm × 1.7 mm × 3 mm, data matrix = 128 × 128 × 41. The T1-weighted structural MRI acquisition parameters were: TR = 6 ms, TE = 2.9 ms, FOV = 256 mm × 256 mm × 176 mm (ap, fh, rl), flip angle = 9°, voxel dimension = 1 mm × 1 mm × 1 mm, data matrix = 256 × 256 × 176. Note that despite some minor differences, both protocols used with the HCs included standard parameters for structural MRI and rs-fMRI acquisition.

As described in Ficek et al. (2018) and Tao et al. (2020), the MRI data preprocessing and FC calculations were carried out with MRICloud,³ a publicly accessible cloud-based platform for automatic neuroimaging data analysis (Mori et al., 2016). MRICloud provides standardized data processing service with pre-tuned parameters, therefore here we only summarize the preprocessing procedure as provided by the developer. For details of the implementation, we refer the readers to the published work by the development team (Faria et al., 2012; Mori et al., 2016). First, the T1 structural image of each individual (in native space) was parcellated into 283 anatomical structures (atlas version “Adult50_90yrs_283Labels_19atlases_M2_V9B”⁴), from which 76 gray matter structures were selected to construct the functional connectome (see below). The parcellation was conducted with a multi-atlas fusion label algorithm (MALF) and large deformation diffeomorphic metric mapping (LDDMM), an algorithm that minimizes mapping error due to atrophy or local shape deformation (Faria et al., 2012).

The rs-fMRI images were preprocessed with MRICloud’s rs-fMRI processing pipeline,⁵ which includes routines imported from SPM5. The preprocessing steps were as follows: slice-timing correction, motion correction that realigned the images to the first volume, physiological nuisance removal with CompCor (Behzadi et al., 2007), outlier volume rejection with “ART”.⁶ The functional images were then co-registered to each individual’s T1 scan with rigid-body transformation, and the averaged time-courses corresponding to 76 gray matter structures (38 in each hemisphere) were extracted, and pairwise correlation values (with Fisher’s z-transformed) were calculated to create a 76 × 76 symmetrical connectivity matrix for each participant (and each time-point). We refer to the 76 gray matter structures as “nodes” in the following network analyses.

In the current study we focused on the network measure, PC (Guimerà and Amaral, 2005), which quantifies a node’s connectivity diversity across multiple modules. Thus, to calculate PC, first we identified an aprior reference modular organization based on the HC group. Specifically, we carried out hierarchical clustering (Ward’s criterion, MATLAB implementation) using the averaged functional correlation matrix of the HC group. Note that the clusters (i.e., modules) identified in this way are often called “resting-state functional networks (RSNs)” or “communities” in the literature (e.g., Smith et al., 2009; Power et al., 2011). Although they can be calculated with different methods (e.g., clustering, ICA) these terms all refer to groups of brain regions with highly correlated time-series. In this manuscript we will refer to these functionally defined clusters of brain regions as modules.

The randomization procedures successfully blinded participants to the assigned stimulation condition since they were at chance (53% correct) at reporting whether they got active tDCS or sham at each period of stimulations in the original crossover study.

We examined the treatment effects in the current cohort following our previous study that reported on the treatment efficacy (Tsapkini et al., 2018). The augmentative effects of tDCS over Sham on the improvement scores for the trained words were evaluated at immediate post-treatment, 2-week, and 2-month follow-ups (Figure 1A), taking into account the co-variates of: pre-treatment accuracy, PPA variant, number of treatment sessions, sex, age, years post onset of symptoms, and clinical dementia rating (FTDL-CDR) and its language sub-score. At the immediate post-treatment time-point, tDCS showed a significant augmentative effect over Sham (p = 0.03), which was maintained at 2-month follow-up (p = 0.0002). The behavioral effects of each group are depicted in Figure 1C and details of the analysis and statistics are reported in Supplementary Material 1. Although a few of the participants reported in Tsapkini et al. (2018) were not included in the current analysis due to the lack of a full imaging dataset, the behavioral results for the participants in the current study were consistent with the previous Tsapkini et al. (2018) report.

As described earlier, the global connectivity measure PC is computed based on a reference modular network identified from the HCs. The reference modular organization consisted of seven modules (Figure 2): (1) temporal, (2) frontoparietal (FP), (3) dorsal frontoparietal (dFP), (4) ventromedial prefrontal (vmPFC), (5) occipital-temporal cortex, (6) subcortical, and (7) perisylvian. All the modules were bilateral and the LIFG-tri ROI, which served as the tDCS target, was situated within the perisylvian module.

The amount of in-scanner motion of each individual was measured as the root-mean-squared (rms) of the six motion parameters and was included in all the regression analyses (see Supplementary Materials 2). There was no significant difference in the amount of motion between the HC and the PPA groups at either time-point [Pre: t(49) = 0.06, p = 0.95; Post: t(49) = 0.64, p = 0.52], nor between the tDCS and the Sham groups at either time-point [Pre: t(30) = 1.22, p = 0.23; Post: t(30) = 0.48, p = 0.64]. Motion also did not differ between the two time-points for either the tDCS [t(15) = 0.71, p = 0.49] nor the Sham group [t(15) = 0.48, p = 0.64], and the differences between pre to post treatment motion parameters did not differ across the two groups [t(30) = 0.13, p = 0.90]. In terms of outlier volume rejection (i.e., scrubbing), only two HCs participants and six PPA participants (three tDCS and three Sham) had fMRI volumes marked as outliers which were rejected before calculating the connectivity values (see Supplementary Figure 3).

The LIFG is one of the most consistently activated regions during language processing and considered to be a “language hub” (Hagoort, 2005; Vigneau et al., 2006). However, in terms of FC, the LIFG has not been typically found to be a “well-connected” region in the functional connectome literature (e.g., He et al., 2009; Bullmore and Sporns, 2012; Power et al., 2013; Ellenblum et al., 2019). This disparity between the LIFG’s task-evoked activation and FC profile may reflect a trade-off between specialization and flexibility, such that while the LIFG may be specialized in language processing, this specialization is achieved at the cost of reduced interaction with other systems.

Consistent with the hypothesis that the perisylvian language module’s high degree of segregation favors efficient language performance, in the current study we found that the LIFG-tri tended to be less segregated and more broadly connected across the whole brain in PPA than in HCs. This higher than normal global connectivity (PC) observed in the PPA individuals was associated with greater dementia severity (Figure 3A), indicating that the increased global connectivity of the LIFG-tri was a pathological change with detrimental behavioral consequences. After tDCS, we observed a reversal of this hyper-connectivity for the tDCS but not for the Sham group (Figure 3B). Furthermore, when we examined the topographical distribution of the connections underpinning the global connectivity decrease of the tDCS group, we found that, before treatment, the within-module connectivity of the LIFG-tri (i.e., within the perisylvian module) did not differ from that of the HCs and remained unchanged after treatment while, in contrast, the between-module connectivity between the LIFG-tri and other modules was higher than normal at pre-treatment and decreased to normal levels following treatment (Figures 4, 5). These decreases were most evident for the LIFG-tri’s connectivity with the frontoparietal (FP) and the temporal modules (Figure 5B). Based on these findings, we hypothesize that tDCS targeting the LIFG-tri increases this region’s functional segregation from other modules, and that this reduction in the LIFG-tri’s pathological hyper-connectivity is associated with its increased language processing efficiency.

Anodal tDCS administered to the LIFG has also been shown to improve language performance and induce neural changes both in healthy participants (Holland et al., 2011; Meinzer et al., 2012, 2013) as well as in MCI (Meinzer et al., 2015). Interestingly, the authors in those studies also argued that the improved language performance was due to tDCS-induced increase in the processing efficiency of the LIFG. For instance, Meinzer et al. (2012) applied concurrent anodal tDCS (or sham stimulation) to the LIFG during a semantic fluency task in healthy young participants and found better performance during tDCS compared to sham stimulation. Most interestingly, they also found that the task-evoked activation in the ventral LIFG (similar to the LIFG-tri examined in the current study), was significantly lower in the tDCS compared to the sham condition [Holland et al. (2011) also showed a similar finding]. Furthermore, when they examined the RSFC, they found it to be higher for tDCS than sham within the left-lateralized perisylvian language regions and lower in other areas like the visual and the sensory-motor area. On the basis of these findings, the authors concluded that the behavioral benefits of anodal tDCS resulted from increased efficiency of the ventral LIFG, a functionally specialized region for lexical semantic knowledge retrieval. In follow-up experiments, the research group evaluated the tDCS effects in older individuals (Meinzer et al., 2013) and individuals with MCI (Meinzer et al., 2015), and reached a similar conclusion that tDCS promoted processing efficiency of the LIFG.

The hypothesis that tDCS increases processing efficiency is also generally consistent with proposed neurophysiological mechanisms of anodal tDCS. It has been proposed that anodal tDCS acts to modify synaptic plasticity and induce long-term potentiation in the stimulation area. These changes have been associated with learning in the brain (Stagg and Nitsche, 2011). As a result, tDCS may enhance learning (or re-learning) with the effect of increasing the efficiency of cognitive processing. This increased processing efficiency could result in the reduced fMRI activation reported in the abovementioned studies that examined on-line fMRI task-evoked activation during tDCS (Holland et al., 2011; Meinzer et al., 2012, 2013, 2015). In this study which examined off-line rs-fMRI activities (before and after a multi-session tDCS intervention), increased processing efficiency may have reduced the LIFG-tri’s interaction with other modules, as reflected by the observed reduction of between-module FC. Note that, however, this proposed mechanism does not preclude the possibility that tDCS may also change, directly or indirectly, the neural activities of other distant regions may also change simultaneously.

To understand the underlying mechanism/s of tDCS, one important question is the extent to which observed tDCS effects are specific to a stimulated area. To evaluate this question, we first examined the right hemisphere homolog of the LIFG-tri, i.e., RIFG-tri, assuming that if the stimulation effects were not restricted to the stimulation site, one prime candidate would be its homolog given the strong functional and structural connection between homolog regions. However, as shown in Figure 6A, although the RIFG-tri, like the LIFG-tri, also exhibited elevated global connectivity levels prior to treatment, it did not show significant pre- to post-treatment changes. Moreover, the effects seemed to be specifically associated with the LIFG triangularis subdivision and, to a lesser extent, with the other anterior LIFG subdivision – the IFG-orbitalis. We did not, however, observe consistently reliable effects for the more dorsal and posterior subdivision LIFG-opercularis, despite its proximity to the stimulation site (Figure 7B). Because the LIFG-orb showed similar effects as the LIFG-tri, we speculate that the two ventral anterior subdivisions are likely to be similar in terms of their connectivity profiles and cognitive functions, and are distinct from the more dorsal posterior LIFG-opercularis. A similar anterior/ventral – posterior/dorsal distinction has also been observed in the other studies that applied tDCS during a semantic task (Holland et al., 2011; Meinzer et al., 2012) discussed below.

There are several possible reasons for these apparent spatially selective effects, and we note that these are not necessarily mutually exclusive. One possibility is simply that tDCS current flow is sufficiently focal such that the major impact is reasonably spatially limited. In this study, it would have to have been limited to the LIFG-tri subdivision. However, this seems somewhat unlikely given the wide and unspecific current spread of tDCS (Brunoni et al., 2012, see Figure 1B for the current modeling results). Another possibility concerns the pairing of the behavioral treatment and stimulation site. It has been hypothesized that the effects of tDCS are maximal when the cognitive function of the stimulated area and the treatment task are well aligned, a mechanism that has been referred to as “functional specificity” (Bikson and Rahman, 2013). The proposal is that tDCS is most likely to affect the neuronal activities of the stimulated area if that area is actively engaged by the behavioral task during the stimulation. Therefore, given that the LIFG-tri plays a key role in selection during lexical-semantic retrieval (Vigneau et al., 2006), the target cognitive function of our behavioral treatment, the functional specificity hypothesis seems most consistent with the observed regional specificity effect (see further discussion in the next paragraph). Finally, another related hypothesis is that the FC changes are linked to neurotransmitter changes caused by anodal tDCS (e.g., reduction in the inhibitory neurotransmitter GABA), which have been shown to be regionally specific (Clark et al., 2011; Hunter et al., 2015; Harris et al., 2019). For example, a previous study by our group found that the tDCS group showed a significant decrease in GABA in the LIFG, something which was not seen either in the Sham group or in the control ROI right sensory-motor cortex (Harris et al., 2019).

Notably, similar regional specificity has been reported in other anodal tDCS studies that also stimulated the LIFG. Both of these prior studies also highlighted the hypothesis that the regional specificity effects resulted from an optimal pairing of task and stimulation site (Holland et al., 2011; Meinzer et al., 2012). In both studies, healthy participants received anodal tDCS (or sham stimulation) to the LIFG while performing a semantic task (picture naming or semantic fluency task). Subsequent tDCS-induced fMRI activation reduction was specifically found in the anterior/ventral part of the left IFG, similar to the LIFG-tri in the current study, and no changes were seen in the right IFG and regions in the vicinity of stimulation (e.g., the dorsal IFG, the precentral gyrus). Both studies attributed the regional specificity effect to the role that the anterior/ventral LIFG plays in semantic processing. Thus, this regional specificity effect is consistent with the “functional specificity” hypothesis that anodal tDCS increases the activity of the task-relevant neurons. Nevertheless, given the scarcity of research on this topic, further study is needed to further evaluate this hypothesis.

We should note that the regional specificity we observed does not necessarily contradict the notion often found in the neurostimulation literature that tDCS (and other neurostimulation methods) can induce network-wide changes distal to the stimulation site (e.g., see Pini et al., 2018 for a review). In fact, the decrease in the LIFG-tri’s global connectivity that we reported here did involve changes in connectivity between the LIFG-tri and various regions/modules across the brain (Figures 4, 5). However, given that, of all the ROIs examined, only the LIFG-tri (and to a lesser extent the LIFG-orbitalis) showed decreased overall global connectivity (PC), the results indicate that the widespread connectivity changes induced by tDCS are likely to stem from the ventral/anterior part of the LIFG, in particular the triangularis. In addition, secondary connectivity changes in other areas or networks are also possible. For instance, the decoupling between the LIFG-tri and the FP module that we observed might introduce additional changes within the FP module (that were not evaluated in this study).

First, in the current study we focused on the stimulation site LIFG, hence our conclusion only apply to the LIFG (triangularis) and not to other regions or to the whole-brain network organization, which may respond differently to stimulation. Indeed, in a recent study (Tao et al., 2020) that examined the global, whole-brain network properties in PPA, we found that the average global integration across the brain was reduced relative to HCs (as measured by global network measures, such as global efficiency and clustering coefficient). Therefore, it is possible that such whole-brain characteristics, as well as other brain regions that play very different roles than the LIFG-tri, may show different tDCS-induced changes. Second, in this study, we focused on the PC measure which specifically evaluates a node’s system segregation level as the connectivity diversity across multiple modules (Eq. 1, also see Figure 4 for the modules used in this study). As a result, however, we can only draw conclusions regarding the aspect of system segregation that is quantified by PC. There are a number of other centrality measures that may capture different aspects of a node’s connectivity profile (e.g., degrees, betweenness, see review by Rubinov and Sporns, 2010) and which may provide additional understanding of the effects of tDCS on FC. A third limitation is the possibility that PPA variants may show different responses to LIFG tDCS. We did not examine this in the current study due to the relatively small sample sizes for each variant. One potential factor that could cause different effects across the PPA variants is that there may be differences in the tDCS current flow due to different distributions of atrophy across the variants. However, this issue was recently investigated by our group and we did not find consistent differences in current flow across the three variants (Unal et al., 2020). Nonetheless even if we assume similar current flow, the underlying disease characteristics across the variants might still play a role in determining neuro-stimulation outcomes. For instance, tDCS might have a larger impact when applied to an already compromised region compared to a healthy one. Thus, it is possible that participants of the non-fluent variant with substantial damage in the LIFG and regions functionally or anatomically connected to the LIFG would show the greatest tDCS benefit. Indeed, this possibility is consistent with results reported by Tsapkini et al. (2018) that the non-fluent variant showed the most robust treatment gains across both arms of the clinical trial (from which the current data set was drawn) and generalized well to untrained words, whereas the logopenic and semantic variants, whose atrophy is relatively distant from LIFG, showed less robust treatment gains. Furthermore, in another study we found that greater tDCS-induced treatment gains were associated with smaller gray matter volume of several left frontal-parietal areas that were functionally and/or structurally connected with LIFG (de Aguiar et al., 2020). Alternatively, it could be the case that tDCS exerts a greater effect when the underlying neural tissue is less (rather than more) damaged, which would also result in different responses to stimulation across the variants. In sum, the interaction between tDCS-induced neural changes and different subtypes of PPA is an important issue that should be examined in the future with larger sample sizes.