Precision Inhibitory Stimulation of Individual-Specific Cortical Hubs Disrupts Information Processing in Humans

Abstract

Introduction

There is widespread interest in using noninvasive brain stimulation (NIBS) as a treatment for psychiatric conditions (Guo et al. 2017), including addiction (Diana et al. 2017), obsessive compulsive disorder (Bais et al. 2014), and depression (Avissar et al. 2017; Dubin et al. 2017). The outcomes of these interventions, however, are variable across treated individuals. In the most widely adopted, Federal Drug Administration approved application of NIBS—repetitive transcranial magnetic stimulation (TMS) for the treatment of medication refractory major depression—only 29% of patients respond positively (Gaynes et al. 2014). Variation in patient response has been attributed in part to uncertainty regarding free parameters inherent to NIBS (Guerra et al. 2017), including the stimulation site. Because the effects of NIBS are believed to propagate from the stimulation site in a manner constrained by the connectivity of the targeted brain area (Muldoon et al. 2016; Gollo et al. 2017), appropriate stimulation site selection is likely critical for therapeutic success, as it will determine whether or not stimulation effects spread throughout clinically relevant neural circuitry.

Stimulation site selection strategies often focus upon anatomical landmarks (Pascual-Leone et al. 1996; Fitzgerald et al. 2009) and local tissue properties (Sack et al. 2009), including activation. The same anatomical area, however, can exhibit different patterns of functional connectivity across individuals (Mueller et al. 2013; Gordon, Laumann, Adeyemo, Gilmore, et al. 2017; Gordon, Laumann, Adeyemo, and Petersen 2017). In other words, stimulating the same anatomical area in different individuals could in practice affect different downstream brain areas—which theoretically could contribute to the variable outcomes of NIBS interventions. The advent of techniques for precisely characterizing the areal (Cohen et al. 2008; Gordon et al. 2016) and network organization (Gordon, Laumann, Adeyemo, Gilmore, et al. 2017; Gordon, Laumann, Gilmore, et al. 2017; Gordon, Laumann, Adeyemo, and Petersen 2017) of individual human brains has set the stage for the development of personalized NIBS protocols, which in principle can increase the likelihood of producing better outcomes in patients (Cocchi and Zalesky 2018).

A network science framework, in which brain areas (“nodes”) engage in networked communication within and across brain networks (“modules”), is theoretically well suited for mapping stimulation sites on an individual basis. This approach can capitalize on the idea that a node’s role in a network can be inferred from its connections (Sporns 2011). Of particular interest are select nodes, termed “connector hubs” (hereafter referred to as hubs), connected to multiple network modules and critical for the function of many networks found in nature (Jeong et al. 2001). Evidence from biophysical models (Honey and Sporns 2008; Misic et al. 2015) and lesions in stroke patients (Gratton et al. 2012; Warren et al. 2014) indicate that hub brain areas could serve critical roles in the human brain as well.

Resting-state fMRI (rsfMRI) functional connectivity within and between functional brain networks is linked with a diverse array of cognitive abilities (Smith et al. 2015), suggesting that these connectivity patterns reflect intrinsic, cognitively relevant information-processing routes in the brain (Raichle 2010; Ito et al. 2017). Because of their unique positioning at the intersection of multiple segregated brain networks, we theorized that administering NIBS to hubs could impact flow of information between brain networks, which in turn would affect performance on a cognitive task more than nonhub stimulation. If this prediction is borne out, it would provide the first causal evidence for the importance of hub brain areas mapped prospectively in individual humans and raise the possibility of leveraging hubs and other features of large-scale brain networks as targets in future NIBS interventions.

We first assessed the feasibility of mapping hubs in single-subjects as NIBS targets using the Midnight Scan Club (MSC) (Gordon, Laumann, Gilmore, et al. 2017), a publicly available dataset of individuals that underwent 10 rsfMRI sessions over a period of 10 consecutive days. We discovered that the spatial positioning of hubs is variable across individuals but highly reproducible within an individual across imaging sessions when mapped using large quantities of per-subject rsfMRI data.

In a prospective, double-blind, within-subject NIBS experiment, we then mapped an individual-specific hub brain area in 24 healthy individuals. We predicted that administering an inhibitory form of TMS to hubs would disrupt information processing during an N-back task, as measured using a drift diffusion model, when compared with nonhub inhibition. Although hub inhibition theoretically should disrupt information processing during any cognitive task, the N-back task was selected given empirical evidence that it requires communication between segregated brain networks (Cohen and D’Esposito 2016; Liang et al. 2016), which is thought to be facilitated by hubs (Liu et al. 2017). We constrained our search space to a single gyrus in order to minimize differences in the gross anatomical location of hub and nonhub stimulation sites. Right middle frontal gyrus was selected for this purpose, a relatively conservative choice in light of its functional association with working-memory (Rottschy et al. 2012; Daniel et al. 2016; Hoy et al. 2016).

Materials and Methods

Midnight Scan Club

Participants

The MSC dataset (Gordon, Laumann, Gilmore, et al. 2017) was downloaded from OpenfMRI.org (ds000224). This dataset is comprised of 10 participants aged 24–34 years (mean age = 29.1 years ± 3.3, 5 F/5 M) that underwent a total of 5 h of rsfMRI over a period of 10 consecutive days (10 × 30-min sessions). Further details regarding data acquisition and sample demographics are reported by (Gordon, Laumann, Gilmore, et al. 2017).

Reproducibility and Interindividual Variation of Single-subject Hub Estimates

To quantify the reproducibility of single-subject estimates, nonoverlapping epochs were randomly sampled from each MSC participant’s motion-censored 5-h rsfMRI time series. A participation coefficient for parcels with a centroid in right middle frontal gyrus was calculated using each epoch, separately. Reproducibility was quantified as the rank spatial correlation of the 2 sets of participation coefficients. This procedure was repeated 10³ times per epoch, with epoch lengths ranging from 1 to 60 min, in 1-min steps. We note that with high-motion participants, such as MSC08 and MSC09, the iterative procedure was only performed up until 30 and 45 min, respectively. This was due to an insufficient amount of time series remaining post motion censoring. Interindividual variation in the spatial distribution of participation coefficients among MSC subjects required comparison of participants with unique areal parcellations. Because node to node comparisons were not possible, a best match for each parcel in participant i was defined in participant j as the parcel with a centroid closest in geodesic space. As an example, consider a scenario where subjects i and j have M and N parcels, respectively. When matching parcels from subject i to those in subject j, this procedure yields a pair of 1 × N participation coefficient vectors. Similarly, when matching parcels from subject j to those in subject i, this procedure will yield a pair of 1 × M participation coefficient vectors. The distance between hubs in single-subjects and their collective group-average was calculated as the distance in euclidean space between the Talairach coordinates corresponding to parcel centroids.

NIBS experiment

Participants

Twenty-four participants aged 18–28 years (mean age = 20.5 years ± 2.5, 11 F/13 M) were recruited from the Georgetown community after complying with the consenting guidelines of the Georgetown University IRB. Participants were screened for history of neurologic and psychiatric conditions, epilepsy, contraindications for MRI, and use of medications that increase likelihood of side effects following TMS.

Data acquisition

All data was acquired on a Siemens Trio 3 T with the participant’s head immobilized using head cushions. A high-resolution structural T1 scan was acquired with the following parameters: MPRAGE: TR/TE = 1900/2.52 ms, 90° flip angle, 176 sagittal slices with a 1.0 mm thickness. Functional echo-planar images were acquired with the following parameters during each imaging sessions: 3 mm isotropic resolution, TR = 2000 ms, TE = 30 ms, flip angle = 90°, FOV = 192 × 192 mm. A T1 and 3 resting-state runs, each lasting 15 min and acquired successively, were collected during the initial imaging visit. Subjects self-reported the percentage of time they felt sleepy during the resting-state scans on a scale of 1–5 (1 = “not sleepy at all,” 2 = “sleepy less than 25% of the time,” 3 = “sleepy 25–50% of the time,” 4 = “sleepy more than 75% of the time,” 5 = “sleepy all the time”). On average, our subjects reported feeling sleepy less than 25% of the time (M = 2.14 ± 0.66), indicating that subject drowsiness was infrequent. An N-back task consisting of 20 blocks (5 blocks each of 1-, 2-, 3-, and 4-back loads, in pseudorandomized order) was administered during each study visit. Blocks consisted of 9 letters, each presented for a duration of 500 ms and with an intertrial interval of 1500 ms. Participants were instructed to provide a right-hand button press for targets and a left-hand button press for nontargets as quickly and accurately as possible. Of the 180 trials, 32 were targets, with either one or 2 targets present in any given block. Stimuli were presented on a backprojection screen using the E-Prime software.

Data preprocessing

Functional images were corrected for differences in motion and slice timing acquisition, and coregistered into each participant’s anatomical image using SPM12 (Wellcome Department of Cognitive Neurology, London, United Kingdom). Functional data were denoised using the aCompCor strategy in the CONN toolbox (https://web.conn-toolbox.org/). Denoising steps included linear detrending and nuisance regression (5 principal components from white matter and cerebrospinal fluid masks from an MPRAGE segmentation; 6 motion parameters and first-order temporal derivatives; and point-regressors to censor time points with mean frame-wise displacement > 0.2 mm). Residual time series were band-pass filtered (0.01 Hz < f < 0.1 Hz). Temporal masks were created to flag motion-contaminated frames for scrubbing. Contaminated volumes were identified by frame-by-frame displacement (FD) calculated as the sum of absolute values of the differentials of the 3 translational and 3 rotational motion parameters. On average, 76 ± 3% (34 min) of rsfMRI time series was retained after motion censoring.

Surface file generation

Following volumetric coregistration, white and pial anatomical surfaces were generated from each participant’s native-space MPRAGE using FreeSurfer’s recon-all pipeline (version 5.0). The fsaverage-registered left and right hemisphere surfaces were then brought into register with each other in fs_LR space (Van Essen et al. 2012) and resampled to a resolution of 32k vertices using Caret tools. Denoised fMRI time series within the cortical ribbon were mapped onto each individual’s midthickness surface and spatially smoothed (σ = 2.55). Both left and right surfaces were combined into the Connectivity Informatics Technology Initiative (CIFTI) format using Connectome Workbench (Glasser et al. 2013), yielding time courses representative of the entire cortical surface, excluding nongray matter tissue, and subcortical structures.

Automated pipeline for identifying stimulation targets

Because the participation coefficient of a node with few edges is difficult to interpret, we eliminated from consideration parcels with a degree in the bottom quartile of whole-brain values. We then identified from the remaining parcels those with a centroid falling within right middle frontal gyrus, defined using gyral labels from the Desikan–Killiany FreeSurfer atlas. The hub was defined as the parcel with the highest participation coefficient value. The euclidean distance between the hub and all other parcels in right middle frontal gyrus was calculated. The nonhub was defined as the parcel with the lowest participation coefficient value >20 mm from the centroid of the hub parcel in euclidean space. This minimum distance was enforced to help facilitative selective targeting of the 2 parcels, given evidence that the spatial resolution of TMS is approximately 0.5–2 cm (Wassermann et al. 1992; Thielscher and Kammer 2002). Native-space MPRAGE coordinates for the hub and nonhub were pseudorandomly assigned to the 2 follow-up sessions.

cTBS and MRI-guided neuronavigation

cTBS was applied following (Huang et al. 2005) using a MagPro x100 device (MagVenture, Inc.) with a passively coiled MCF-B70 figure 8 coil. The orientation of the coil handle was perpendicular to the right middle frontal gyrus in each subject. Trains of three 50 Hz pulses were repeated at 200-ms intervals (Huang et al. 2005) for a total of 600 pulses over a period of 40 s. Stimulation intensity was 80% of active motor threshold. Active motor threshold was defined as the intensity required to induce an evoked potential (≥100 μV in peak-to-peak amplitude) in the contralateral FDI muscle when pulses were applied to right motor cortex during a mild sustained contraction in 5 out of 10 consecutive stimuli. Muscle contractions were measured using surface electrodes placed on the FDI muscle, which are connected to an electromyography device incorporated into the BrainSight system. Parcel centroids were targeted using the Brainsight 2 Frameless stereotactic system for image guided TMS research (Rogue Research). This system uses infrared reflectors attached to a headband worn by the subject to coregister the MPRAGE with the participant’s head. The coil was coregistered via infrared reflectors. The aftereffects of cTBS are thought to last up to 50 min (Wischnewski and Schutter 2015), which is long enough to complete the 12-min N-back task. Self-reports indicated that cTBS was well tolerated (average discomfort rating = 1.33 ± 0.63, with the possible range being 1–10, and higher values indicating more discomfort).

Electric field modeling

We calculated realistic estimates of the electric fields induced by cTBS in our intervention based on the finite elements method using the SimNIBS (v2.1) software package [www.simnibs.org; (Thielscher et al. 2015)]. Tetrahedral volume meshes were constructed for each subject using the “headreco” function, which segments a T1 image into 5 tissues types (white matter, gray matter, cerebrospinal fluid, skin, and bone) using SPM12 and CAT12. We then estimated the normal component of electric field (nE) in hub and nonhub parcels during each cTBS session. Tissue conductivities were set to previously established values [white matter = 0.126 S/m, gray matter = 0.28 S/m, cerebrospinal fluid = 1.79 S/m (Thielscher et al. 2011), skin = 0.25 S/m (Truong et al. 2013), and bone = 0.01 S/m (Dannhauer et al. 2011)]. We evaluated whether the magnitude of the normal electric field differed by target site and by hub status using repeated-measures 2 × 2 ANOVA [target site (targeted vs. not targeted) × hub status (hub vs. nonhub)] was performed using nE [V/m] values.

EZ drift diffusion modeling of task performance

Clustering hubs into discrete subtypes

We calculated the average functional connectivity between hub parcels and all parcels of each brain network that were >30 mm in geodesic space. The similarity of these resultant hub cross-network connectivity profiles to 3 hub-type templates (Gordon et al. 2018) was calculated (details regarding the creation of the hub-type templates is described briefly in Supplementary Fig. S5). This resulted in a 24 × 3 (participant × hub type) array of Fisher-transformed correlation coefficients. A stepwise linear regression analysis was performed using these coefficients as predictor variables and the change in drift rate (drift rate_Hub – drift rate_Nonhub) averaged across loads as a dependent variable. Stepwise linear regression was used because the average hub-type cross-network functional connectivity profiles are not independent from one another.

Data availability

E-prime outputs for each subject and code to perform EZ diffusion modeling of N-back performance are available at https://github.com/cl968/NIBS_Hubs.

Results

Hub Estimates are Reproducible with Sufficient Per-subject rsfMRI Data

We estimated the degree to which discrete cortical areas (“parcels”) in the right middle frontal gyrus of each MSC participant function as hubs using the graph theory metric, participation coefficient (Guimera and Nunes Amaral 2005). Nodes with higher participation coefficient values (“hubs”) have edges that are distributed across more network modules than those with lower values (“nonhubs”) (Fig. 1A,B: participation coefficients in an example MSC subject calculated using all rsfMRI data). We first sought to determine the amount of rsfMRI data necessary for achieving reproducible hub estimates in the right middle frontal gyrus of MSC subjects over a period of several days.

Figure 1.

Spatial distribution of participation coefficients in the right middle frontal gyrus (A) of an example MSC subject (MSC04) calculated using her entire 5-h rsfMRI dataset (B). Participation coefficients calculated using randomly selected 5-min (C) and 45-min (D) epochs from this MSC subject’s 5-h rsfMRI dataset. In matrices C and D, rows represent parcels and columns represent unique epochs (100 epochs selected for visualization). An iterative reproducibility analysis reveals that, across all MSC subjects, the reproducibility of participation coefficients increases with greater quantities of rsfMRI data utilized. Unique colored lines represent different MSC subjects (E). Similarity matrix summarizes interindividual variation in the spatial distribution of participation coefficients, calculated using each subject’s entire rsfMRI dataset (F).

Open in new tabDownload slide

Nonoverlapping epochs were randomly sampled from each MSC participant’s 5-h rsfMRI time series, which was created by concatenating ten 30-min sessions acquired over a period of 10 consecutive days (10 × 30 min = 5 h total). A set of participation coefficients for parcels with a centroid in right middle frontal gyrus was calculated separately using each epoch, and reproducibility quantified as the rank spatial correlation of the 2 sets. This procedure was repeated 10³ times per epoch, with epoch lengths ranging from 1 to 60 min, in 1-min steps. We found that participation coefficients were variable across iterations of this analysis when calculated using smaller, commonly utilized quantities of per-subject rsfMRI data (Fig. 1C) relative to those calculated using larger quantities of data (Fig. 1D). The reproducibility of participation coefficients improved with larger quantities of per-subject rsfMRI data (Fig. 1E). For example, we observed an average r_s of 0.82 ± 0.10 when using 45 min of rsfMRI data. A second reproducibility analysis focused upon reproducibility between imaging sessions and is included in the supplementary information (see Supplementary Mateial 1).

Hubs are Idiosyncratic Features of Functional Brain Organization

We next quantified interindividual variation in the spatial distribution of hub estimates using pairwise rank spatial correlations (Fig. 1F). The spatial distribution of hub estimates across individuals was not similar (average r_s = −0.01 ± 0.26). Furthermore, single-subject hub estimates were not similar to their collective group-average (r_s = 0.09 ± 0.21). The practical implications of this finding for the proposed NIBS experiment was assessed by estimating the effective stimulation zone surrounding the group-average hub (the highest participation coefficient parcel). Administering cTBS to this target would have theoretically failed to inhibit a hub in 70% of MSC subjects, assuming a liberal spatial resolution of 0.5–2 cm. Thus, administering cTBS to a group-average hub could fail to impact hubs in some individuals. Collectively, these findings motivated the decision to map hub and nonhub stimulation sites on an individual basis using large quantities of rsfMRI data in our NIBS experiment.

Precision Mapping and Inhibitory Stimulation of Hubs

Twenty-four healthy participants completed three study sessions (see Fig. 2A for experimental design summary). An automated pipeline mapped a hub (the highest participation coefficient parcel) and nonhub (the lowest participation coefficient parcel ≥ 20 mm from the hub) stimulation site in the right middle frontal gyrus of each subject. The numerical difference in participation coefficient between the stimulation sites was large (Fig. 2B: PC_Nonhub = 0.21 ± 0.13, PC_Hub = 0.72 ± 0.05, and ΔPC = 0.51 ± 0.12), indicating that the pipeline successfully mapped a putative hub and nonhub target in each subject. Notably, we determined retrospectively that hub and nonhub stimulation sites did not significantly differ from one another in terms of their anatomical positioning, nodal degree, baseline N-back activation, or distance to the stimulating coil (see Supplemenatry Material 2). The native-image space coordinates corresponding to the hub and nonhub were then pseudorandomly assigned as targets for 2 follow-up sessions (average interval between sessions = 5.8 ± 5.3 days). Immediately prior to performing an N-back task during these sessions, cTBS (Huang et al. 2005) was administered offline and guided by neuronavigation. A repeated-measures 2 × 2 ANOVA revealed a main effect of targeting [F(1,23) = 91.05, P < 0.001, ω² = 0.47] (Fig. 2C), such that hub and nonhub parcels both exhibited greater nE (V/m) when targeted relative to when they were not targeted, indicating selective target engagement. There was no significant main effect of target [F(1,23) = 0.29, P = 0.59, ω² = 0.00] or interaction [F(1,23) = 0.23, P = 0.62, ω² = 0.00], indicating that nE (V/m) in hub and nonhub parcels did not significantly differ when either was targeted. Finally, findings from the MSC dataset regarding intra and interindividual variability were replicated using this new dataset (see Supplementary Material 3).

Figure 2.

Summary of the NIBS experimental design (A). Participation coefficients associated with hub and nonhub targets mapped by an automated pipeline (B). Strength of the normal electric field (nE) in hub (red) and nonhub (gray) parcels when targeted and when not targeted (C, left). nE component of the electric field vector for the nonhub (first column) and hub (second column) stimulation sites in an example subject under targeted (first row) and not targeted (second row) conditions (C, right). Error bars denote standard error. Hub and nonhub target centroids are shown as black foci on the pial surface of the brain. Example functional connectivity of a nonhub (D, left) and hub (D, right) stimulation site. Drift diffusion modeling of N-back performance following hub and nonhub inhibition with cTBS. A correlation matrix, where entries denote the pairwise relationships of input (acc = mean accuracy, mrt = mean response time, vrt = variation in response time) and output (v = drift rate, a = response boundary, ter = nondecision time) parameters of the model, was constructed to aid in the interpretation of changes in drift rate (E). A 2 × 4 repeated-measures ANOVA (target × load) performed on drift rates revealed that hub cTBS disrupted drift rates more than nonhub cTBS (F).

Open in new tabDownload slide

Hub Inhibition Disrupts Information Processing During Working-memory

We tested our hypothesis that inhibiting hubs with cTBS would disrupt information processing using a drift diffusion model. This model takes the mean and variance of response times for correct trials, and mean accuracy as inputs and calculates cognitively relevant latent variables indexing the rate of information processing (“drift rate”), response conservativeness (“boundary separation”), and nondecision time. Of these three diffusion parameters, drift rate was most relevant for our prediction that hub inhibition will disrupt information processing, as slower drift rates index worse N-back performance (Shine et al. 2016). Slower drift rates can be interpreted as more variable response times and fewer correct responses (Fig. 2E).

A 2 × 4 repeated-measures ANOVA (target × load) performed using drift rates revealed a main effect of target [F(1,23) = 9.13, P_corrected = 0.006, ω² = 0.04], such that hub inhibition reduced drift rates when compared with nonhub inhibition (Fig. 2F). Post hoc paired t-tests revealed that drift rates were slower following hub inhibition relative to nonhub inhibition for the 1-back [t(23) = 2.70, P = 0.01, Cohen’s d = 0.45], 2-back [t(23) = 2.51, P = 0.02, Cohen’s d = 0.43], and 4-back [t(23) = 2.88, P = 0.008, Cohen’s d = 0.51] conditions. This finding demonstrates, as predicted, that hub inhibition disrupts information processing more than nonhub inhibition. Notably, this effect remained in models where potentially confounding variables were included as between participant covariate of no interest, including anatomical positioning, baseline activation, and distance to TMS coil (see Supplementary Material 4). As expected, we also observed a main effect of load [F(1,23) = 21.33, P_corrected < 0.001, ω² = 0.09], confirming that higher cognitive loads were more difficult than lower loads. Target and load did not interact [F(1,23) = 0.03, P_corrected = 1.00, ω² = 0.00], however, indicating that the difference in drift rate after hub and nonhub inhibition did not vary significantly by cognitive load. The response boundary and nondecision time parameters were analyzed as well (see Supplementary Material 5), but neither revealed a main effect of target.

Hubs may Belong to Discrete, Functionally Relevant Categories

Our experimental design assumed that hubs belong to a single category of objects with global connections, akin to a hub-spoke network. Recent evidence, however, indicates that hubs instead tend to link control networks with discrete subsets of nonrandom brain networks and can be clustered using their cross-network profiles of functional connectivity into one of the 3 categories (Gordon et al. 2018). The 3 hub categories are termed according to their putative information processing roles—control-processing, control-default, and cross-control (Supplementary Material 6). We hypothesized post hoc that the degree to which hubs targeted with cTBS in the present investigation resembled these hub types could account for variation in drift rate change (drift rate_Hub – drift rate_Nonhub) averaged across cognitive loads. We tested this possibility in an exploratory analysis, using a stepwise linear regression model with hub-type resemblances (Fisher-transformed correlation coefficients) as predictors. Resemblance with the control-processing hub type was the only significant predictor (β = −0.41, ΔR² = 0.17, P = 0.04), indicating that inhibition of hubs induced greater information-processing deficits when those hubs were more similar to the control-processing hub type. The control-processing hub type is distinguished by linking select control networks, including the cingulo-opercular network and dorsal attention network, with processing systems representing external information, including the visual and somatomotor networks.

Discussion

We present 2 findings in this investigation. First, cortical hubs mapped using large quantities of per-individual rsfMRI are reproducible, but idiosyncratic features of functional brain organization. Second, inhibiting a hub area with cTBS disrupted working-memory performance, as measured using a drift diffusion model, more than inhibiting a nonhub area of the same gyrus. The theoretical and potential translational implications of these points are considered in the following.

Intra and Interindividual Variation in Single-subject Hub Estimates

The reproducibility curves for single-subject hub estimates reported here began low and rapidly reached asymptote, resembling sigmoid functions reported in similar analyses (Laumann et al. 2015). Measurement error, due to the amount of per-individual data utilized, is thought to be a primary source of intraindividual variance in rsfMRI functional connectivity estimates. Other studies have also reported that small quantities of per-individual rsfMRI data fail to reliably characterize functional brain organization (Anderson et al. 2011; Laumann et al. 2015), likely due to the poor signal-to-noise ratio in BOLD fMRI data (Welvaert and Rosseel 2013). Consistent with this perspective, we found that large quantities of rsfMRI data are needed for reproducible single-subject hub estimates. The practical significance of interindividual variation in topological features functional brain organization, including hubs, for future work will depend on whether precision is necessary for the context at hand. A group-average map may reflect a central-tendency in the spatial positioning of hubs sufficient for purposes of cartography (Van Essen 2013), but interventions employing high-resolution techniques, such as TMS (Wassermann et al. 1992; Thielscher and Kammer 2002), could benefit from targeting individual-specific hubs. With this concern in mind, we mapped hubs on an individual basis in our NIBS experiment.

Hub Inhibition Slows Drift Rates During Working-memory Performance

The computational model applied in our investigation assumes an accumulation of noisy information supporting a decision process—whether or not the current stimulus matches one occurring N trials ago. Conceptually, the drift rate parameter indexes the average amount of information accumulated per unit of time during this decision process (Forstmann et al. 2016). It is noteworthy that changes in drift rate, despite being conceptualized in terms of speed (i.e., faster or slower), are due to changes in both accuracy and response time distributions, such that slower drift rates reflect more variable and less correct responses. Thus, drift rates can be interpreted as an index of subject ability to perform a behavioral task. As an example, prior studies have found that sleep deprivation (Ratcliff and Van Dongen 2009) or alcohol consumption (van Ravenzwaaij et al. 2012) slow drift rates during the performance of cognitive tasks. Slower drift rates induced by hub inhibition could result from a disruption in the integration of task-relevant information distributed among brain networks bridged by the inhibited hub (Shine et al. 2016). This scenario could explain why inhibiting some hubs disrupted information processing more than others (i.e., if task-relevant information was represented in discrete subsets of brain networks). While working-memory was identified a priori as one form of complex cognition theoretically involving hubs, we believe that it is unlikely that hubs subserve a single cognitive process. Instead, the extent to which hubs are functionally specialized could be with respect to broader domains of cognitive processing, such as externally or internally oriented cognition (Gordon et al. 2018).

Inhibition of Different Hub Types Differentially Affects Behavior

The network neuroscience literature generally conceptualizes hubs as a single class of nodes with global connections (Cole et al. 2013; van den Heuvel and Sporns 2013). Consistent with this perspective, our NIBS experiment was designed to inhibit a hub in each individual, without explicitly considering which networks it bridges. By retrospectively comparing hubs with independently identified categories of hubs linking discrete subsets of brain networks (Gordon et al. 2018), however, we assessed whether the variation in the behavioral effects of hub inhibition could be related to differences in cross-network connectivity. The degree to which hubs resembled the “control-processing” hub type was a significant predictor; however, without control for multiple comparisons. Although speculative, the topological positioning of the control-processing hub type, at the intersection of select control and sensorimotor networks, is well suited for enabling information processing during a visual working-memory task requiring a motor response. An item maintained in working-memory is theoretically represented broadly in sensorimotor cortex and influenced by signals from control-related brain areas (Eriksson et al. 2015; Christophel et al. 2017). Thus, it is tempting to conclude that hub types could be specialized for specific forms of information processing but this finding should be interpreted as preliminary. In principle, a future experiment could test this possibility empirically by prospectively mapping different hub types within an individual and testing for a dissociation.

Implications for Translational Network Neuroscience

Network theory is a powerful and increasingly widespread conceptual framework in neuroscience (Medaglia et al. 2015; Bassett and Sporns 2017). This approach distills the complexity of the brain into simpler mathematical representations (Rubinov and Sporns 2010), which in turn allows investigators to form and test tractable hypotheses regarding how a brain might process information in a networked fashion. Validating the core predictions of this framework is a necessary step towards establishing its translational value (Castellanos et al. 2013; Matthews and Hampshire 2016). One of the most useful predictions in network neuroscience is that the role of a brain area can be inferred from its connections (Sporns 2011). From this perspective, hub nodes should be more important than more peripheral nodes for network function. Evidence supporting this prediction to date in the human brain has come from biophysical models (Honey and Sporns 2008; Misic et al. 2015) or lesions in stroke patients (Gratton et al. 2012; Warren et al. 2014), with no precise, causal manipulations of hubs mapped prospectively in single-subjects performed until the present investigation. Thus, beyond providing evidence for the importance of hub brain areas, this investigation highlights how network neuroscience could be leveraged in the future to inform personalized interventions in humans (Deco and Kringelbach 2014), including NIBS (Fornito et al. 2015), but also neurofeedback or rehabilitation.

Implications for rsfMRI-guided NIBS Interventions

There is growing interest in using rsfMRI to guide stimulation site selection in NIBS therapies (Gong and He 2015; Fischer et al. 2016; Dubin et al. 2017). This is because rsfMRI can efficiently characterize a functional brain organization present across many task states (Cole et al. 2014) but is not affected by task-related confounds (Fox and Greicius 2010). A notable limitation of the BOLD fMRI signal, however, is that it is an indirect measure of neuronal activity (Raichle 2010). Despite this limitation, we found that inhibiting hub and nonhub areas, which differed only in their rsfMRI connectivity, produced significantly different behavioral outcomes, despite being separated by only a few centimeters on the same gyrus in each subject. It is noteworthy that NIBS investigations generally use either a sham condition or an active control that is anatomically distinct (Polania et al. 2018). Thus, utilizing a nonhub area of the same gyrus as an active control is a highly rigorous element of our experimental design. The success of this approach highlights the precision at which rsfMRI and select NIBS techniques may be able to map and manipulate functionally discrete areas of cortex in patients. Finally, future work can evaluate whether NIBS interventions could benefit from stimulating hubs mapped using rsfMRI in single-subjects. This strategy would be a significant departure from existing proposals for network-centric targeting (Fox et al. 2012; Dubin et al. 2017), as hubs are positioned at the intersection of multiple networks, which could afford an opportunity to modulate a wider range of symptoms. This strategy is conceptually well suited for clinical populations, such as depression (Drysdale et al. 2017), with heterogeneous clinical profiles and abnormal functional connectivity between multiple brain networks (Liston et al. 2014).

Conclusions

This investigation joins an emerging field of precision neuroscience (Laumann et al. 2015; Braga and Buckner 2017; Gordon, Laumann, Gilmore, et al. 2017) that treats idiosyncrasies in functional brain organization as neurobiologically informative rather than epiphenomena. Our findings highlight how precisely mapping individual-specific features of a connectome could be leveraged in the future to guide interventions in the human brain. NIBS therapies may be particularly well suited to adopt this approach.

Funding

Dean Toulmin’s Pilot Project Award from Georgetown University Medical Center (to P.E.T. and C.J.V.).

Notes

The authors would like to thank the research staff at the Center for Functional and Molecular Imaging and Junaid Merchant for their assistance with data collection. The authors also thank Conor Liston and Xiaozhen You for their feedback on a revised version of this manuscript. Conflict of Interest: None declared.

References