1. Neuroscience
Download icon

Right inferior frontal gyrus implements motor inhibitory control via beta-band oscillations in humans

  1. Michael Schaum  Is a corresponding author
  2. Edoardo Pinzuti
  3. Alexandra Sebastian
  4. Klaus Lieb
  5. Pascal Fries
  6. Arian Mobascher
  7. Patrick Jung
  8. Michael Wibral
  9. Oliver Tüscher
  1. Systemic Mechanisms of Resilience, Leibniz Institute for Resilience Research, Germany
  2. Department of Psychiatry and Psychotherapy, University Medical Center of the Johannes Gutenberg University, Germany
  3. Ernst Strüngmann Institute (ESI) for Neuroscience in Cooperation with Max Planck Society, Germany
  4. Donders Institute for Brain, Cognition and Behaviour, Radboud University Nijmegen, Netherlands
  5. Campus Institute Dynamics of Biological Networks, Georg-August University, Germany
Research Article
  • Cited 2
  • Views 883
  • Annotations
Cite this article as: eLife 2021;10:e61679 doi: 10.7554/eLife.61679

Abstract

Motor inhibitory control implemented as response inhibition is an essential cognitive function required to dynamically adapt to rapidly changing environments. Despite over a decade of research on the neural mechanisms of response inhibition, it remains unclear, how exactly response inhibition is initiated and implemented. Using a multimodal MEG/fMRI approach in 59 subjects, our results reliably reveal that response inhibition is initiated by the right inferior frontal gyrus (rIFG) as a form of attention-independent top-down control that involves the modulation of beta-band activity. Furthermore, stopping performance was predicted by beta-band power, and beta-band connectivity was directed from rIFG to pre-supplementary motor area (pre-SMA), indicating rIFG’s dominance over pre-SMA. Thus, these results strongly support the hypothesis that rIFG initiates stopping, implemented by beta-band oscillations with potential to open up new ways of spatially localized oscillation-based interventions.

Introduction

Response inhibition is an essential component of cognitive control. Many neuropsychiatric disorders like attention deficit hyperactivity disorder, obsessive-compulsive disorder, and Parkinson’s disease demonstrate the severe impact of impaired response inhibition on mental health (Schachar and Logan, 1990; de Wit et al., 2012; Voon and Dalley, 2011). Response inhibition has been operationalized via the classical stop-signal task (SST, Logan et al., 1984) in many studies; thus the SST plays a major role in defining research on cognitive control (Verbruggen et al., 2019). In the ‘GO’ condition of this task, participants have to respond to a GO signal by executing an immediate motor response, for example, performing a button press. In the less frequent ‘STOP’ condition, the GO signal is followed by a STOP signal after a variable delay and participants have to cancel the planned or ongoing motor response. A well-developed theoretical framework describing the cognitive processes underlying response inhibition relies on a so-called independent race model, which assumes the independence of go and stop processes (Logan et al., 1984; Verbruggen et al., 2019). Depending on which process finishes first, the stop process, which is initiated after the variable stop-signal delay, or the go process, which is initiated after the GO signal, the response can be inhibited successfully or not. Based on this framework, the latency of the unobservable stop response, the stop-signal reaction time (SSRT), can be estimated by subtracting the mean stop-signal delay (SSD) from the mean GO RT (in the simplest way).

So far, numerous studies using the SST provided consistent evidence that response inhibition crucially depends on two cortical regions: the inferior frontal cortex (IFC) and the pre-supplementary motor area (pre-SMA) (Wessel and Aron, 2017). The particular functional roles of these two regions within response inhibition have been controversially debated for more than a decade now (Aron et al., 2004; Hampshire and Sharp, 2015a; Aron et al., 2015; Hampshire and Sharp, 2015b). While a range of studies on response inhibition point toward a single inhibitory module within the IFC that directly initiates stopping (Aron et al., 2003), other studies assign this role to pre-SMA (Li et al., 2006; Nachev et al., 2007). As an alternative, some authors suggest that the functional role of IFC within response inhibition is either the encoding of context-specific task rules or simply the attentional detection of the need for response inhibition, whereas its initiation is triggered by the pre-SMA (Duann et al., 2009; Sharp et al., 2010; Rae et al., 2015; Xu et al., 2017). A more critical view questions the behavioral construct of response inhibition in general (Hampshire and Sharp, 2015a). Here, the authors suggest that there is no unitary response inhibition module and claim that the IFC is part of a domain-general control network (Erika-Florence et al., 2014). However, in the attempts to experimentally define the differential roles of IFC and pre-SMA, the temporal activation order of those regions has become a decisive, yet unresolved question in the field (Allen et al., 2018; Swann et al., 2012).

Two methodological issues may explain why this topic has so far remained unresolved. One is a lack of temporal resolution in many respective studies, the other is the specific design of the standard SST. The majority of studies used fMRI, while only few studies in humans addressed the temporal activation order of IFC and pre-SMA using techniques providing both, high spatial and high temporal resolution at the same time. One study used electrocorticography (ECoG) in a single patient and concluded that pre-SMA precedes the IFC activation (Swann et al., 2012). In contrast, two recent MEG studies suggest that both regions may be simultaneously active and that there is no temporal or functional primacy (Allen et al., 2018; Jha et al., 2015). However, the latter two studies may be limited by the design of the standard SST, which does not allow to disentangle attentional from inhibitory processes, and hence, to unequivocally define the functional roles of IFC and pre-SMA. To specifically control for the attentional load involved in the SST, Sharp et al., 2010 used a selective stopping task. This task comprises additional attentional capture go (AC-GO) trials that are identical to STOP trials in terms of timing and frequency, but participants have to continue with their already prepared go response (Figure 1A). Contrasting both conditions (i.e., STOP versus AC-GO trials) has two main advantages. First, it allows us to identify the neural signature of response inhibition independently of other cognitive processes that are related to attention or conflict (Sharp et al., 2010). Second, this contrast enables a comparison between conditions where stop and go processes are disturbed in the same way due to a salient stimulus. More importantly, the timing of the initiation of stop and go processes is comparable between conditions. This is not the case when successful stop (sSTOP) trials are contrasted with GO or unsuccessful stop (uSTOP) trials (Sánchez-Carmona et al., 2016). Note that the race model assumes that in sSTOP trials, the go process is slower than the stop process, while in uSTOP trials it is vice versa. Since we aimed to analyze the temporal activation order of IFC and pre-SMA with respect to the initiation of response inhibition, sSTOP trials need to be compared to an appropriate control condition that has on average the same signal delay (Xu et al., 2017) and is matched in terms of the onset of the related processes. This is particularly important for high temporal resolution studies.

Experimental design and behavioral performance of the selective stopping task.

(A) The task comprised three conditions: a GO condition (white arrow), a STOP condition (arrow changes color to blue), and an attentional capture GO (AC-GO) condition (arrow changes color to green). The variable stop signal delay (SSD) was adapted to the participants’ performance to yield a probability of 50% of successful response inhibitions per block. SSRT, stop signal reaction time, ACSD, attentional capture signal delay (analog to SSD). (B) Temporal region of interest (orange box) for latency and connectivity analysis starts at 100 ms, where early visual processing is likely complete and ends at SSRTmax (maximal SSRT across participants, 350 ms). (C) Reaction time (RT) and accuracy rate in correct GO (cGO), correct attentional capture GO (cAC-GO), and unsuccessful stop trials (uSTOP); mean ± standard error of the mean. ***p<0.001, based on Bonferroni-corrected post hoc comparisons of a repeated-measures ANOVA, n = 59.

Figure 1—source data 1

Behavioral data used for the statistics in panel (C).

https://cdn.elifesciences.org/articles/61679/elife-61679-fig1-data1-v2.csv

So far, to the best of our knowledge, only three studies used a selective stopping task with an AC-GO (‘continue’) condition in connection with a high temporal resolution technique. Two studies used ECoG in a single patient (Swann et al., 2012), or in four patients (Wessel et al., 2013), respectively, but Wessel et al., 2013 could only cover the right IFC region, not pre-SMA. Sánchez-Carmona et al., 2016 used EEG, but did not analyze the temporal activation order of both regions. However, the majority of studies providing high temporal resolution relied on a standard SST (Allen et al., 2018; Bartoli et al., 2018; Fonken et al., 2016; Jha et al., 2015).

The current study aims to answer long-standing unresolved questions concerning the roles and timing of IFC and pre-SMA during response inhibition, that is, which region initiates inhibitory control and which region exerts a putative causal influence on the other. We therefore used a stimulus selective stopping task in combination with high temporal resolution MEG recordings. To validate MEG source reconstruction, we used the high spatial resolution of fMRI scans in a multimodal imaging approach. Based on previous findings, we hypothesized a predominant role of beta-band activity in both regions during the time of stopping (Kühn et al., 2004; Swann et al., 2009; Fonken et al., 2016) and aimed to define the temporal activation order and connectivity of IFC and pre-SMA during response inhibition.

Results

Behavioral results

Healthy participants (n = 62; three of them were excluded from further analysis, see 'Materials and methods') performed a stimulus selective stopping task with an attentional control condition (Figure 1A). Depending on different signal colors, participants were instructed to respond with a button press, to try to inhibit or to continue their already initiated go response. The mean SSD was 294.5 ± 120.6 ms (mean ± SD) and led to a probability of responding on a STOP trial close to 50% (49.0 ± 3.0%) proving the adherence of participants to the task rules and the successful operation of the staircase procedure (see 'Materials and methods' for more information). While for sSTOP trials the mean SSD was 288.1 ± 124.6 ms, it was 300.9 ± 117.1 ms for uSTOP trials. The mean attentional capture signal delay (ACSD) for correct attentional capture go (cAC-GO) trials was 296.3 ± 118.4 ms. A repeated-measures ANOVA based on RT revealed a main effect of condition (F = 471.5, p<0.001). Bonferroni-corrected post hoc comparisons revealed that mean RT in cAC-GO trials (588.2 ± 14.0 ms, group mean ± standard error of the mean) was significantly longer as compared to GO trials (550.2 ± 13.5 ms) and uSTOP trials (499.6 ± 12.6 ms). Mean RT in uSTOP trials was significantly shorter as compared to GO trials (p<0.001 for all post hoc tests; Figure 1C). Participants performed accurately as indicated by low omission error rates in GO (2.3 ± 2.8%, mean ± SD) and AC-GO trials (1.2 ± 1.4%). Accuracy was 97.2 ± 2.8% in GO and 97.0 ± 2.8% in AC-GO trials (Figure 1C). All these behavioral results of the MEG study are in line with the ones of the fMRI study (Sebastian et al., 2016). SSRT was calculated using the integration method (see 'Materials and methods' for more information), resulting in a median SSRT of 244.2 ± 36.0 ms. The maximal SSRT across participants (SSRTmax) was 349.9 ms.

To exclude an effect of signal color, the experiment was designed in a cross-balanced manner, that is, the attribution of signal color for stopping (blue/green) to trial type (STOP/AC-GO) was balanced across subjects. A two-sample t-test with signal color as between factor revealed that the stopping latency as measured by the SSRT did not differ significantly between groups (t = 0.088, p = 0.930). A mixed-design ANOVA with signal color as between factor and RT (GO versus AC-GO) as within factor further revealed no influence of attribution of color to trial type, as no interaction of these two factors was present (F = 0.847, p = 0.361).

Source activation

To analyze the oscillatory dynamics and connectivity of response inhibition, we first identified a set of cortical sources involved in successful stopping. Hence, we contrasted spectral source power of sSTOP and cAC-GO trials. This approach also enabled us to identify coordinates at which stopping-related sources were actually active in the acquired MEG data. This is in contrast to seed-based analyses which rely on a priori defined coordinates from an anatomical atlas (mask) or other studies as reference. Of these two approaches, contrasting appropriate conditions is the recommended best practice (Gross et al., 2013). Contrasting conditions enabled us to identify coordinates of peak activity for rIFG and pre-SMA, that is, the locations with the strongest effects across subjects; relying on this approach, compared to a seed-based one, also excluded that the effects we analyzed were due to leakage from other sources.

To find appropriate parameters for source reconstruction using a beamformer method, the spectral power at sensor level was analyzed for both conditions pooled, sSTOP and cAC-GO trials: We a priori defined a temporal region of interest (tROI) starting after the presumed completion of early visual processing (Amassian et al., 1989; Swann et al., 2009; Allen et al., 2018) and ending with SSRTmax (100–350 ms, see 'Materials and methods' for more information). We then compared spectral power between the tROI and an equally long baseline epoch (see Figure 1BFigure 2—figure supplement 1A) to identify frequency bands of interest for further analyses. The cluster-based permutation test revealed one cluster with significant power decrease (12.0–31.9 Hz, beta band) and one cluster with significant power increase (63.8–87.7 Hz, gamma band).

Table 1
MEG source reconstruction in the beta band.

MNI coordinates obtained as peak voxels when contrasting successful stop (sSTOP) versus correct attentional capture GO (cAC-GO) trials.

 Coordinates
Region (label)xyzt-Value
Inferior frontal gyrus BA44/45 R (rIFG)5020203.654
Inferior frontal gyrus BA44/45 L (l-IFG)–5010202.820
Anterior insula/left IFG (lAIns)–403002.894
Pre-supplementary motor area (pre-SMA)020603.878
Middle frontal gyrus (lMFG)–400503.360
Premotor cortex BA6 R (rPMC)10–10504.015
Premotor cortex BA6 L (lPMC)–20–10603.571
Fornix00103.740

To then reveal effects related to response inhibition, we reconstructed source power during the tROI and contrasted it between both conditions (sSTOP and cAC-GO) for each of the two frequency bands identified above. This contrast revealed that beta-band power was increased in sSTOP compared to cAC-GO trials in a network of pre-motor and pre-frontal areas (Figure 2A). One main cluster of increased beta-band power was located in the IFC with its maximum in the right inferior frontal gyrus (rIFG), pars opercularis. Increased beta-band power was also present in the pre-SMA, left middle frontal gyrus (lMFG), and bilateral premotor regions (Table 1, Figure 2A). The same analysis performed for the gamma band did not reveal a significant cluster. To further validate the MEG source reconstruction, a subset of participants in the MEG experiment (n = 31) plus 45 additional participants were recorded using fMRI (resulting in n = 76). Here, the contrast sSTOP versus cAC-GO revealed broad activation of bilateral prefrontal, sensorimotor, parietal, and occipital areas (Figure 2B). In particular, rIFG was activated in two subregions (pars triangularis, BA45, x = 52, y = 22, z = 4, t = 10.96; pars opercularis, BA44, x = 52, y = 18, z = 14, t = 9.42), while pre-SMA (x = 16, y = 14, z = 64, t = 9.45) was identified as sub-maxima in the biggest cluster obtained (13,438 voxels) as well as in a smaller cluster (85 voxels) (x = –‍14, y = 18, z = 66, t = 5.63). For a whole-brain comparison between both methods, MEG and fMRI, we generated an overlap visualization (Figure 2C). The overlap was defined by the product of absolute, normalized t-values for each voxel i, |ti,MEG|×|ti,fMRI|. MEG data were interpolated to match the fMRI resolution. Normalization was performed for each method separately by division with the maximum t-value obtained by the respective method, resulting in normalized t-values between 0 and 1. The comparison of activity revealed by the MEG and fMRI contrasts showed an overall good correspondence. Given the 1 cm grid used for the MEG source reconstruction, the distance between peaks of both methods were rather close for rIFG (7 mm), pre-SMA (15 mm), lAIns (11 mm), and lMFG (15 mm). The peaks of premotor areas (rPMC, lPMC) as well as l-IFG were less close to the fMRI results (20–24 mm difference between methods), but their activity was rather widespread. Yet for the key areas that were most important for this study, rIFG and pre-SMA, we found a very good correspondence and thus consider the MEG source reconstruction as quite accurate for these areas.

Figure 2 with 1 supplement see all
Sources involved in successful stopping.

(A) Source reconstruction of MEG data in the beta band (12–32 Hz, 100–373 ms after STOP/AC-GO [attentional capture go] signal onset). Surface plots show t-values of significant clusters when contrasting successful stop (sSTOP) and correct attentional capture go (cAC-GO) trials (cluster-based permutation test, two-tailed, αcluster = 0.05, n = 59). Peak voxels (local extrema) of these clusters are highlighted and labeled (MNI coordinates are shown in Table 1). Dashed circles indicate that the peak voxel is not directly visible on the surface. Anatomical regions: rIFG, right inferior frontal gyrus, pre-SMA, pre-supplementary motor area, l-IFG, left inferior frontal gyrus, lAIns, left anterior insula, lMFG, left middle frontal gyrus, lPMC, left premotor cortex, rPMC, right premotor cortex. (B) fMRI activation maps for the contrast sSTOP versus cAC-GO; map thresholded pFWE < 0.05, cluster extent k = 5 voxels, n = 76. Positive activation, that is, positive t-values (red color), is obtained by the contrast sSTOP > cAC-GO, while negative activation, that is, negative t-values (blue color), is obtained by the contrast sSTOP < cAC-GO. Data are taken from Sebastian et al., 2017. (C) Overlap between MEG (A) and fMRI (B) activity. To compare the activity revealed by both methods, the product of normalized t-values from both methods is displayed. Peak voxels of right inferior frontal gyrus (rIFG) and pre-supplementary motor area (pre-SMA) as found in the MEG analysis are highlighted and labeled.

Temporal precedence of rIFG over pre-SMA activation

To identify the cortical region that initiates response inhibition, we analyzed the temporal development of beta-band power across the sources obtained by contrasting z-transformed sSTOP with cAC-GO trials. The earliest beta-band power difference was found in the rIFG source compared to all other sources (Figure 3C). According to our hypothesis, we then tested explicitly whether the beta-band power difference occurred earlier or later at the rIFG than at the pre-SMA. Since cluster-based permutation tests on time-frequency representations (TFRs) do not establish significance of latency differences between conditions (Sassenhagen and Draschkow, 2019), we tested the temporal lead of rIFG activity onset by contrasting z-transformed spectral power time courses in sSTOP trials with those in cAC-GO trials. On the spectral power time courses, onset times were defined by using a threshold defined by certain percentage of the range between zero power and the first local power maximum found in the tROI (see 'Materials and methods' for more information). These onsets were then used as input for a permutation test that revealed a significant latency onset difference between both sources (p=0.0240, two-tailed test, mean onset latency rIFG: 137 ± 46 ms, and pre-SMA: 159 ± 52 ms). The significant difference reported here was based on a 25% onset threshold, but the onset difference was also significant for 10%, 30%, and 50% thresholds. For this test, seven subjects had to be excluded because no positive peak could be found within the tROI (n = 52).

Beta-band time-frequency representations of virtual channels placed at sources identified.

First and second solid lines indicate stop signal reaction time (SSRT) and SSRTmax; first and second dotted lines indicate 10% and 50% percentiles of RTAC-GO for selected AC-GO trials with RTAC-GO > SSRT (RTAC-GO is the duration between AC-GO signal and button press, see Figure 1A). Time axis locked to STOP and AC-GO signal (0 s). All plots show the results of a cluster-based permutation test, two-tailed, αcluster = 0.05, n = 59, significant clusters are outlined. (A) Task-versus-baseline power for successful stop (sSTOP) trials. (B) Task-versus-baseline power for correct attentional capture go (cAC-GO trials). (C) Contrast of z-transformed sSTOP versus cAC-GO trials. (D) Averaged beta-band power (12–32 Hz). Orange curve (solid line): sSTOP trials, green curve (dashed line): cAC-GO trials, bounded lines: standard error of the mean.

To further ensure that the earlier onset latency of rIFG compared to pre-SMA does not rely on the specific time-frequency transformation method used and the parameters applied for onset latency definition, we used an orthogonal approach to define the onset latency of both regions. Here, we used broadband signal time courses instead of a frequency-specific analysis. We performed a time-resolved support vector machine (SVM) analysis in order to classify sSTOP versus cAC-GO trials and to identify the discriminability onset between both conditions, separately for rIFG and pre-SMA. This SVM-based approach revealed a robust above-chance level classification starting from 129 ms (p<0.0001) after STOP/AC-GO signal onset for the rIFG source and from 163 ms (p<0.0001) for the pre-SMA source (Figure 4A, tested within tROI, 100–350 ms). To statistically compare individual classification onsets between the two sources, we defined onsets using the final classification time series within the tROI. This analysis showed a significantly earlier rIFG onset compared to pre-SMA at a threshold of 25% with p=0.0377 (p=0.0062). Mean onset in rIFG was 178 ms (186 ms) and 200 ms (220 ms) in pre-SMA (first value refers to an analysis that relied on classification peaks with an accuracy above one standard deviation of the mean baseline accuracy, while the second value in parentheses refers to two standard deviations). Three (17) subjects were excluded because they did not meet inclusion criteria, see 'Materials and methods' for more information. These results were not biased by the selected threshold since all other thresholds gave similar results. Since the onset definition relied on low-pass filtered data for smoothing, onsets in average are later than the values obtained by the above-chance level test. Additionally, to check if this effect was beta-band driven, we performed the same analysis on beta-band filtered data, but still in the time domain. Accuracy was significantly above chance level in the rIFG at 132 ms (p<0.0001) and in pre-SMA at 142 ms (p<0.0001), tested within tROI (Figure 4B). However, since the beta-band filtered data was quite noisy, most of the subjects were excluded from the individual onset statistics when applying the same onset definition as for the broadband data. Thus, the onset of rIFG compared to pre-SMA using beta-filtered data was nominally earlier, but this difference did not reach statistical significance. This result indicates that the temporal precedence effect in broadband SVM analysis we found was not only driven by beta-band oscillations. Yet, the beta-band filtered data allowed to reliably classify sSTOP trials, indicating that beta-band oscillations are carrying the crucial information to reliably classify sSTOP trials.

Time-resolved support vector machine (SVM) analysis.

Both trial types (successful stop [sSTOP] and correct attentional capture go [cAC-GO]) were robustly classified above chance within the temporal region of interest (tROI) (100–350 ms) across subjects at right inferior frontal gyrus (rIFG) (blue curve, solid line) and pre-supplementary motor area (pre-SMA) (orange curve, dashed line). As input, time-embedded vectors of 14 consecutive samples were used, corresponding with one beta cycle at the center frequency of 22 Hz (see 'Materials and methods'). Significance for above chance-level accuracy was tested between 100 and 350 ms. Significant classifications within this range are highlighted by colored bars above the curves and onsets are indicated by arrows. Bounded lines: 95% confidence intervals of the mean, n = 59. First and second dashed lines indicate stop signal reaction time (SSRT) and SSRTmax. (A) Broadband time course data as input for the SVM. While above-chance level onset for rIFG was at 129 ms (p<0.0001), pre-SMA was above chance at 163 ms (p<0.0001). (B) Same analysis applied on beta-band filtered data (12–32 Hz). While above-chance level onset for rIFG was at 132 ms (p<0.0001), pre-SMA was above chance at 142 ms (p<0.0001).

In sum, the obtained results were comparable between all three analyses, that is, broadband and beta-filtered SVM as well as the onsets obtained by the TFR analysis with respect to an earlier onset in rIFG. Overall, the analysis of time domain data corroborates the significantly earlier activation of rIFG in the beta-band power analysis, suggesting that this area plays a leading role in initiating response inhibition.

Beta-band source power predicts inhibitory performance

To understand the effects of changes in beta power on the SSRT in more detail, we performed a Bayesian multiple regression analysis with per subject SSRT values as outcome variable and beta power (STOP–cAC-GO) in pre-SMA and rIFG in successful and failed trials, as well as the directed influence asymmetry index (DAI) (see below) as regressors. We compared three models for their expected posterior log-probability using a leave-one-out cross-validation (LOO-CV) procedure as suggested in Gelman et al., 2014. Model 1 used all regressor variables and their first-order interaction terms, model 2 used all regressor variables, but no interaction terms, and model 3 additionally averaged together the beta power in pre-SMA for successful and failed trials as the values of these regressors were highly correlated (r > 0.6). Thus, model complexity was reduced successively from model 1 to model 3.

Model 3 achieved the best LOO-CV score (expected log posterior predictive density elpdloo=64.65±5.32, higher is better, maximum is at zero), but model 2 was within one standard error of model 3 (elpdloo=65.24±5.10). Model 1 had a considerably worse LOO-CV score (elpdloo=71.67±4.53). In the winning model 3, there was as 91% marginal posterior probability for a negative regression coefficient of the beta power in rIFG for successful trials, and a 87% marginal posterior probability for a negative regression coefficient for the DAI, but only a 64% marginal posterior probability for a negative regression coefficient for the average beta power in pre-SMA, and a 40% marginal posterior probability for a negative regression coefficient for the beta power in rIFG in failed trials (for more details, consult the linked notebook in Box 1). Given these results, we consider it most plausible that beta power in rIFG in successful trials and the DAI predict SSRT outcomes, with all other regressors having no or a much smaller influence. This finding strongly supports our hypothesis that modulation of beta-band activity is related to response inhibition.

Box 1.

Toolboxes and custom functions used for analysis of the MEG and fMRI data.

Toolboxes

MEG preprocessing, source reconstruction, and TFRs

     FieldTrip (Oostenveld et al., 2011), FieldTrip: RRID:SCR_004849

     https://www.fieldtriptoolbox.org

fMRI preprocessing and activity map

     SPM8 toolbox, SPM8: RRID:SCR_007037

     http://www.fil.ion.ucl.ac.uk/spm/

Statistics of behavioral data

     JASP version 0.12.2, JASP: RRID:SCR_007037 PMID:28685272 (Allen et al., 2018)

     https://jasp-stats.org/

Support vector machine

     Multivariate pattern analysis for MEG (Guggenmos et al., 2018)

     https://github.com/m-guggenmos/megmvpa , SVM / Megmvpa: no RRID found

     LIBSVM: A Library for Support Vector Machines (Chang and Lin, 2011)

     https://www.csie.ntu.edu.tw/~cjlin/libsvm/, LIBSVM: RRID:SCR_007037

Conditional Granger causality

     As implemented in FieldTrip, using a blockwise approach (Wang et al., 2007)

     https://www.fieldtriptoolbox.org/example/connectivity_conditional_granger/

Bayesian multiple regression

     PyMC3 (Salvatier et al., 2016)

     https://docs.pymc.io, PyMC3: RRID:SCR_018547

Custom functions

Sensor statistics

     FFTAnalysis.m, FFTStatsTaskvsBaselinePooled.m

Source statistics

     Statistics_Step1_Analytic_BC.m, Statistics_Step2_DepSamplesT_BC.m

     SourceLocalization.m

TFRs and onset latency analysis

     TFRStatsConditionContrast.m, TestOnsetLatencies.m

fMRI activity map

     ttest_stopVSac_SPM_job.m

Support vector machine

     run_mvpa.m

Conditional Granger causality

     masterGC.m

Bayesian multiple regression

MultiRegressionSSRT.ipynb

Source code is stored in a GitHub repository:

https://github.com/meglab/acSST (copy archived at swh:1:rev:ea0bf4acc0f11cdc78ad31b6c1285f1851389312; Schaum, 2021)

Source data (MATLAB files) are stored at the Dryad open-access repository:

     https://doi.org/10.5061/dryad.x3ffbg7gp

For more information, please consult the readme.txt file in the GitHub repository.

rIFG is the dominant sender of information during response inhibition

To further characterize the functional role of rIFG and pre-SMA during response inhibition, we analyzed their connectivity pattern using nonparametric conditional Granger causality (cGC). We contrasted spectral cGC of sSTOP and cAC-GO trials in the tROI. Therefore, we computed the cGC from rIFG to pre-SMA and vice versa, conditional on all other five active cortical sources as obtained by the source reconstruction contrast. The cluster-based permutation test (see 'Materials and methods' for more information) revealed statistical differences between sSTOP and cAC-GO conditions within the (high) beta band in the direction rIFG to pre-SMA (pcluster = 0.0074, corrected for the two links tested), but not in the opposite direction (Figure 5A).

Connectivity analysis.

(A) Spectral-resolved conditional Granger causality (cGC) between right inferior frontal gyrus (rIFG) and pre-supplementary motor area (pre-SMA). Orange curve (solid line): cGC for successful stop (sSTOP) trials, green curve (dashed line): cGC for correct attentional capture go (cAC-GO) trials, bounded lines: standard error of the mean. A cluster-based permutation test was used to identify significant differences between both conditions. Significant differences (p<0.05) are highlighted in blue if Bonferroni corrected, that is, the two links tested, and in gray if uncorrected. The frequency range tested (8–44 Hz) is indicated by vertical dashed lines. n = 58 for the link rIFG to pre-SMA, n = 57 for pre-SMA to rIFG (subjects with average cGC below bias level in both conditions were excluded). (B) Stop-signal reaction time (SSRT) correlated with directed influence asymmetry index (DAI) for rIFG and pre-SMA cGC values (sSTOP trials, cGC at 34 Hz in the temporal region of interest [tROI], 100–350 ms). Positive DAI corresponds to links from rIFG to pre-SMA, while negative DAI corresponds to links from pre-SMA to rIFG. The negative correlation indicates that participants with higher cGC from rIFG to pre-SMA are the better inhibitors (faster SSRT). n = 55 with average cGC reliably above bias level (note that only sSTOP trials has been used here). (C) Temporal evolution of spectral-resolved cGC was analyzed post hoc between rIFG to pre-SMA (top) and pre-SMA to rIFG (bottom). A sliding window was shifted with 100 ms steps around the tROI. cGC from rIFG to pre-SMA (but not in the opposite direction) was significantly higher for sSTOP compared to cAC-GO trials in the tROI itself and in a later time window tested. Here, we corrected for testing six time windows and for testing in two directions (rIFG to pre-SMA and vice versa). Same formatting as in (A). Spectra of the tROI, highlighted with a gray-colored box, are also shown in (A).

To assess if the connectivity between rIFG and pre-SMA was correlated with SSRT, we used the DAI (Bastos et al., 2015) that captures the predominant direction of GC and allows comparison of DAIs across frequencies (see Materials and methods). We correlated the DAI of the sSTOP condition with SSRT in the significant frequency range identified (in steps of 2 Hz, in the range from 30 to 40 Hz, resulting in six discrete frequencies; Figure 5A). Using a bootstrap method (Efron and Tibshirani, 1993), we found a significant negative correlation between DAI and SSRT at 34 Hz (r=0.332, within a 99.17% confidence interval, equivalent to Bonferroni correction for testing six frequencies, [–‍0.558; –‍0.053]). Hence, the better subjects were able to inhibit their response, that is, the faster their SSRT was, the higher, positive DAI values they had, indicating a stronger connectivity from rIFG to pre-SMA, whereas slower inhibiting subjects showed the reverse pattern (negative DAI, stronger connectivity from pre-SMA to rIFG, Figure 5B). Additionally, this finding was supported by our Bayesian analysis, which showed that the DAI regressor indeed provided relevant information on the SSRT outcomes, even when combined with other relevant regressors in the full Bayesian model. This demonstrates that the asymmetry of communication between rIFG and pre-SMA affects the stopping process.

Finally, to test whether rIFG to pre-SMA connectivity was specific to the tROI of response inhibition, we tested the temporal evolution of the cGC contrast between sSTOP and cAC-GO trials. Therefore, we performed a sliding window analysis, moving backward and forward from the main tROI (100–350 ms) in steps of 100 ms (Figure 5C). This post hoc analysis revealed decrease in rIFG to pre-SMA connectivity in two additional windows. In the first additional time window from 200 to 450 ms, the flanks of the beta band showed a difference in cGC between conditions. Note that this first effect was not significant after the correction for multiple comparisons of six time windows and of each direction tested (p(lower cluster) = 0.0052, p(upper cluster) = 0.0207, both uncorrected). However, in a second, later window (300–550 ms), the upper flank of the beta band showed a significant difference between conditions (p(upper cluster) = 0.0384, corrected, and p(lower cluster) = 0.0064, uncorrected). No significant differences were found in any time window before the tROI and after 550 ms, nor in the direction of pre-SMA to rIFG. Overall, our cGC findings suggest that response inhibition is initiated within a crucial time range by specific rIFG to pre-SMA connectivity, the strength of which correlates with behavioral performance.

Discussion

The neural network mechanisms of response inhibition are central to the understanding of the neurobiology of cognitive control and have been the matter of a long-standing debate. Here, we tested specifically which of the two main regions proposed as the cortical initiator of response inhibition, rIFG and pre-SMA, respectively, is activated first and which exerts a putative causal influence on the other.

Our study, using the high temporal resolution of MEG and a relatively large cohort of subjects, is the first to show a significantly earlier activation in the rIFG compared to the pre-SMA. The onset of beta-band activity in the rIFG was found at ∼140 ms after the STOP signal, which is in the same range (∼120 ms) as revealed by an independent component analysis (ICA) of a recent EEG study (Jana et al., 2020). The onset of beta-band activity in the pre-SMA started at least 22 ms later – as independently verified by a latency analysis of spectral power and a time-resolved SVM analysis. Connectivity analysis showed that rIFG sends information in the beta band for successful stopping to pre-SMA but not vice versa. The behavioral significance of beta-band power in rIFG as well as of the connectivity from rIFG to pre-SMA in the beta band was demonstrated by their prediction of stopping performance.

rIFG initiates response inhibition

So far, only a limited number of previous studies have investigated the temporal activation order of rIFG and pre-SMA in response inhibition at source level with high temporal resolution (MEG or ECoG) by using an SST (Allen et al., 2018; Jha et al., 2015; Swann et al., 2012). Two of these studies (Allen et al., 2018; Jha et al., 2015) reported that the pre-SMA and rIFG were simultaneously activated in the time window between STOP signal and SSRT. In contrast to these MEG studies, one ECoG study in a single patient concluded that pre-SMA precedes the IFC activation (Swann et al., 2012). There may be three reasons for these diverse findings. First, these previous studies rely on a comparably small sample sizes (n = 19, n = 9, n = 1, respectively) compared to the current study (n = 59). A second reason is the use of different SST paradigms, that is, non-selective (standard) SSTs (Allen et al., 2018; Jha et al., 2015) versus a selective stopping task (Swann et al., 2012 and current study). Contrasting sSTOP trials with uSTOP or GO trials in non-selective SSTs entails several issues related to separating attentional effects from inhibition (see end of this discussion). A third reason is that results may depend on the way onset latencies are defined (see discussion of limitations by Bartoli et al., 2018). We therefore validated our main findings that were based on TFRs via an additional approach. In this additional approach, sSTOP and cAC-GO trials were classified using a time-resolved SVM. The sSTOP and cAC-GO trials could be classified 22–34 ms earlier from rIFG activity than from pre-SMA activity in the SVM analysis, independently replicating our result on latency differences in the TFR analysis. Since the temporal onset of discriminability between conditions was found already before the SSRT, this suggests that rIFG activity is indeed pivotal for initiating stopping.

Beta-band power in rIFG relates to inhibitory performance

We further hypothesized that if activity in the rIFG is a neural correlate for response inhibition, its power should also predict the inhibitory performance at behavioral level, that is, SSRT. Indeed, SSRT was better predicted by the beta-band power of successful trials in rIFG compared to pre-SMA. In the rIFG, the beta-band power of successful trials was negatively correlated with the SSRT. This is in line with several fMRI studies, where a negative correlation between rIFG activation and SSRT was found (Aron and Poldrack, 2006; Aron et al., 2007; Rubia et al., 2007), suggesting that subjects with greater activation in rIFG inhibited more quickly (shorter SSRT). More importantly, this relationship was also revealed by recent EEG/transcranial magnetic stimulation (TMS) studies (Hannah et al., 2020; Sundby et al., 2021), which found that the timing of beta bursts correlated with SSRT. Hence, the result of our Bayesian multiple regression additionally supports the dominant role of the rIFG for initiating stopping. Notably, both, sSTOP and uSTOP, trials had the same temporal activation pattern, that is, rIFG before pre-SMA. However, in the sSTOP trials, the rIFG beta-band power predicted the SSRT better than in the uSTOP trials. Hence, the temporal precedence of rIFG over pre-SMA may indeed be relevant for stopping in general, while at the same time, the degree of control by the rIFG is responsible for the success and the performance of stopping.

Dominance of rIFG during response inhibition supports its role as controller

To further elucidate the role of rIFG in response inhibition, we analyzed the connectivity between rIFG and pre-SMA using spectrally resolved conditional Granger causality analysis (cGCA). This method measures directed information transfer between two regions, that is, it quantifies predictive relationship within the network (Granger, 1969; Geweke, 1982). If the rIFG indeed has the function of a top-down controller of response inhibition, then it should exert an ongoing causal influence on the pre-SMA across the time window of response inhibition. Indeed, a significant difference between sSTOP and cAC-GO in cGC influence from rIFG to pre-SMA was detected, but not in the opposite direction. In agreement with our hypothesis, significant results were found within the beta band in particular. The observed directionality is in line with cGCA in a comparably high-powered fMRI study (Duann et al., 2009) (n = 57), yet inconsistent with GCA of a somewhat lower powered MEG study (Allen et al., 2018) (n = 19). In addition to the finding of significant beta-band cGC from rIFG to pre-SMA, we found that the strength of the directional asymmetry in cGC (DAI) predicted stopping performance as measured by SSRT. This is compatible with a recent DCM (dynamic causal modeling) analysis of an EEG study available in preprint (Fine et al., 2019) that also found (inhibitory) connectivity from rIFG to pre-SMA during STOP trials to be negatively related to SSRT when transcranial focused ultrasound was applied. These findings suggest that connectivity directed from rIFG to pre-SMA per se reflects a state in which rIFG is prepared to send a signal to pre-SMA to execute stopping, while the strength of this directionality influences the performance of stopping. To test whether the directionality from rIFG to pre-SMA was specific for response inhibition, or just related to other processes, we analyzed connectivity between rIFG and pre-SMA before and after the (a priori defined) tROI (100–350 ms) using a sliding window approach and indeed found the dominant effects centered on the tROI. Thus, we could confirm that inhibition-related beta-band connectivity between rIFG and pre-SMA was specific to the tROI and a subsequent time window, particularly only in the direction from rIFG to pre-SMA.

In sum, the results of the spectrally resolved cGCA further support the leading role of rIFG in response inhibition and the importance of beta-band oscillations in this process. Since connectivity from rIFG to pre-SMA was specifically established during the tROI and correlated with inhibitory performance, rIFG may be characterized as a brake in response inhibition. This interpretation is in line with Aron et al., 2014 and was recently corroborated by an ECoG study that revealed evidence for the prefrontal-subthalamic hyperdirect pathway, where stopping-related potentials in the rIFG preceded the ones in the subthalamic nucleus (STN) (Chen et al., 2020).

Involvement of beta-band oscillations in response inhibition

Our results strongly support the interpretation that the rIFG signals the need to stop an already initiated action in the sense that it triggers a cascade of events that ultimately results in action cancelation and that beta-band oscillations are critically involved in this process. In general, beta-band oscillations are thought to be related to the maintenance of the current sensorimotor or cognitive state as hypothesized by a prominent model (Engel and Fries, 2010). In contrast, desynchronization of beta-band oscillations enables the transition into an active processing state (Miller et al., 2012; Little et al., 2019), for example, pressing a button. Our TFR results can be explained within the framework of this model, that is, beta-band desynchronization (BBD) is related to motor response preparation and BBD should already set in early, after the initial GO signal. And indeed, until approximately 100 ms after the STOP/AC-GO signal, we observe almost the same strength of BBD for both conditions (so that it cancels out in the contrast, Figure 3C). While in cAC-GO trials, the BBD continues until the median RT in the rIFG (Figure 3B, D), BBD is aborted in sSTOP trials already before the median SSRT, as we see an increase of beta-band power (Figure 3A, D). As soon as the stopping or motor response, respectively, is completed, beta-band oscillations resynchronize (Fonken et al., 2016).

To interpret the role of beta-band oscillations during the initiation of response inhibition, it is important to focus on the early changes in beta-band power, which set in before SSRT and the later so-called beta rebound. With the BBD signaling an ongoing motor response, a slowing or plateauing of BBD in the time period before SSRT could be interpreted as a slowing or pausing of this ongoing motor initiation process. Hence, the reduction of BBD during an ongoing motor initiation process could not be regarded as a distinct active stopping process. Yet, what we found in the rIFG was a rather early termination of BBD in sSTOP trials in the sense that beta-band power started to increase shortly after the STOP signal (Figure 3D). However, this increase in synchronization was not strong enough such that beta power was increased above baseline, but values reached baseline levels in rIFG. Thus, the relative change in beta power that we found in the contrast originates genuinely in the sSTOP trials, where BBD is terminated early, that is, before the SSRT, while in the cAC-GO trials no such termination was observed. This finding suggests an active change of processing in rIFG that is related to stopping, while processing proceeds ‘as usual’ in cAC-GO trials, and the motor response will be executed.

Furthermore, while we are fully aware that absence of evidence is not evidence of absence, the finding of significant Granger causality in time windows relevant for stopping, but not earlier, suggests that active signaling from rIFG to pre-SMA specifically in the beta band is an important part of the neural dynamics underlying stopping. In sum, the two findings, that is, rIFG as source of stopping relevant information transfer in the beta band and a termination of BBD specifically in rIFG already before SSRT, suggest that response inhibition is implemented in part by beta-band modulations. Although the present data and evidence from previous research (Wessel et al., 2013; Castiglione et al., 2019) indicate that the stopping process is related to a modulation of oscillatory power in the beta band, it still needs to be determined whether beta changes result from a pausing of action initiation or whether they have a genuine stopping function.

Controlling for attention in a selective stopping task suggests that rIFG does not serve an attentional role

A limitation of most previous neuroimaging studies on response inhibition is that sSTOP trials are usually contrasted with GO or uSTOP trials to identify inhibitory processes. Both contrasts are indeed capable of removing basic perceptual activity due the presentation of the stimuli (Aron and Poldrack, 2006; Sharp et al., 2010), but leave other confounding processes in place: Contrasting sSTOP with uSTOP trials, for example, involves error processing (Rubia et al., 2003) and an early response in motor cortices that has the potential to obscure activity related to the adjacent pre-SMA (Allen et al., 2018; Jerbi et al., 2007). Additionally, this contrast was criticized as overly conservative, because it assumes that inhibitory processes are also present in uSTOP trials (Boehler et al., 2010). Contrasting sSTOP with GO trials instead also introduces two issues: First, a less frequent event like the STOP signal is perceptually more salient and causes significant brain activation that is not directly based on inhibitory processing alone (Sharp et al., 2010). Second, STOP signals are delayed compared to the GO signals (i.e., SSD), hence directly contrasting both trial types entails the comparison of different time points within the stopping process (Figure 1A). This is particularly important when high temporal resolution methods are used. Additionally, attentional and discriminative processing of the STOP signal are presumed to disturb the timing of go and stop processes compared to GO trials (Sharp et al., 2010).

All of the above issues can be resolved by contrasting sSTOP with cAC-GO trials. First, the influence of error processing was reduced to a minimum when the contrast did not rely on unsuccessful trials (Rubia et al., 2003). Second, compared to the contrast with GO trials, frequency and stimulus properties of STOP and AC-GO signals could be matched precisely. As stimuli for STOP and AC-GO signals, blue and green arrows were used, respectively, and balanced across subjects. And indeed, the signal color had no effect on SSRT or with respect to RTs, on trial type (GO versus AC-GO). Additionally, trial number was balanced between both conditions, STOP and AC-GO. Hence, both trial types were presented with the same frequency, but still unexpected compared to GO trials. Thus, the stimuli of STOP and AC-GO could be regarded as equally salient. Third, AC-GO stimuli were presented after a signal delay that was adapted using a staircase procedure as for the STOP trials. Thus, timing of cAC-GO trials was matched to the one of sSTOP trials. Additionally, the attentional influence of the signal delay could be confirmed by our behavioral data. We found an increase in RTs after the appearance of an AC-GO signal compared to simple GO trials as predicted. There is independent evidence that the prolonged time required for the perceptual processing of the AC-GO signal is approximately the same as for the STOP signal compared to GO trials (Xu et al., 2017). In this study, the authors demonstrated at least the behavioral similarity of both conditions by comparing SSRT and ‘continue-signal reaction time’ (i.e., approximately, the time between AC-GO signal and response, a measure adopted from Mayse et al., 2014).

Hence, the contrast sSTOP versus cAC-GO is the most appropriate to identify the neural correlates of ‘pure’ response inhibition, independent of attentional and error processing (Sánchez-Carmona et al., 2016). The dissociation of inhibitory from attentional processes by an appropriate task design suggests that the rIFG does not, at least not primarily, serve an attentional role in response inhibition. However, although the rIFG has often been regarded as candidate for an inhibitory module and our results showed that response inhibition is indeed initiated by the rIFG, the rIFG might still be part of a more domain-general cognitive control mechanism, implemented by frontoparietal networks (Erika-Florence et al., 2014; Hampshire and Sharp, 2015a).

Limitations

One limitation of this study is that we could not analyze activity of the STN which is an important subcortical node within the response inhibition network because of its direct relation to IFG (Aron et al., 2007; Duann et al., 2009; Jahfari et al., 2011; Rae et al., 2015; Xu et al., 2017; Chen et al., 2020). No STN activity was detected by our source reconstruction contrast, most likely because MEG is less sensitive to subcortical compared to cortical sources (Gross, 2019). Nonetheless, several studies with intracranial recordings detected beta-band oscillations within the STN (Kühn et al., 2004; Ray et al., 2012; Zavala et al., 2018; Fischer et al., 2018; Chen et al., 2020), also supporting the importance of beta-band oscillations in response inhibition.

Conclusion

Our findings strongly support the hypothesis that response inhibition is initiated by the rIFG and implemented as a top-down control via beta-band oscillations (Aron et al., 2014). This brake is turned on in a time window around 100–200 ms after the STOP signal. Within the framework of the race model, rIFG may promote stopping processes and then signals to pre-SMA, among other regions, to proceed and execute the stopping process. Due to our task design, that is, contrasting sSTOP with cAC-GO trials, we could exclude a primarily attentional role of the rIFG in stopping. Thus, these results may open up new ways of spatially localized oscillation-based interventions, that is, modulating beta-band activity by real-time EEG-triggered TMS. Additional causal evidence for the mechanisms of response inhibition may open up new avenues of novel therapeutic strategies for diseases related to impaired response inhibition.

Materials and methods

Participants

Sixty-two healthy subjects participated in the MEG experiment. As a result of applying behavioral outlier criteria (Congdon et al., 2012), one participant had to be excluded because the inhibition stop rate was below 40%. A second participant had to be excluded because the mean RT of uSTOP trials was higher than the mean RT of cGO trials, indicating that the assumptions underlying the horse race model (Logan et al., 1984) were violated in this participant. A third participant had to be excluded because no head model could be obtained from the available anatomical MR image. All of the remaining 59 participants (39 women; mean age ± SD, 25 ± 6 years) had normal or corrected-to-normal vision. All individual participants included in the study were screened for factors contradicting MRI and MEG scanning and provided written informed consent before participation and consent to publish any research findings based on their provided data in anonymized form. The study was approved by the local ethics committees (Johann Wolfgang Goethe University, Frankfurt, Germany, and Medical Board of Rhineland-Palatinate, Mainz, Germany), and participants were financially compensated for their time.

Behavioral

Request a detailed protocol

SSRT was calculated using the integration method (Verbruggen et al., 2019). This method estimates the SSRT by finding the point at which the integral over the RT distribution equals the actual p(respond|signal) and then subtracting the mean SSD. The integration approximately corresponds to the nth RT of the distribution of GO trials (sorted by RT), multiplied by the actual p(respond|signal). For instance, p(respond|signal) is 0.48 across 1000 GO trials acquired, then the nth RT is the 480th fastest GO RT. Then SSRT is calculated by subtracting the mean SSD from the 480th RT. The distribution of GO trials also includes incorrect GO trials; GO omissions were replaced with the maximum RT in order to compensate for the lacking response (Verbruggen et al., 2019). Behavioral data was extracted from the MEG triggers and analyzed with the JASP software (version 0.12.2, https://jasp-stats.org).

Tasks and procedure

Request a detailed protocol

Participants performed a stimulus selective stopping task with an additional attentional control condition, that is, AC-GO trials. The task and procedure was the same as described in Sebastian et al., 2016. For stimulus presentation, we used Presentation software (version 13.1, http://www.neurobs.com).

Before the acquisition session, participants had to read the instructions and practice the task on a laptop computer for approximately 5 min. The acquisition session was split into 10 blocks. Throughout the acquisition, participants were asked to respond to the stimuli by pressing a response button with the left or right thumb (MEG) or index finger (fMRI), respectively.

The task comprised three conditions: a GO condition (50% of all trials), a STOP condition (25% of all trials), and an AC-GO condition (25% of all trials). Trials of all conditions began with the presentation of a white fixation cross in the center of the screen with a randomly varied duration between 1500 and 2000 ms in the MEG experiment and 500 ms in the fMRI experiment, respectively (Sebastian et al., 2016). Then, a white arrow (GO signal) pointing to the left or right was displayed. Participants were instructed to respond with a left button press to a left pointing arrow and with a right button press for a right pointing arrow. In the GO condition, the white arrow was displayed for 1000 ms (equivalent to the maximum permitted RT) or until a button press was performed. In the STOP condition, the white arrow was presented first, followed by a change of its color from white to blue after a variable SSD. Participants were instructed to try to inhibit their initiated button response after the GO signal. The SSD was adapted to the participants’ performance to yield a probability of 50% of successful inhibitions per block. Therefore, a staircase procedure was implemented with the following properties: The initial SSD was set to 210 ms. If the response was not successfully inhibited (unsuccessful stop trial, uSTOP), the SSD in the next STOP trial was decreased by 30 ms with a minimum SSD of 40 ms. If a response was successfully inhibited (successful stop trial, sSTOP), the SSD in the next STOP trial was increased by 30 ms. The maximum SSD was limited by the maximum permitted RT of 1000 ms. In the AC-GO condition, the white arrow was presented first, followed by a change of its color from white to green after a variable AC-GO signal delay (ACSD), and participants were instructed to continue their response. The ACSD was varied in accordance with the staircase in the STOP condition.

The attribution of color (green/blue) to trial type (STOP/AC-GO) was counterbalanced across participants. In case of an omission error (no button press) in the GO or AC-GO condition, participants were given a short feedback (‘oops’ – no button press’) to maintain the participants’ attention and to limit proactive slowing. The length of the intertrial interval (presentation of a blank screen) was 700 ms in the MEG experiment. In the fMRI experiment, the length of the intertrial interval was varied randomly between 2500 and 3500 ms. One block consisted of 112 trials presented in a randomized order.

fMRI experiment

Request a detailed protocol

A subset of 31 participants of the MEG experiment performed the fMRI task and additionally data from 45 participants from the Sebastian et al., 2017 study were included for analysis (n = 76). For all participants of the MEG experiment (n = 62), an anatomical MRI was acquired on a Magnetom Trio Syngo 3 T system (Siemens Medical Solutions) at two sites (8-channel head coil at site 1 and a 32-channel head coil at site 2). For recording details and analyses, refer to Sebastian et al., 2017.

MEG experiment

MEG data acquisition and preprocessing

Request a detailed protocol

MEG data acquisition was conducted in line with the good practice guidelines for MEG recordings recommended by Gross et al., 2013. MEG signals were recorded using a whole-head system (Omega 2005; VSM MedTech) with 275 channels. The signals were recorded at a sampling rate of 1200 Hz in a third-order synthetic gradiometer configuration and were filtered with fourth-order Butterworth 300 Hz low-pass and 0.1 Hz high-pass filters. In addition to the MEG, we recorded the electrooculogram (EOG) and electrocardigram (ECG) for offline artifact rejection. All analyses were performed in MATLAB (MathWorks Inc, Natick, MA), mainly using the FieldTrip toolbox (Oostenveld et al., 2011) and custom scripts (see Box 1).

Trial definition and selective filtering was directed by our aim to identify neural networks related to response inhibition. To eliminate attentional processes, we intended to directly contrast sSTOP with cAC-GO trials. However, we had to take into account that cAC-GO trials might in turn involve button press-related motor activity that is not present in the sSTOP trials. Therefore, the definition of a tROI should preferably end before button presses in cAC-GO trials set in. However, regarding sSTOP trials, the tROI should not end before response inhibition is initiated and finished. To also account for participants with longer SSRTs compared to the average SSRT, we additionally sought to ensure that for these participants the inhibition response process is still captured by the tROI, so we a priori defined a tROI from 100 ms to an SSRTmax of 350 ms relative to the STOP signal. To reduce the contamination of our tROI with button press-related motor activity, we rejected cAC-GO trials with an RTAC-GO < SSRT, where RTAC-GO is the ‘attentional capture GO RT’, defined as duration between AC-GO signal and button press, and SSRT is the median SSRT across subjects (244 ms). More importantly, when contrasting sSTOP and cAC-GO trials, these trials should also be matched in terms of the underlying distribution of go and stop processes according to the independent horse race model. This is important because the neural correlates of response inhibition revealed by this contrast should not rely on mere timing differences between the underlying go and stop processes (Sánchez-Carmona et al., 2016). Although the distribution of go and stop processes cannot be directly accessed, the horse race model implies that in sSTOP trials, the response can be successfully inhibited because the go process is slower than the stop process. Thus, to reduce processing speed differences to a minimum, it is necessary to select cAC-GO trials with longer than average RTs, where the underlying distribution consists of slower go processes, comparable with sSTOP trials. Thus, rejecting cAC-GO trials with an RTAC-GO < SSRT additionally supports a close match between the underlying distribution of go and stop processes in both trial types. After this trial selection was performed, the 10% percentile of RTAC-GO was 273 ms, and the median was 359 ms (as indicated in Figure 3). To exclude that these results were biased by the cAC-GO trial selection (i.e., only including trials with RTAC-GO > SSRT), we performed all analyses on randomly selected cAC-GO trials. Overall, these results were comparable to the reported ones.

Trials containing sensor jump or muscle artifacts were rejected using automatic artifact detection functions provided by FieldTrip. Power line noise was reduced using a discrete Fourier transform filter at 50, 100, and 150 Hz. In addition, to remove artifacts related to eye movements and heart beat, ICA was performed on all trial types (Makeig et al., 1995). Data was downsampled to 400 Hz before the extended infomax (runica/binica) algorithm was applied to decompose the data (as provided by FieldTrip/EEGLAB). Independent components (ICs) were removed from the data if their spatial topography corresponded with the artifact type (Fatima et al., 2013), which was visually inspected and, when in doubt, correlation coefficients with EOG and ECG data were consulted. Typically, five ICs were rejected (range 2–8 ICs). Since minimization of head movement is crucial to MEG data quality (Gross et al., 2013), trials were rejected when the head position deviated more than 5 mm from the mean head position over all blocks for each participant.

Spectral analysis at sensor level

Request a detailed protocol

The only statistics performed at the sensor level was to define appropriate frequency bands as beamformer parameters for subsequent source reconstruction. We therefore compared the spectral power between 4 and 120 Hz (based on Hanning tapers) over averaged channels during the tROI (100–350 ms) with the spectral power of corresponding baseline segments for all participants. For this purpose we used a dependent-samples permutation t-test and a cluster-based correction method (Maris and Oostenveld, 2007) to account for multiple comparisons across frequencies. Samples whose t-values exceeded a threshold of αcluster = 0.01 were considered as candidate members of clusters of spectrally adjacent samples. The sum of t-values within every cluster, that is, the ‘cluster size’, was calculated as test statistics. These cluster sizes were then tested (two-sided) against the distribution of cluster sizes obtained for 5000 partitions with randomly assigned task and baseline data within each subject. Absolute cluster values above the 99.5th percentiles of the distribution of cluster sizes obtained for the permuted datasets were considered significant. This corresponds with an alpha threshold of α=0.01 that was Bonferroni-corrected for a two-tailed test. Since we aimed to contrast neural power of sSTOP and cAC-GO trials on source level to obtain a set of sources related to response inhibition, we had to use an (orthogonal) statistical test at sensor level in order to avoid circularity (Kriegeskorte et al., 2009). We therefore combined both conditions (sSTOP and cAC-GO trials) for analysis on the sensor level and contrasted these ‘pooled’ task segments with corresponding baseline segments of both conditions.

Source imaging

Request a detailed protocol

The anatomical MRI of each participant was linearly transformed to a segmented standard T1 template of the SPM8 toolbox (http://www.fil.ion.ucl.ac.uk/spm/) in MNI space (Collins et al., 1994). This template was overlaid with a regular dipole grid (spacing 1 cm). The inverse of the obtained linear transformation was then applied to this dipole grid and the lead field matrix was computed for each of the grid points of the warped grid using a single shell volume conductor model (Nolte, 2003). Since all grid locations of each subject were aligned to the same anatomical brain compartments of the template, corresponding brain locations could be statistically compared over all subjects.

To reconstruct sources in the beta and gamma band, a frequency domain beamformer source analysis was performed by using the dynamic imaging of coherent sources algorithm (Gross et al., 2001) implemented in the FieldTrip toolbox. Since our experimental setup did not contain any (coherent) external reference, filter coefficients were constrained to be real-valued to restrict our analysis to local source power. This can be done, because we assume that the magnetic fields propagate instantaneously from source to sensor, and therefore, no phase shifts can occur that would lead to complex coefficients (Nunez and Srinivasan, 2006). Beamformer analysis uses an adaptive spatial filter to estimate the power at every specific (grid) location of the brain. The spatial filter was constructed from the individual lead fields and the cross-spectral density (CSD) matrix for each subject. CSD matrices were computed for the task period of 100–373 ms after STOP/AC-GO signal onset and a baseline period of same length, with an offset of −100 ms relative to GO signal onset (see Figure 1B for timescale). To avoid spectral leakage, the length of the time-frequency window was extended from 350 ms so that it matches an integer number of oscillatory cycles of the center frequency of the frequency band width (Harris, 1978). Frequency bands were defined based on the statistical analysis of spectral power at the sensor level. Thus, CSD matrices were computed in the beta band for 22 Hz (±10 Hz) and the gamma band for 76 Hz (±12 Hz), where spectral smoothing is indicated in brackets. CSD matrix calculation was performed using the FieldTrip toolbox with the multitaper method (Percival and Walden, 1993) using four Slepian tapers (Slepian, 1978) for the beta band and five Slepian tapers for the gamma band.

Source statistics

Request a detailed protocol

To reveal sources related to response inhibition at group level, we statistically tested differences between source power of sSTOP and cAC-GO trials. First, we calculated an activation-versus-baseline t-statistic at single-subject level by using an analytic dependent-samples within-trial t-test. The t-values obtained from this activation-versus-baseline test were then used as input for a cluster-based permutation test for contrasting (‘baseline corrected’) sSTOP and cAC-GO trials (dependent-samples t-test, Monte Carlo estimate, same two-sided test as described for the spectral analysis at sensor level, but for clusters of adjacent voxels as samples and α=0.05). In the clusters obtained, we searched for local extrema to identify peak voxels. All of the peak voxels obtained were reported, but for subsequent TFR analysis, only physiologically plausible voxels, that is, voxels in gray matter according to the Jülich Histological atlas provided by the FSL toolbox (FMRIB’s Software Library) (Smith et al., 2004), were used.

Latency analysis

Request a detailed protocol
TFRs on virtual channel time courses
Request a detailed protocol

To analyze the temporal activation pattern of each peak voxel identified by contrasting sSTOP and cAC-GO trials, we computed TFRs of these sources. Therefore, time courses of these sources were reconstructed as ‘virtual channels’ using bandpass-filtered raw data (2–150 Hz). For time course reconstruction, a time-domain beamformer was used (linear constrained minimum variance) (Van Veen et al., 1997), based on common filters analog to the ones used in the frequency domain. In contrast to the frequency domain, uSTOP trials were additionally included in the common filter computation when a post hoc analysis with uSTOP trials was performed. A principal component analysis on the three reconstructed time courses in x, y, and z direction was performed for each grid point in order to determine the dominant dipole orientation (direction with the largest variance). The time course of the first principal component was used for subsequent TFR analysis (8–44 Hz, Hanning taper with a sliding time window of three cycles were used). We computed three different TFRs for each source: (1) task-versus-baseline power for sSTOP trials, (2) task-versus-baseline power for cAC-GO trials, and (3) the contrast of sSTOP versus cAC-GO trials to reveal significant inhibition-related power changes at group level. In case of the latter TFR, for baseline correction, we z-transformed sSTOP and cAC-GO trials of each participant separately by subtracting the mean of the corresponding baseline and dividing by the standard deviation of combined (sSTOP and cAC-GO) baseline trials. For all of the three TFRs, we used a cluster-based permutation test as described for the spectral analysis at sensor level, except for (time,frequency)-pairs serving as samples and α=0.05. However, since cluster-based permutation tests on TFRs do not establish significance of latency differences between conditions (Sassenhagen and Draschkow, 2019), we used individually defined onset values to test for a significant latency difference between rIFG and pre-SMA. These individual onset latency (and power) values were defined by subtracting averaged z-transformed cAC-GO trials from averaged z-transformed sSTOP trials (see below).

TFR-based latency analysis
Request a detailed protocol

To answer the question of the temporal order in which the two candidate sources, rIFG and pre-SMA, are activated, we performed an onset latency analysis based on difference TFRs (z(sSTOP) – z(cAC-GO)). Source power was averaged over the beta band (12–32 Hz), smoothed and analyzed for peaks within the tROI. Onset latency (for each participant and source) was defined as the time point at which the averaged source power exceeded a threshold of 25% of the range between the first positive peak within the tROI and the (positive) minimum found until this peak (in most cases, zero). If there was no positive peak within the tROI for at least one source, we excluded this subject for the TFR-based latency analysis. To test for differences between onset latencies of rIFG and pre-SMA at group level, a permutation statistics was performed (50,000 permutations). Onset latency differences were also tested with thresholds of 10%, 30%, 50%, 75%, and 100%.

Alternative latency analysis using time-resolved SVM
Request a detailed protocol

To determine the onset of discriminability in the time-domain between sSTOP and cAC-GO trials, we used a time-resolved multivariate support vector machine analysis (t-SVM) with a linear kernel. The t-SVM was trained and tested separately for each subject, and for rIFG and pre-SMA. First, we z-normalized each source signal relative to baseline period (samples before the GO cue, 500 ms length). Next, to reduce the computational cost and increase the signal-to-noise ratio, the data were smoothed with a Gaussian Kernel of ±10 ms and downsampled to 300 Hz, similar to the approach in Hebart et al., 2018. For time embedding, we used 14 consecutive samples, corresponding to one beta cycle (center frequency 22 Hz). After preprocessing, the embedded source data were randomly assigned to one of eight supertrials (per condition) and averaged (MATLAB code was adopted from Guggenmos et al., 2018, which builds on the library for SVMs by Chang and Lin, 2011). Finally, we separated these supertrials into training and testing data with one supertrial per condition serving as test data and all the other as training data. The binary classification was performed for each time point from −100 ms before to 700 ms after signal onset. To obtain a more robust estimate of the classification accuracy, we performed 100 iterations of supertrials averaging and classification. The final classification time series reflects the average across these iterations. For statistical testing, we employed a nonparametric cluster permutation approach (Maris and Oostenveld, 2007), with clustering of subsequent time points (10,000 permutations). The null hypothesis of no experimental effect for the classification time series was equal to 50% chance level and tested within the tROI (100–350 ms, where 350 ms corresponds to SSRTmax). Using a bootstrap approach (1000 iterations), 95% confidence intervals of the mean were estimated. For each subject we computed thresholded onset and peak latencies of the classification time course. For this, we adopted the method from Marti et al., 2015. First, the data were low-pass filtered at 10 Hz. Second, to measure the latency of the peak, we identified the first peak within the tROI that was above one or two standard deviations, respectively, of the mean baseline accuracy. If no peak was found, we considered all time points within the tROI that exceeded the 95th percentile of the distribution of the classification performance. Then the median of these points was considered as the peak latency. Finally, from the peak latency, the onset of the peak was defined by going backward and identifying the time point at which the classification performance exceeded a certain threshold percentage of the peak. We performed this analysis with different percentage values (10%, 25%, 30%, 50%) of the difference between the mean decoding accuracy during baseline and the accuracy peak value (results were similar across different threshold values). If the peak latency criteria could not be met for one source, rIFG or pre-SMA, the subject was excluded. Statistical differences between onset latencies of rIFG and pre-SMA at group level were assessed as in the TFR-based latency analysis with permutation statistics (50,000 permutations). To check if the effects we found using broadband data as input for the SVM were specific to beta-band oscillations, we repeated the above described analyses on beta-band filtered data. The source signal was filtered between 12 and 32 Hz using a one-pass Butterworth filter.

Bayesian multiple regression for SSRT analysis

Request a detailed protocol

To analyze how SSRT is influenced by changes in beta-band power or connectivity between rIFG and pre-SAM, we performed a Bayesian multiple regression with per-subject SSRT values as outcome variable and beta power (sSTOP–cAC-GO) in pre-SMA and rIFG in successful and failed trials, as well as the DAI (see below) as regressors. All participants that showed a positive beta power peak during the tROI (as described for the onset definition based on TFRs) in both trial types and both regions were included into the analysis (outliers above three standard deviations were excluded, n = 43).

We chose to model the SSRT as normally distributed across subjects, as it is (despite its name) not a directly observable RT but the result of an extraction procedure that involves multiple samples per SSRT value. We z-normalized all variables before modeling and used a normal distribution with zero mean and a standard deviation of 0.5 as the prior for each regressor. The likelihood was modeled as a normal distribution centered on the regression model with a standard deviation drawn from a half normal prior with standard deviation of 1. Posterior distributions of the parameters were computed in PyMC3 (Salvatier et al., 2016) using the Hamiltonian Monte Carlo No U-Turn sampler (NUTS), with 2000 tuning samples and 20,000 samples for the joint posterior distribution.

We compared three models for their expected posterior log-probability using an LOO CV procedure as suggested in Gelman et al., 2014. The three models are described in the 'Results' section.

Connectivity

Request a detailed protocol
Nonparametric Granger causality
Request a detailed protocol

For the computation of cGC, we employed a multivariate nonparametric spectral matrix factorization. We computed the CSD matrix of the source signals on the tROI (100–350 ms) using the fast Fourier transform in combination with multitapers (5 Hz smoothing). Using the nonparametric variant of cGC (Dhamala et al., 2008) avoids choosing a multivariate autoregressive model order, which can introduce a bias. Specifically, we used a blockwise approach (Wang et al., 2007), considering the first two PCs of each source signal as a block, and we estimated the cGC that a source X exerts over a source Y conditional on the remaining areas (Bastos et al., 2015). According to our hypothesis, testing for differences between both trial conditions was done for two links, from rIFG to pre-SMA, and vice versa, conditional on all other five active cortical sources as obtained by the source reconstruction. Even though unobserved sources pose an irresolvable problem (Bastos and Schoffelen, 2015), and we cannot totally exclude this scenario, we made use of all the information from the sources observed to distinguish between direct and indirect effects and thus avoiding possible spurious results.

Statistical testing of cGC
Request a detailed protocol

First, we assessed whether the average cGC (in the frequency range of 8–44 Hz) of the source-target pairs (rIFG to pre-SMA and vice versa) was reliably above the bias level, for each condition (sSTOP and cAC-GO) separately. In order to estimate the bias, we randomly permuted the trials 1000 times in each condition to create a surrogate distribution of mean cGC values. We tested if the found cGC value was in the upper 97.5% extreme (equivalent p<0.05 with a Bonferroni correction for two possible source-target pairs) of the surrogates distribution. If the average cGC exceeded the bias level, this source-target link was considered significant. These steps were repeated for each subject separately. Second, for both source-target pairs, cGC values in the sSTOP condition were contrasted with cGC value in the cAC-GO condition at the group level on the subjects that showed at least a significant link in one of the two conditions. The statistical comparison was performed in the range 8–44 Hz using a dependent-samples permutation t-metric. A cluster-based correction was used to account for multiple comparisons across frequencies (Maris and Oostenveld, 2007). Adjacent frequency samples with uncorrected p-values of 0.05 were considered as clusters. Fifty-thousand permutations were performed and the critical α value was set at 0.025. A Bonferroni correction was applied to account for multiple comparisons across links.

Directed influence asymmetry index
Request a detailed protocol

The DAI was defined by Bastos et al., 2015 as

DAI=GC(AB)GC(BA)GC(AB)+GC(BA)

where in our analysis brain area A was rIFG and brain area B was pre-SMA. The numerator represents the net direction of GC (or cGC, respectively) while the denominator normalizes the net GC by the sum of GC in both directions such that it is possible to compare it across frequencies (Michalareas et al., 2016).

Data availability

All data generated or analysed during this study are included in the manuscript and supporting files. All source data files are available on Dryad Digital repository (https://doi.org/10.5061/dryad.x3ffbg7gp). All custom Matlab codes used in these analyses are available at https://github.com/meglab/acSST (copy archived at https://archive.softwareheritage.org/swh:1:rev:ea0bf4acc0f11cdc78ad31b6c1285f1851389312).

The following data sets were generated
    1. Schaum M
    (2021) Dryad Digital Repository
    Right inferior frontal gyrus implements motor inhibitory control via beta-band oscillations in humans.
    https://doi.org/10.5061/dryad.x3ffbg7gp

References

    1. Chang C-C
    2. Lin C-J
    (2011) LIBSVM: a library for support vector machines
    ACM Transactions on Intelligent Systems and Technology 2:1961199.
    https://doi.org/10.1145/1961189.1961199
  1. Book
    1. Efron B
    2. Tibshirani R
    (1993)
    An Introduction to the Bootstrap
    Chapman & Hall/CRC Press.
  2. Book
    1. Gelman A
    2. Carlin JB
    3. Stern HS
    4. Dunson DB
    5. Vehtari A
    6. Rubin DB
    (2014) Evaluating, comparing, and expanding models
    In: Gelman A, editors. A Simple Method for Comparing Complex Models: Bayesian Model Comparison for Hierarchical Multinomial Processing Tree Models Using Warp-III Bridge Sampling. Chapman & Hall/CRC Press. pp. 165–196.
    https://doi.org/10.1007/s11336-018-9648-3
  3. Conference
    1. Makeig S
    2. Bell AJ
    3. Jung T-P
    4. Sejnowski TJ
    (1995)
    Independent component analysis of electroencephalographic data
    In NIPS’95 Proceedings of the 8th International Conference on Neural Information Processing Systems. pp. 145–151.

Decision letter

  1. Simon Little
    Reviewing Editor; UCSF, United States
  2. Richard B Ivry
    Senior Editor; University of California, Berkeley, United States
  3. Simon Little
    Reviewer; UCSF, United States

In the interests of transparency, eLife publishes the most substantive revision requests and the accompanying author responses.

Acceptance summary:

This manuscript is a valuable addition towards understanding the neural mechanisms of inhibitory control. The combination of MEG with fMRI in a large cohort of subjects using a behavioral paradigm that takes account of attentional confounds, provides a rich dataset in which to determine the temporally primary cortical node in action stopping – the pre-supplementary motor area (pSMA) or right inferior frontal gyrus (rIFG). Here it is shown using a number of techniques, that β activity in the rIFG begins prior and is granger causal to the pSMA, and predicts stopping behavior. As detailed in the discussion, this significantly improves our understanding of this network and provides a platform for consideration of future therapies directed towards rIFG inhibitory β.

Decision letter after peer review:

Thank you for submitting your article "Right inferior frontal gyrus implements motor inhibitory control via β-band oscillations in humans" for consideration by eLife. Your article has been reviewed by four peer reviewers, including Simon Little as the Reviewing Editor and Reviewer #1, and the evaluation has been overseen by Richard Ivry as the Senior Editor.

The reviewers have discussed the reviews with one another and the Reviewing Editor has drafted this decision to help you prepare a revised submission.

As the editors have judged that your manuscript is of interest, but as described below, see the need for additional analyses and revisions are required before it could be considered for publication. We would like to draw your attention to changes in our revision policy that we have made in response to COVID-19 (https://elifesciences.org/articles/57162). First, because many researchers have temporarily lost access to the labs, we will give authors as much time as they need to submit revised manuscripts. We are also offering, if you choose, to post the manuscript to bioRxiv (if it is not already there) along with this decision letter and a formal designation that the manuscript is "in revision at eLife". Please let us know if you would like to pursue this option. (If your work is more suitable for medRxiv, you will need to post the preprint yourself, as the mechanisms for us to do so are still in development.)

This study uses a multimodal neuroimaging approach to investigate the cortical inhibitory control network in a significant number of subjects with a modified version of the stop signal reaction time task. The authors examine the process of stopping a movement with the aim to show that this process is initiated in the right IFG rather than pre-SMA. To this end, they perform MEG source localization analysis of response latencies in the source data and directionality analysis using a variant of Granger causality. The central question that is addressed is whether the IFG is the hub of inhibitory control, or does response inhibition initiate within the SMA. Previous studies have been inconclusive. Some have been limited by poor temporal resolution (fMRI), while recent MEG studies suggest both are active. Another important question and unresolved issue regarding stop inhibition is how to disentangle attentional from inhibitory processes. To address this, this study utilizes attentional capture go trials and compares activity between those trials and stop trials. This allows to partially distinguish between an attentional signal and an inhibitory signal.

The main findings was of a significantly earlier β band response latency in rIFG than pre-SMA and that rIFG β but not pre-SMA β is correlated with stop signal response time. Additional analyses showed a significant correlation for a modified Granger causality from rIFG to pre-SMA but not vice versa and a partially overlapping activation network in the same task using fMRI compared to MEG.

Although there was generally some enthusiasm for the manuscript, writing and the analysis, there was also convergence on a number of concerns across reviewers. Most specifically, it was felt that the main finding, that rIFG precedes pre-SMA needs further evidence to support it. The other findings of rIFG and pre-SMA involvement in stopping support previous findings. The main reviewers concerns are summarized here below and particular attention should be made to addressing these:

The question was repeatedly raised of the use of contrasts in the β power analysis (rather than showing the non-contrasted β amplitudes). Specifically, it is not possible to know if there was a genuine increase in β related to stopping, or simply a slowing of β event related desynchronization (reduction) in the stopping condition compared to the attentional capture condition. This would have important consequences for mechanistic interpretations of the signal and whether it can be truly considered as an initiator of the stopping network, but wasn't clear from the results or figures.

The appearance of low frequency activity in the spectrogram in Figure 3 and the question of whether the IFG – SMA inhibitory effect is also mediated by non – oscillatory evoked activity (which also appears in the low frequency range). This would require a further analysis looking to determine if the effects were truly specific to β or could also be found in non – oscillatory (evoked) activity. Furthermore, multiple reviewers felt that a better mechanistic account was needed of how such short timing effects (e.g. 20 ms difference in the SVM model classification difference) could be driven by β with a cycle length of 50ms and much clearer explanation of why the β frequency ranges were so different across analyses. Notably – the 30-40 Hz signal in the Granger causality analysis does not appear to be the same signal being analyzed in the latency analysis. This either requires further analysis of matching frequencies or significant further explanation and more measured interpretation.

Correspondence between fMRI versus MEG was not felt to be as strong as described in the write up. This raised the question for some reviewers as to whether the regions of interest where truly chosen in a data driven way. A further control analysis of another region such as pre-motor cortex and left IFG was felt to be necessary to ensure that these effects seen are genuinely restricted and specific to the rIFG and pre-SMA in the time period of interest.

Elaboration on the above issues:

1. The author use a comparison of statistics rather than a direct statistical comparison of correlations. This comparison was felt to be invalid by a number of reviewers and a direct statistical comparison was requested (see below for details). Statistics – The argument for the difference between rIFG and pre-SMA (Figure 3D) is based on comparison of statistics rather than a statistic for comparison (see https://www.nature.com/articles/nn.2886). Even ignoring that, the p-value of 0.042 with N=50 does not seem to be very convincing and looking at the figure the two point clouds are not very different. It is unclear how the results will stand up if this essential comparison is done..

2. Statistic and corrections – Correlation analysis: For demonstrating the difference between rIFG and preSMA correlations these must be directly compared to each other rather than showing a significant correlation in one, and the absence of that in the other (please see Nieuwenhuis et al. 2011, Nature Neuroscience). Also, though not relevant when comparing preSMA and rIFG directly, when using multiple correlations these should be corrected for multiple comparisons. As we understand the analyses, four correlation analyses were performed preSMA and IFG for different contrasts (uSTOP and sSTOP). The rIFG correlation seems to only survive when not correcting for multiple comparisons and using a one-tailed p.

3. Contrasts – "he authors almost exclusively report relative changes in MEG activity in one condition (successful stop, sSTOP) compared to another condition (attentional capture GO, cAC-GO). Even though the authors try to take into account that in one condition people pressed a button (cAC-GO) while they do not execute a motor response in the other (sSTOP) by specifying a certain time window, they still cannot control for differences related to the well-known β-decrease during a motor response, which starts several hundred ms before a response. Looking at the median RT, there is presumably strong β-desynchronisation during the time window of interest. So any 'β-increase' in sSTOP vs cAC-GO, which is interpreted as stopping-related, might simply reflect that there is less β decrease in sSTOP compared to cAC-GO, because there is not motor response. In fact, supplementary figure 2 shows that β strongly decreased (when pooled across conditions), and not increased, after the cue / before the response. In order to support the claim that a β increase (and not the lack of a decrease) is related to stopping, the authors would have to show that β in fact increases in sSTOP-GO compared to baseline activity (e.g. before the cue).

4. Contrasts – "Figure 3 and Methods L. 1248 – I was not able to follow the authors' description of the difference between what is shown in 3A and 3B other than that 3A shows t-value and 3B – z-value. When you talk about z-transforming trials, did you z-transform the raw data or the TF power? If it is the former, then the colour in 3B would not correspond to z-value. What makes sense to me is that in both cases you averaged the power across trials for each subject but in 3A looked at differences in power whereas in 3B – at differences in latency but either I am wrong or the description is unclear."

5. fMRI – "Description of Figure 2B in the text implies that there was specific activation of rIFG and preSMA in the fmri study, but the figure shows broad activation including bilateral prefrontal, parietal and occipital areas. Thus, even though the authors state that they did the source reconstruction for analyzing sources based on the current experiment and not based on previous literature picking out preSMA and rIFG for the next analyses seems to be largely related on previous literature and not the current study. For example, how would these analyses look like for left IFG or premotor areas? While I understand the rationale for looking at rIFG more closely based on the current data (due to the early onset of spectral differences), picking out preSMA from the other sources seems to be based on previous literature."

6. fMRI – "In Figure 2 the correspondence between MEG and fMRI results does not seem to be very convincing with the MEG results surprisingly being much clearer. Particularly the absence of activation of bilateral motor strips and pre-SMA in fMRI is puzzling.

a. Would it be possible to quantify the overlap and say how it would compare to other published MEG/fMRI comparisons with the same task?

b. Do you have any explanation for the discrepancies?

c. Also for me it is not obvious that higher β power should necessarily correspond to higher BOLD as at least for movement higher β is associated with movement inhibition. Have you tried looking at the other direction of the fMRI contrast, perhaps the missing sensorimotor cortex is there?

d. Is the IFG latency also significantly different than the latency in the other brain regions in which β band activity is observed?"

7. SVM – "The β latency analysis is supported by the SVM analysis which uses a wide-band frequency input. To strengthen the case that the β band is mechanistically primary here, it would be useful to repeat the SVM analysis using a β filtered input to the SVM. This would confirm that the effects seen in the SVM are driven by the β band.

8. SVM – "Just as a point of clarification, it appears that the onset analysis of the classification reveals a much later time of onset. It would be helpful to understand why this classification time course appears slower than the time course of β power, and whether the difference in classification between IFG and SMA should be interpreted in the same way as the difference in power latency."

9. SVM – The authors report that the SVM was point-by-point. One reviewer commented "I don't understand what that means because the whole idea of SVM is that it's multivariate so could possibly be sensitive to things like β. Point-by-point would mean that somehow the raw signal amplitude at each point is different between trial types so unless there is an evoked response there, I don't see how this would be the case. If it's really point-by-point the latency difference with β power would make sense because the time resolution of TF analysis is limited, in this case to 150ms windows around 20Hz."

10. Frequency of interest – "There is an early 'relative' β-increase in rIFG, which already starts ~100ms after the cue. Its spectral components seem quite different from the later occurring β change and it includes a low-frequency component. Did the authors assess whether this could be an evoked response with different frequency components in the spectral domain? The cluster's duration seems <100 ms which would correspond to less than 2 cycles of a 20Hz oscillation. Also, would this cluster be significant when not 'spilling over' into the later β-cluster and the low-frequency cluster?"

11. Frequency of interest – "Recent evidence, already cited in this manuscript (Chen et al. 2020) suggests that an earlier evoked response from IFG might precede β changes during reactive stopping. The Figure 3A low frequency activity also supports this. Given the SVM changes on broadband data, the findings in Figure 3A and the previous literature, it would be informative if the authors could analyze if evoked (non – oscillatory +- very low frequency δ which is likely equivalent) activity is also earlier (and potentially) causal in IFG --> pre-SMA."

12. Frequency of interest – "Much of the discussion and the initial results regarding spectral power are focused on the broad β band from 12-30 Hz. Yet the analysis of Granger causality and DAI are focused on a slightly higher frequency range, 30-40 Hz, which is not usually considered as β when discussing the stopping nature of IFG (in fact, some would consider this in the range of low γ). How should one reconcile these different frequencies?

a. Can the authors re-analyze in matched frequency bands and discuss the non – concordance in frequency given.

b. Moreover, the relation between DAI and SSRT is specific to 34 Hz. I understand that this has been corrected for multiple comparisons, but does this correction just account for the six discrete frequencies between 30-40 Hz, or does this account for the entire β band (or even all frequencies for that matter)? "

13. Frequency of interest – "This relates to interpretation – could the authors discuss and reference how an oscillation with a cycle length ~ 50ms could causally impact one region in the very short time frames demonstrated in this study (e.g SVM difference is 20ms in Figure 3C, ie less than half a cycle)."

14. Granger causality – "Figure 4 – why does Granger causality go down with frequency? I would expect it to be flat except where there is a significant relation between signals. Could it be some kind of power confound or normalisation issue?"

15. DAI – "L. 475 – Was the correlation of DAI with SSRT independent of the IFG power correlation i.e. was the DAI regressor significant in the presence of rIFG power? If not, perhaps the DAI result could be explained by varying rIFG SNR that is known to confound directional measures."

16. Clusters – "L. 1152 – the description of the cluster test seems to be inaccurate. If you took cluster sizes below 0.5th percentile of the distribution that would mean that you took very small clusters as well as very large. What usually happens is that people take large clusters for both positive and negative direction and apply Bonferroni correction for the two directions. Also why was a conservative threshold used p<0.01? Would there be additional clusters for p<0.05?"

17. Code – "Although the authors included links to the data and the custom code in the cover letter, they do not appear in the paper itself. It would be helpful to include a supplementary table pointing to the toolboxes and custom functions used for each analysis. I was wondering particularly about cGC as no reference for the implementation was given and the conditional part sounds like more than is implemented in Fieldtrip."

[Editors' note: further revisions were suggested prior to acceptance, as described below.]

Thank you for resubmitting your work entitled "Right inferior frontal gyrus implements motor inhibitory control via β-band oscillations in humans" for further consideration by eLife. Your revised article has been evaluated by Richard Ivry (Senior Editor) and a Reviewing Editor.

We appreciate your resubmission along with your thorough responses to the suggestions from our reviewers. Many of the issues have been addressed eg. comparison of statistics with your new modelling approach, clarification around ROIs and fMRI with the new analysis of the contract between the two neuroimaging approaches, SVM methodological expansion, frequencies of interest, including evoked responses and others. Overall, this is now more compelling and a stronger paper.

A remaining issue relates to the new plots in Figure 3 and also the interpretation of the β signal. The new plots showing spectral activity per condition are welcome and confirm that there is indeed a reduced β desynchronization in the Stop condition. However, with the spectral (color) plots it is still difficult to accurately determine the time course of the β activity, which here is particularly important given the mechanistic interpretations of β and was a concern of reviewers.

It would be advisable to include one further plot of a time series of the β filtered power in each condition separately as a separate time series. To put it another way – to show a line plot of the time resolved β power (rather than a spectrogram) with error bars. This would give a clearer view of whether the β plateaus, slows its desynchronization, slightly increases or whether the variance increases, all of which could be driving change in significance and are difficult to determine from the spectrogram. This is important for the mechanistic interpretation and particularly for the "Inhibitory role of β-band oscillations" Discussion section.

While there is agreement that the study "suggests an active change of processing in rIFG that is related to stopping, while processing proceeds 'as usual' in cAC-GO trials", the role of β here and whether this it has a primary stopping function or these changes are secondary to a different mechanism for stopping in the IFG is still undetermined here. The current Figure 3 and your aligning with the status quo theory suggests that β changes here are more secondary and related to a withdrawal of an ongoing action initiation process rather than new active stopping process. It should be discussed further and potentially highlighted that a reduction of an ongoing motor initiation process (e.g. reducing β ERD) is different from a specific, distinct and active "stopping" process.

Your current discussion was at times still felt to be implying that β was mechanistically a stopping signal (which would have been supported by an early increase in β with stopping) rather the possibility that β desync reflects motor initiation which was arrested (as implied by a slowing or plateauing or the ERD). An analogy might be the different between reducing pressure on an accelerator pedal, versus actively pressing a brake. In general this should be addressed throughout and the following are particular example areas of text which could be clarified to take account of this β discussion above:

"We suggest that these differences then result in the observed net β-band power increase before SSRT". – This implies β increases, whereas the data thus far presented doesn't support this.

"Inhibitory role of β oscillations" – This isn't definitively shown here.

"Our results strongly support the interpretation that the rIFG acts as the initiator of response inhibition and that β-band oscillations play an important role for its neural implementation." – This implies β is mechanistic for response inhibition – but again this isn't shown here definitively, and β changes could secondarily result from a pausing of action initiation.

Finally – whilst Figure 3 is a welcome addition – the curved greyed out area suggests a strong windowing effect from Time Frequency (TF) decomposition on short time window / epochs causing distortion at the edges which are worse for low frequencies (which have been masked out). This is unconventional to analyse and present like this, with normal practice being to perform the TF decomposition on the continuous data and to epoch power estimates afterwards, or to perform TF decomposition on a much wider epoch of data per trial and then re-epoch / chop down to the (same) time period of interest (e.g. -0.2 to 0.5). It is advised to repeat this analysis with better epoching – currently the windowing is artificially constraining the time and frequencies analysed.

https://doi.org/10.7554/eLife.61679.sa1

Author response

[…]

The appearance of low frequency activity in the spectrogram in Figure 3 and the question of whether the IFG – SMA inhibitory effect is also mediated by non – oscillatory evoked activity (which also appears in the low frequency range). This would require a further analysis looking to determine if the effects were truly specific to β or could also be found in non – oscillatory (evoked) activity. Furthermore, multiple reviewers felt that a better mechanistic account was needed of how such short timing effects (e.g. 20ms difference in the SVM model classification difference) could be driven by β with a cycle length of 50ms and much clearer explanation of why the β frequency ranges were so different across analyses. Notably – the 30-40 Hz signal in the Granger causality analysis does not appear to be the same signal being analyzed in the latency analysis. This either requires further analysis of matching frequencies or significant further explanation and more measured interpretation.

Correspondence between fMRI versus MEG was not felt to be as strong as described in the write up. This raised the question for some reviewers as to whether the regions of interest where truly chosen in a data driven way. A further control analysis of another region such as pre-motor cortex and left IFG was felt to be necessary to ensure that these effects seen are genuinely restricted and specific to the rIFG and pre-SMA in the time period of interest.

Elaboration on the above issues:

1. The author use a comparison of statistics rather than a direct statistical comparison of correlations. This comparison was felt to be invalid by a number of reviewers and a direct statistical comparison was requested (see below for details). Statistics – The argument for the difference between rIFG and pre-SMA (Figure 3D) is based on comparison of statistics rather than a statistic for comparison (see https://www.nature.com/articles/nn.2886). Even ignoring that, the p-value of 0.042 with N=50 does not seem to be very convincing and looking at the figure the two point clouds are not very different. It is unclear how the results will stand up if this essential comparison is done..

We agree with the above arguments. To understand the effects of changes in β power on the stop signal reaction time (SSRT) in more detail we performed a Bayesian multiple regression analysis with per subject SSRT-values as outcome variable and β power in presupplementary motor area (pre-SMA) and right inferior frontal gyrus (rIFG) in successful and failed trials, as well as the directionality asymmetry index (DAI) as regressors (please refer to question 15, which was related to correlation with SSRT, too).

We chose to model the SSRT as normally distributed across subjects, as it is (despite its name) not a directly observable reaction time but the result of an extraction procedure that involves multiple samples per SSRT-value. We z-normalized all variables before modelling and used a normal distribution with zero mean and a standard deviation of 0.5 as a the prior for each regressor. The likelihood was modeled as a normal distribution centered on the regression model with a standard deviation drawn from a half normal prior with standard deviation of 1. Posterior distributions of the parameters were computed in PyMC3 (Salvatier et al., 2016) using the Hamiltonian Monte Carlo No U-Turn sampler (NUTS), with 2000 tuning samples and 20.000 samples for the joint posterior distribution.

We compared three models for their expected posterior log-probability using a leave-one-out cross-validation (LOO-CV) procedure as suggested in Gelman et al. (2014). Model 1 used all regressor variables and their first order interaction terms, model 2 used all regressor variables, but no interaction terms, and model 3 additionally averaged together the β power in pre-SMA for successful and failed trials as the values of these regressors were highly correlated (r > 0.6). Thus, model complexity was reduced successively from model 1 to model 3. Model 3 achieved the best LOO-CV score (expected log posterior predictive density elpdloo = –64.65 ± 5.32, higher is better, maximum is at zero), but model 2 was within one standard error of model 3 elpdloo = –65.24 ± 5.10. Model 1 had a considerably worse LOO-CV score (elpdloo = –71.67 _ 4.53).

In the winning model 3, there was as 91.0 % marginal posterior probability for a negative regression coefficient of the β power in rIFG for successful trials, and a 87 % marginal posterior probability for a negative regression coefficient for the DAI, but only a 64 % marginal posterior probability for a negative regression coefficient for the average β power in pre-SMA, and a 40 % marginal posterior probability for a negative regression coefficient for the β power in IFG in failed trials.

For model 2, with individual regressors for β power in the pre-SMA, results were highly similar, with a 89 % marginal posterior probability for a negative regression coefficient for β power in rIFG in successful trials, and a 88 % marginal posterior probability for a negative regression coefficient for the DAI, with all other marginal distributions of regression coefficients being more symmetric around 0.

Given these results, we consider it most plausible that β power in rIFG in successful trials and the DAI predict SSRT outcomes, with all other regressors having no or a much smaller influence. We revised the respective paragraph (results: l. 362 ff, methods: l. 1519 ff) accordingly.

This new analysis supports the claims made previously on the role of β-band power in rIFG for stopping, and replaces the initial analysis which was indeed insufficient.

2. Statistic and corrections – Correlation analysis: For demonstrating the difference between rIFG and preSMA correlations these must be directly compared to each other rather than showing a significant correlation in one, and the absence of that in the other (please see Nieuwenhuis et al. 2011, Nature Neuroscience). Also, though not relevant when comparing preSMA and rIFG directly, when using multiple correlations these should be corrected for multiple comparisons. As we understand the analyses, four correlation analyses were performed preSMA and IFG for different contrasts (uSTOP and sSTOP). The rIFG correlation seems to only survive when not correcting for multiple comparisons and using a one-tailed p.

We agree with this point. Therefore, we decided to use a full Bayesian analysis as described above to address this issue.

3. Contrasts – "he authors almost exclusively report relative changes in MEG activity in one condition (successful stop, sSTOP) compared to another condition (attentional capture GO, cAC-GO). Even though the authors try to take into account that in one condition people pressed a button (cAC-GO) while they do not execute a motor response in the other (sSTOP) by specifying a certain time window, they still cannot control for differences related to the well-known β-decrease during a motor response, which starts several hundred ms before a response. Looking at the median RT, there is presumably strong β-desynchronisation during the time window of interest. So any 'β-increase' in sSTOP vs cAC-GO, which is interpreted as stopping-related, might simply reflect that there is less β decrease in sSTOP compared to cAC-GO, because there is not motor response. In fact, supplementary figure 2 shows that β strongly decreased (when pooled across conditions), and not increased, after the cue / before the response. In order to support the claim that a β increase (and not the lack of a decrease) is related to stopping, the authors would have to show that β in fact increases in sSTOP-GO compared to baseline activity (e.g. before the cue).

We agree with your observation that the reported increases in β power are relative and not absolute increases. We have now added the contrasts (“main effects”) for task vs. baseline to Figure 3 in the revised manuscript (TFRs for both conditions, sSTOP and cAC-GO). Nonetheless, those contrasts support our interpretation of a specific, though relative, increase in β power for the stop condition, when compared side by side with the cAC-GO condition. As you pointed out, β-band desynchronization (BBD) is related to motor-response preparation and BBD already sets in early after the initial GO signal. And indeed, until approximately 100 ms after the STOP/AC-GO signal occurs, we observe almost the same strength of BBD for both conditions (see Figure 3, panels A, B in the revised manuscript). This is confirmed by the contrast between both conditions showing that BBD almost cancels out in this time window (see Figure 3, panel C in the revised manuscript). While in cAC-GO trials, the BBD continues until and beyond the median RT (first dashed line in Figure 3), BBD is aborted in sSTOP trials already approximately 100–150 ms before the median SSRT (first solid line in Figure 3). Hence, from the perspective of an already initiated BBD, what we found is not just a lack of BBD after the stop signal compared to cAC-GO trials, but an early termination of BBD in sSTOP trials within the time window of interest. This means that our results do not just originate from the contrast of sSTOP with cAC-GO trials, but the BBD termination sets in already before the SSRT in the sSTOP condition, while in the cAC-GO trials no such termination was observed. This finding suggests an active change of processing in rIFG that is related to stopping, while processing proceeds ’as usual’ in cAC-GO trials, and the motor response will be executed. Yet, we didn’t find an absolute β-band increase before SSRT in sSTOP trials compared to baseline, which is in line with a recent ECoG study based on 21 patients (Chen et al. (2020), see Figure 4B, Figure S6). Thus, taken together for the results that relied on the contrast between the conditions sSTOP and cAC-GO, as well as their single contrast components (task-versus-baseline TFRs), we found clear evidence that the relative change of β-band oscillations specifically within the sSTOP condition plays a crucial role in response inhibition. We revised the respective paragraph in the discussion (l. 635 ff) accordingly.

4. Contrasts – "Figure 3 and Methods L. 1248 – I was not able to follow the authors' description of the difference between what is shown in 3A and 3B other than that 3A shows t-value and 3B – z-value. When you talk about z-transforming trials, did you z-transform the raw data or the TF power? If it is the former, then the colour in 3B would not correspond to z-value. What makes sense to me is that in both cases you averaged the power across trials for each subject but in 3A looked at differences in power whereas in 3B – at differences in latency but either I am wrong or the description is unclear."

We apologize for the confusion that Figure 3A/B created. Both subfigures rely on the same data. While 3A shows the result of a statistical test that asks for differences between both conditions (non-significant areas outside the clusters were displayed transparent), 3B shows the result of simply subtracting power of cAC-GO from power of sSTOP trials, i. e. it is closer to the raw data.

The data were processed as follows:

1. Trial-wise time-frequency analysis for each subject and condition (task and baseline segments for sSTOP and cAC-GO conditions).

2. z-transformation of these single trials (in the frequency domain, not the raw data), by trial-wise baseline correction (standard deviation of the baseline of both conditions pooled was used).

3. Averaging all z-transformed trials per subject. For the statistics (Figure 3A), the average over trials of each subject and condition was used as input for the cluster-based permutation test (sSTOP vs. cAC). For the simple difference shown in Figure 3B (sSTOP – cAC), we grand-averaged both conditions across subjects.

So why did we show both representations? The appearance of the early significant cluster in rIFG (Figure 3A) simply states that there indeed exists an early difference between both conditions. However, cluster-based permutation tests do not allow a precise statement about the onset of a certain effect (Sassenhagen and Draschkow, 2019), as cluster border locations come without statistical guarantees of any kind (just cluster existence is statistically guaranteed). Since we aimed to test our hypothesis for differences between the onset timing in rIFG and pre-SMA, we therefore used averaged, z-transformed TFR data obtained by step 3 as described above. With Figure 3B we aimed to highlight these different approaches but given the confusion we removed it and added the single condition TFRs (task vs. baseline) instead (see point 1). Additionally, we improved the corresponding section in the methods section (l. 1418 ff). Thus, we hope that the comprehensibility of the manuscript was improved.

5. fMRI – "Description of Figure 2B in the text implies that there was specific activation of rIFG and preSMA in the fmri study, but the figure shows broad activation including bilateral prefrontal, parietal and occipital areas. Thus, even though the authors state that they did the source reconstruction for analyzing sources based on the current experiment and not based on previous literature picking out preSMA and rIFG for the next analyses seems to be largely related on previous literature and not the current study. For example, how would these analyses look like for left IFG or premotor areas? While I understand the rationale for looking at rIFG more closely based on the current data (due to the early onset of spectral differences), picking out preSMA from the other sources seems to be based on previous literature."

It was not our intention to give the impression that only rIFG and pre-SMA activity were found in the fMRI contrast. We revised the manuscript to avoid this impression (l. 239 ff). In the original manuscript we had reduced the comparison to the specific regions we were interested in, based on our hypotheses. We also would like to clarify that the overall approach of this study was highly hypothesis driven with the purpose to investigate timing and role of rIFG and pre-SMA. This is because, both, pre-SMA and rIFG, represent the key areas involved in response inhibition as shown by many previous studies. In other words, the question “Who is first, pre-SMA or rIFG?” has crystallized over many previous studies, but had not been resolved before, hence our strong focus on this hypothesis/comparison.

The term “data-driven” referred to the identification of coordinates of sources at which oscillatory activity can be actually observed in the acquired MEG data. This is in contrast to a literature-based selection of coordinates. The word “coordinates” was inserted in the misleading sentence (l 191 ff.) to emphasize this approach. Many former MEG studies a priori defined anatomical ROIs and their coordinates (or masks) instead and extracted data using virtual channels without testing or at least without reporting whether these locations actually revealed significant activity during response inhibition given their paradigm and measurement data or not. This seed-based approach, however, is explicitly discouraged in the current best practice guidelines for MEG analyses (Gross et al., 2013). The recommended approach of contrasting appropriate conditions that we used enabled us not only to identify coordinates of peak activity for rIFG and pre-SMA, i. e. the location with the strongest effect across subjects, but also excluded that the effects we analyzed were due to leakage from other sources. Indeed, other sources were active at the same time window, too, and we decided to present them in the figure as well to provide a better view of the overall activity. Note that rIFG and pre-SMA were also revealed by the contrast using a smaller (cluster) _-value of 0.01, while other sources would not have survived this threshold (like l-IFG, lAIns, Fornix). Still, the overall approach of this study was highly-hypothesis driven and aimed at revealing timing and roles specifically of rIFG and pre-SMA during response inhibition. Both key areas could be localized quite accurately using fMRI and MEG (based on the same task), and within close distance to each other – as revealed by the comparison between both methods.

6. fMRI – "In Figure 2 the correspondence between MEG and fMRI results does not seem to be very convincing with the MEG results surprisingly being much clearer. Particularly the absence of activation of bilateral motor strips and pre-SMA in fMRI is puzzling.

a. Would it be possible to quantify the overlap and say how it would compare to other published MEG/fMRI comparisons with the same task?

In order to compare the source activation between both methods, MEG and fMRI, we created another plot that shows the overlap (see Figure 2C in the revised manuscript). The overlap was defined by the product of absolute, normalized t-values for each voxel i, |ti,MEG| _ |ti,fMRI| (MEG data were interpolated to match the fMRI resolution). Normalization was performed for each method separately by division with the maximum t-value obtained by the respective method, resulting in normalized t-values between 0 and 1.

Comparing the sources revealed by the MEG and fMRI contrasts showed an overall good correspondence. Given the 1 cm grid used for the MEG source reconstruction, the distance between peaks of both methods were rather close for rIFG (7 mm), pre-SMA (15 mm), lAIns (11 mm), and lMFG (15 mm). The peaks of premotor areas (rPMC, lPMC) as well as l-IFG were less close to the fMRI results (20–24 mm difference between methods). However, particularly the sensorimotor areas revealed a rather wide-spread distribution of source activation and referring just to the MEG peak voxel may be over-conservative. Yet for the key areas that were most important for this study, rIFG and pre-SMA, we found a very good correspondence and thus could validate that the MEG source reconstruction was quite accurate for these areas. We revised the respective paragraph (l. 248 ff) accordingly.

So far, to our best knowledge, there are no other combined MEG/fMRI studies yet that used a stop signal task. However, if we receive a hint with a reference, we will be happy to include it in the manuscript and also discuss the comparison.

b. Do you have any explanation for the discrepancies?

We agree that the absence of bilateral motor strips in fMRI was confusing. Since we were focused on the comparison of rIFG and pre-SMA for successful stopping (i.e., sSTOP > cAC-GO) between both methods, we didn’t show the fMRI contrast cAC-GO > sSTOP in the original manuscript (also see next point, c.). Now we included this contrast in Figure 2B in the revised manuscript and it can be seen that sensorimotor cortex shows activation, too. However, we disagree that pre-SMA activation was absent in the fMRI analysis. In the Results section, we reported the exact coordinates of pre-SMA activation revealed by the contrast (l. 211f., original manuscript). Yet, due to the surface projection, it was not fully visible. Thus, discrepancies should be eliminated now (see also discussion in a./c.).

c. Also for me it is not obvious that higher β power should necessarily correspond to higher BOLD as at least for movement higher β is associated with movement inhibition. Have you tried looking at the other direction of the fMRI contrast, perhaps the missing sensorimotor cortex is there?

Indeed, when looking at the fMRI contrast cAC-GO > sSTOP, the sensorimotor cortex is active then as expected, also see b. However, please note that the sign is different compared to the MEG β-band activity. Since the BOLD response cannot be limited to the same tROI as in the MEG, the motor response (button presses) in cAC-GO condition might dominate the contrast in the fMRI activity map.

d. Is the IFG latency also significantly different than the latency in the other brain regions in which β band activity is observed?"

Our original research question was to find out which of the two key areas, rIFG and pre-SMA, comes first during response inhibition. The limitation to analyze only these two brain regions was highly hypothesis driven (see also point 5.). For the sake of completeness, we presented all sources obtained by contrasting sSTOP and cAC-GO trials during the tROI. Indeed, rIFG and pre-SMA were active as hypothesized and allowed us to obtain the most accurate (data-driven) coordinates for our analysis. The analysis of all other sources would be just exploratory, and the same applies to the connectivity analysis. Still, it is interesting to see that even at this level, rIFG seems to be the dominant sender within the motor/inhibitory network (see Author response image 1). However, these results were not significant when correcting for multiple comparisons (42 links in total).

The mean onset latency of rIFG was indeed the earliest compared to all other sources. For a statistical comparison of the latencies of all brain regions in which β band activity was observed, we additionally computed a repeated measures ANOVA. Since onset latency was defined (roughly) by a relative positive peak, subjects that did not show such a peak at least in one of the seven sources were excluded (to be consistent with the latency-statistics reported in the original manuscript). Thus, the number of subjects still included would decrease to n = 40 and only the latency difference between rIFG and lAI was found to be significant by post hoc tests. However, we would like to point out that interindividual differences in the neural response of stopping in the rIFG and in the β band may vary a lot (see, e. g. in Wessel et al. (2013), Figure 2, Swann et al. (2009), Figure 5). This said, statistical power for the latency analysis of all sources might be too low. Yet, our study was strongly focused on revealing the timing specifically of rIFG and pre-SMA during response inhibition, and we clearly found a temporal precedence of rIFG over pre-SMA activation.

7. SVM – "The β latency analysis is supported by the SVM analysis which uses a wide-band frequency input. To strengthen the case that the β band is mechanistically primary here, it would be useful to repeat the SVM analysis using a β filtered input to the SVM. This would confirm that the effects seen in the SVM are driven by the β band.

We would like to thank the reviewer for suggesting this additional analysis. In the original analysis, we used broadband filtered data for the SVM analysis. Yet, the result of temporal precedence of rIFG over pre-SMA as found for the TFR analysis was confirmed. To check if the effect revealed by the SVM broadband analysis was β-band driven, we additionally performed the same SVM analysis on β-band filtered data (still in the time-domain), but with time-embedded vectors as input as suggested in point 9. Then the above-chance level classification was significant in rIFG at 132 ms (p < 0.0001), and in pre-SMA at 142 ms (p < 0.0001), tested between 100–350 ms. We added these results as additional plot in Figure 4 in the revised manuscript (panel B) and updated the respective paragraph (l. 335 ff) accordingly.

These findings are close to the one of the broadband analysis (rIFG at 129 ms, pre-SMA at 163 ms) and support the timing of the TFR analysis. However, we observed that the accuracy of the β-band filtered analysis decreased compared to the broadband analysis. We guess that the reduction of accuracy might be due to the missing contribution of evoked responses (as indicated by a reviewer in point 9). This reduction was in particular found after the median SSRT. Since we found significant differences between conditions in the evoked response also after SSRT (Author response image 4A) , we guess that indeed evoked responses are crucial for the high classification accuracy after the SSRT, that decreased in the β-filtered compared to the broadband SVM analysis. Yet, before SSRT, we found an early onset of significant above-chance classification when using β-band filtered data, indicating that β-band activity may indeed play an important role for the initiation of response inhibition in the rIFG. However, β-band filtering reduced the power of the signal and individual accuracy curves became quite noisy. Thus, individual onset definition would result in too many exclusions and we were not able to perform the same individual onset statistics as for the broadband data. Hence we could not provide strong support for the temporal precedence of rIFG over pre-SMA when using β-band filtered data. Nonetheless, since the additional SVM analysis on β-band filtered data still revealed an early above-chance classification onset in the rIFG, we conclude that β-band oscillations are carrying the crucial information to reliably classify sSTOP trials and hence, play a leading role in initiating response inhibition in rIFG.

Author response image 1
Exploratory connectivity analysis during the tROI.

Spectral-resolved conditional Granger causality (cGC) between all source target pairs (links) was analyzed. Blue curve: cGC for sSTOP trials, green curve: cGC for cAC-GO trials, bounded lines: standard error of the mean. There were no significant differences between both conditions in the frequency range of 8-44Hz when correcting for all 42 links. However, significant differences were found if correction for multiple comparisons was reduced to six links (highlighted in yellow). Uncorrected significance is highlighted in grey.

8. SVM – "Just as a point of clarification, it appears that the onset analysis of the classification reveals a much later time of onset. It would be helpful to understand why this classification time course appears slower than the time course of β power, and whether the difference in classification between IFG and SMA should be interpreted in the same way as the difference in power latency."

The discrepancy can be explained by the differences between the two methods employed. First, the onset values revealed by the latency analysis using TFR as well as the SVM approach depend on the thresholds applied with respect to the peak latency, i. e., higher thresholds would result in later onset values, and the peak characteristics might differ between both analyses. Second, in contrast to the TFR onset analysis, the SVM accuracy curves were low-pass filtered to reduce noise before determining the onset. We therefore believe that the later time onset is due to this filtering step in the analysis pipeline. To check this, we performed an additional analyses using a jacknife approach (Miller et al., 2009) that did not rely on low-pass filtered accuracy curves. In consistency with the SVM analysis of the revised manuscript, we used time-embedded data as SVM input (see questions 7 and 9). This analysis confirmed a mean IFG onset of 130 ms (with low-pass filtering: 178 ms) and a mean pre-SMA onset of 178 ms (with low-pass filtering: 200 ms), p = 0.003. Thus, we could confirm that the slower latency values of the individual onset SVM statistics compared to the TFR onset analysis might be due the low-pass filtering for onset definition. More importantly, at least for the rIFG, the onset value found using unfiltered data in combination with the Jacknife approach matched quite well with the onsets found by the TFR analysis, revealing a substantial agreement between both methods (difference was only 7 ms). However, we are aware that the particular method employed might result in slight shifts of the onset values. Yet, our main finding was that the SVM analysis revealed additional evidence for the temporal precedence of rIFG over pre-SMA during response inhibition, and this was found independently of the method used.

9. SVM – The authors report that the SVM was point-by-point. One reviewer commented "I don't understand what that means because the whole idea of SVM is that it's multivariate so could possibly be sensitive to things like β. Point-by-point would mean that somehow the raw signal amplitude at each point is different between trial types so unless there is an evoked response there, I don't see how this would be the case. If it's really point-by-point the latency difference with β power would make sense because the time resolution of TF analysis is limited, in this case to 150ms windows around 20Hz."

In the original analysis we indeed employed a univariate analysis and the classification was done point-by-point, resulting in a linear discriminant analysis in the end. Hence, it’s likely that an evoked response was involved in the effects we found. To be sensitive in particular to β-band oscillations, we performed an additional analysis on β-band filtered data using time-embedding, as suggested in point 7. Using time-embedded vectors as input in general increased the classification accuracy. Hence, we updated the manuscript with results revealed by this multivariate approach (Figure 4A in the revised manuscript). However, even when using a time-embedded multivariate approach, the β-band filtered analysis resulted in lower classification accuracies compared to the broadband analysis (see point 7), which is obviously more heavily influenced by evoked responses. This could be interpreted in a way that successful stopping is also supported by evoked responses. Nevertheless, the early above-chance classification in rIFG was found in the β-band filtered analysis as well. Thus, evoked responses do not seem to be required for the classification of successful stop trials, and most importantly both SVM analyses (using broadband data as well as β-band filtered) confirmed our onset analysis based on TFRs.

10. Frequency of interest – "There is an early 'relative' β-increase in rIFG, which already starts ~100ms after the cue. Its spectral components seem quite different from the later occurring β change and it includes a low-frequency component. Did the authors assess whether this could be an evoked response with different frequency components in the spectral domain? The cluster's duration seems <100 ms which would correspond to less than 2 cycles of a 20Hz oscillation. Also, would this cluster be significant when not 'spilling over' into the later β-cluster and the low-frequency cluster?"

We agree with the reviewer that both questions should be addressed. First, to ensure that the early relative β-increase in rIFG is not just a spectral component of an evoked response, we computed the ERF of the virtual channel and then created a TFR out of the averaged data. Thus, we could identify the spectral components of the ERF (Author response image 2A, B). However, when contrasting both conditions as done with the original TFR analysis, we could not find a significant difference in the rIFG (Author response image 2C). Hence, we could not find evidence that the early β-band cluster at the rIFG revealed by the time-frequency contrast could be related to evoked-like activity.

Second, we tried to exclude that the early cluster was just a spilling over from other clusters like the low-frequency cluster that was found within the same time range and the later β-rebound cluster (Author response image 3A). Note that in contrast to the original manuscript, this new figure shows the TFR statistics performed between 2–44 Hz (as suggested in point 11). In this reanalysis the low-frequency cluster was not significant anymore for the contrast of interest. However, a larger cluster like the later β-rebound would survive permutation-based statistics more easily than a smaller, separate cluster, potentially masking other effects. Hence, we additionally tested in a smaller time-frequency window around the early cluster (100–200 ms, 12–35 Hz) and found that the cluster was still there (Author response image 3B). Thus, it seems to be related to a distinct neural process. We interpret the different frequency composition of the early and later relative β-increase in a way that the later β-band increase (after median SSRT) likely reflects the well-known β-rebound effect, while the earlier increase might be specific to the initiation of response inhibition.

In conclusion the additionally performed analyses support our interpretation that early relative β-increase in rIFG represents a distinct oscillatory neural correlate that is related to the initiation of response inhibition and is not related to spectral components of an evoked response.

Author response image 2
Time-frequency representations of evoked responses in rIFG.

First and second solid line indicate SSRT and SSRTmax, first and second dotted line indicate 10% and 50%=percentiles of RTAC-GO for selected AC-GO trials with RTAC-GO > SSRT (RTAC-GO is the duration between AC-GO signal and button press). Time axis locked to STOP and AC-GO signal (Os). (A) To obtain evoked responses, successful stop (sSTOP) trials were averaged (for baseline correction, the mean of a 500ms long baseline segment that ended 100ms before the Go signal was used, no low-pass filter applied) and then transformed to TRFs. (B) Same as in (A), but for correct attentional capture go trials (cAC-GO). (C) Contrast of sSTOP versus cAC-GO trials cluster-based permutation test, tow-tailed, αcluster = 0.05, n=59, no significant clusters found

11. Frequency of interest – "Recent evidence, already cited in this manuscript (Chen et al. 2020) suggests that an earlier evoked response from IFG might precede β changes during reactive stopping. The Figure 3A low frequency activity also supports this. Given the SVM changes on broadband data, the findings in Figure 3A and the previous literature, it would be informative if the authors could analyze if evoked (non – oscillatory +- very low frequency δ which is likely equivalent) activity is also earlier (and potentially) causal in IFG --> pre-SMA."

Although Chen et al. (2020) found an early evoked potential in rIFG, it remains unclear whether this EP was specific to stopping or not and whether it set in before the early β changes we found. There are three issues to consider in this context. First, since the results of that study rely on a simple SST, not a selective SST, it was not possible to compare them with an approriate control condition. Second, Chen et al. (2020) showed that IFG–STN interaction is mediated primarily by a fast EP rather than by oscillatory synchronization. Yet, the onset of the EP was compared to a β increase that was observed after SSRT. As mentioned in our discussion, the β increase after SSRT may reflect a rebound of activity related to action termination, not the actual initiation of response inhibition. From this perspective it seems quite likely that the EP set in earlier than the increase of the late β power due the rebound. Thus it remains an open question whether their EP effect precedes our early β-band effect. Third, Chen et al. (2020) showed that cross-correlation between the EPs from rIFG and STN correlated with SSRT, revealing an important function of the hyper-direct pathway for stopping, but it was not the rIFG EP alone that could be related to stopping. Hence, although the cited study revealed several remarkable results, it should be considered with care when making conclusion about the relationship between evoked and oscillatory responses during the initiation of response inhibition. However, we agree with the reviewer that evoked responses and very low frequency δ oscillations should be analyzed in addition to the oscillatory results we presented.

First, to analyze evoked responses, we generated event-related fields (ERFs) from virtual channel time course data. Since these data originate from a principal component analysis (PCA), signs of the ERFs may be flipped randomly across subjects. To correct for this, we applied the method described by Vidaurre et al. (2018). In this analysis we did not find any differences between conditions before the SSRT at rIFG and pre-SMA (Author response image 4A). Second, to analyze very low frequency oscillations, we created TFRs starting at 2 Hz and using a 500 ms time window. We couldn’t find any significant differences between conditions for low frequency activity either (Author response image 4B). Also note that for point 10., we checked whether effects in the low-frequency components of the TFRs could originate from effects in evoked responses, which could not be confirmed. Hence, when using the contrast with the cAC-GO control condition, neither effects in evoked responses nor in very low frequency oscillations were setting in already before the early relative β-band increase were found in rIFG. Yet, since absence of evidence is no evidence for absence, we cannot exclude that evoked responses contribute to the initiation of response inhibition. However, in our high-powered study we could not find any evidence that evoked responses or very low frequency oscillations play a crucial role for the initiation of response inhibition.

Author response image 3
Time-frequency statistics focused on early β-band cluster in rIFG.

Contrast of z-transformed sSTOP versus cAC-GO trials (cluster-based permutation test, two-tailed, αcluster = 0.05, n=59, significant cluster is outlined). (A) Tested between -200-500ms and 2-44Hz as in the analysis showed in Figure in the revised manuscript. (B) Statistics performed in a smaller time window (100-200ms) and only between 12-35Hz. Thus, we tried to exclude that the early significant cluster revealed by the original analysis is just spilling over from the larger and later β-rebound that would survive permutation-based statistics more easily than a smaller, separate cluster. And indeed, although we specifically tested around the outline of the early cluster in rIFG, it was still significant and not a spill-over from other clusters.

12. Frequency of interest – "Much of the discussion and the initial results regarding spectral power are focused on the broad β band from 12-30 Hz. Yet the analysis of Granger causality and DAI are focused on a slightly higher frequency range, 30-40 Hz, which is not usually considered as β when discussing the stopping nature of IFG (in fact, some would consider this in the range of low γ). How should one reconcile these different frequencies?

a. Can the authors re-analyze in matched frequency bands and discuss the non – concordance in frequency given.

We agree with the reviewer that the parts of the TFR and connectivity (cGC) results were significant in different frequency bands. Yet, at the first step, both statistical analyses that aimed to test for differences between conditions, sSTOP and cAC-GO, had been actually performed in the same range, 8–44 Hz. Allow us to recapitulate how the analysis pipeline looked like for both methods before we try to reconcile the different outcomes.

Author response image 4
Additional analysis of evoked responses and very low frequency oscillations.

(A) Statistical comparison of evoked responses, sSTOP red curve, cAC-GO blue curve. A cluster-based permutation test was performed between 0 and 500 ms. Significant difference are highlighted by grey boxes. (B) Oscillator analysis including very low frequencies (2-44Hz). Here, the contrast of z-transformed sSTOP versus cCA-GO trials is show (Cluster-based permutations test, two-tailed, αcluster = 0.5, n=59, significant clusters are outlined).

In case of the TFR analysis (Figure 3A in the original manuscript), the group statistics in the range 8–44 Hz revealed an early significant cluster in the rIFG (approximately in the range 10–34 Hz, peak at 25 Hz) and we supposed that this cluster could represent early processes of response inhibition. Since this cluster was spilling over to a lower frequency cluster, it was necessary to limit the definition of individual latency and power values to a frequency range in which not two obviously different neural effects take place simultaneously (i. e., α- and β-band oscillations). To draw a line, we decided to rely on the initial β band, 12–32 Hz, that was obtained by the sensor statistics during the tROI (100–350 ms) and that also was used for the source reconstruction (the question incorrectly stated a range from 12–30 Hz). Thus we could ensure that the TFR effect we analyzed matched best with the source reconstruction that revealed the rIFG and pre-SMA sources. Also note that in our reply to point 10 we showed that the β-band cluster was revealed independently of the lower-frequency cluster during the same time window.

In case of the connectivity analysis, we tested for differences in cGC, also within 8–44 Hz as for the TFR analysis at group level. Here, we found a broad cluster in the range 16–40 Hz with its peak between 24 to 28 Hz depending on the time window (at 28 Hz in the tROI). However, the cGC curves of both conditions were significantly different only between 30–40 Hz. Yet one should keep in mind two issues when interpreting these results. First, significance boundaries should not be overinterpreted. As pointed out by Sassenhagen and Draschkow (2019), frequency precision of statistical claims should not be overestimated when using cluster-based permutation tests. Given a higher statistical power, a wider range would probably be significant and hence, overlapping with the classical β-band range. Second, we don’t know exactly why the statistical results only reveal the right hand slope of the GC peak within the tROI. We guess that the data is just too noisy because of the limited amount of trials. Also note that the noise that comes with the GC analysis differs from the one of the TFR analysis. Additionally, as DAI is based on the estimated cGC values, the same considerations apply to the correlation between DAI and SSRT (also see b.).

In conclusion we would like to highlight the following aspects:

• The range of significant cGC differences was in fact overlapping with the cluster revealed by the TFR analysis, even if only in a small range between 30–32 Hz.

• Although the frequency band that showed significant differences between conditions (30-40 Hz) is not usually considered as β when discussing the stopping nature of IFG, the peaks obtained by both methods (28 Hz by cGC and 25 Hz by the TFR analysis) still could be assigned to the higher β band.

• The noise is not the same for the TFR and GC analysis – and this may differentially affect different frequencies.

• The constraints of cluster-based permutation tests should be considered when interpreting the significance of frequency boundaries.

Considering these facts, the slightly different spectral results of connectivity and TFR analysis can be reconciled. As a consequence, we interpret the effects revealed by both methods in a way that they are related to the same neural processes of response inhibition.

b. Moreover, the relation between DAI and SSRT is specific to 34 Hz. I understand that this has been corrected for multiple comparisons, but does this correction just account for the six discrete frequencies between 30-40 Hz, or does this account for the entire β band (or even all frequencies for that matter)? "

First, we identified differences between conditions in cGC between 8–44 Hz, using a cluster-based permutation test. This test revealed a significant frequency range between 30–40 Hz. Second, DAI was correlated with SSRT for each frequency in this significant range (30, 32, 34, 36, 38, 40 Hz). To identify whether a correlation was significant, a bootstrap method was applied, and the α threshold used for the CI was corrected by the number of the frequencies tested, i. e., corrected for six comparisons. Hence, since the correlation between DAI and SSRT was only tested for frequencies that were found to be significant at the first level, it did not seem appropriate to us to apply the correction for multiple comparison to the whole frequency range tested at the first level.

13. Frequency of interest – "This relates to interpretation – could the authors discuss and reference how an oscillation with a cycle length ~ 50ms could causally impact one region in the very short time frames demonstrated in this study (e.g SVM difference is 20ms in Figure 3C, ie less than half a cycle)."

We try to answer this question by illustrating a basic principle in physics. Allow us to use an example from simple electromagnetism and take a radio station that suddenly interrupts the playing of a song in order to announce breaking news. As soon as the sending radio station stops the song, this change has an instantaneous impact on the receiving device (only delayed by the propagation time through the medium, which is extremely short). The stopping of the song is not dependent on its frequency composition, i. e., low frequencies will stop at the same time as higher frequencies. Hence, the frequency of a process has nothing to do with the physical latency between sender and receiver. Likewise, the difference between two observed latencies is not limited to a cycle or multiples of a cycle.

As an additional demonstration, we simulated an auto-regressive model of a coupled system, where process A is oscillating with 20 Hz and process B with 30 Hz, and where A is driving B with a delay of 2 samples (corresponding to 16.67 ms with the sampling rate of 120 Hz used). When analyzing spectral Granger causality, we would expect no GC from A→A, B→A, and B→B. Only A→B should reveal a 20 Hz component. Indeed, this is what we obtain as results by our simulation (Author response image 5). Thus, it is well possible to detect Granger causality of signals with a period (50 ms) that is considerably larger than the coupling delay (16.67 ms).

Author response image 5
Granger causality simulation.

Simulation results of an auto-regressive model of a coupled system, where process A is oscillating with 20Hz and process B with 30 Hz, and where A is driving B with a delay of 16.67 ms. When analysing spectral Granger causality, we would expect no GC from A→A, B→A, and B→B. Only A→B should reveal a 20 Hz component. Indeed, this is what we obtained as results by our simulation. Thus, it is well possible to detect Granger causality of signal with a period (50 ms) that is considerably larger than the coupling delay (16.67 ms).

14. Granger causality – "Figure 4 – why does Granger causality go down with frequency? I would expect it to be flat except where there is a significant relation between signals. Could it be some kind of power confound or normalisation issue?"

We think that the GC does not go to zero because multiple frequencies contribute to the GC estimate, this makes the GC curves not flat outside the β-band peak. We additionally checked the influence of the estimation bias and ERF like frequency contributions. So far, they did not seem to have a substantial impact on the GC distribution.

Furthermore, there are many other MEG studies that show a similar GC distribution over frequency when GC estimates were computed in a similar manner. Please refer to e. g., Recasens et al. (2018) (Figure 5A, B), Rassi et al. (2019) (Figure 2C) or Seymour et al. (2019) (Figure 2).

However, we think that even if there were a potential disadvantage for higher frequencies (i.e., smaller GC values than for lower frequencies) due to normalization issues, this would not affect our conclusions because of two reasons: First and most importantly, we mainly interpret the GC results regarding the directionality. Second, we do not compare frequencies with each other within one GC analysis, but the difference between the conditions, sSTOP and cAC-GO. In conclusion, we don’t expect that the GC distribution as reported effects the interpretation of our results in any way.

15. DAI – "L. 475 – Was the correlation of DAI with SSRT independent of the IFG power correlation i.e. was the DAI regressor significant in the presence of rIFG power? If not, perhaps the DAI result could be explained by varying rIFG SNR that is known to confound directional measures."

Yes – an independent influence of DAI on SSRT had a probability of approximately 90% when controlling for rIFG power. This can be seen from our new Bayesian analysis that included rIFG power and DAI as additional regressors at the same time. For the results, please refer to our reply to question 1.

16. Clusters – "L. 1152 – the description of the cluster test seems to be inaccurate. If you took cluster sizes below 0.5th percentile of the distribution that would mean that you took very small clusters as well as very large. What usually happens is that people take large clusters for both positive and negative direction and apply Bonferroni correction for the two directions. Also why was a conservative threshold used p<0.01? Would there be additional clusters for p<0.05?"

We apologize for the incorrect description. The text was corrected accordingly: “Absolute cluster values above the 99.5th percentiles of the distribution of cluster sizes obtained for the permuted datasets were considered significant. This corresponds with an α threshold of _ = 0.01 that was Bonferroni-corrected for a two-tailed test.”

We used a more conservative threshold as we aimed to get crisp clusters and even more reliable results. This is possible due to the relatively large sample size used in our study. A threshold of p < 0.05 did not reveal additional clusters.

17. Code – "Although the authors included links to the data and the custom code in the cover letter, they do not appear in the paper itself. It would be helpful to include a supplementary table pointing to the toolboxes and custom functions used for each analysis. I was wondering particularly about cGC as no reference for the implementation was given and the conditional part sounds like more than is implemented in Fieldtrip."

We already provided a detailed list as readme file in the GitHub repository (https: //github.com/meglab/acSST). This list connects results (ordered by figure numbers) with code/scripts and data. However, we additionally created an overview box with all toolboxes used and the most important custom scripts (see Box 1 in the revised manuscript).

Conditional GC (cGC) is indeed implemented in FieldTrip, so we didn’t add an additional reference. Detailed script configurations can be found on our GitHub repository, too. For more information also consult the FieldTrip online documentation (https://www.fieldtriptoolbox.org/example/connectivity_conditional_granger/). The block approach we applied was cited. If you think a specific reference would be helpful in this context, please tell us and we’re happy to add it.

[Editors' note: further revisions were suggested prior to acceptance, as described below.]

A remaining issue relates to the new plots in Figure 3 and also the interpretation of the β signal. The new plots showing spectral activity per condition are welcome and confirm that there is indeed a reduced β desynchronization in the Stop condition. However, with the spectral (color) plots it is still difficult to accurately determine the time course of the β activity, which here is particularly important given the mechanistic interpretations of β and was a concern of reviewers.

It would be advisable to include one further plot of a time series of the β filtered power in each condition separately as a separate time series. To put it another way – to show a line plot of the time resolved β power (rather than a spectrogram) with error bars. This would give a clearer view of whether the β plateaus, slows its desynchronization, slightly increases or whether the variance increases, all of which could be driving change in significance and are difficult to determine from the spectrogram. This is important for the mechanistic interpretation and particularly for the "Inhibitory role of β-band oscillations" Discussion section.

We thank you for suggesting this additional data representation. Indeed the color coding of the spectral plots is hard to interpret between the transition from negative to positive values. Hence, we averaged the β-band power within the frequency range used for the latency and power analyses, i. e., 12 to 32 Hz, based on z-transformed spectral data as used as input for the cluster-based permutation statistics in panels A–C. We added the results to Figure 3 as panel D. What we found was an increase of β power specifically for the rIFG in the sSTOP condition at the beginning of the temporal region of interest (100–350 ms) for the rIFG in the sSTOP condition, independent of a potential influence of the lower frequency cluster. To ensure that the shape of the averaged curve was not biased by lower frequency components, we also averaged for 14–32 Hz and 16–32 Hz, and indeed, the characteristics of the averaged curve particularly for the rIFG in the sSTOP condition did not change. Thus, there is evidence that there is indeed a termination of β-band desynchronization (BBD) in the rIFG in the sense that it was not just a slowing or plateauing of the BBD. We did not find an increase in synchronization strong enough such that β power was increased above baseline, but values reached baseline levels in rIFG. In this context, we would like to point out that there is big variation of β-band power changes across subjects. Some of them actually indeed showed a strong β increase above baseline levels before SSRT. Additionally, please note that the change of β-band power in the sSTOP condition in the rIFG is different from the one in pre-SMA and in the other areas (Figure 3, panel D). The rIFG is the only source revealing an early difference in the sSTOP condition compared to the AC-GO control condition. It is also particularly different from the premotor areas (rPMC, lPMC) that show the typical β-band modulation expected during planning and execution of a motor response.

While there is agreement that the study "suggests an active change of processing in rIFG that is related to stopping, while processing proceeds 'as usual' in cAC-GO trials", the role of β here and whether this it has a primary stopping function or these changes are secondary to a different mechanism for stopping in the IFG is still undetermined here. The current Figure 3 and your aligning with the status quo theory suggests that β changes here are more secondary and related to a withdrawal of an ongoing action initiation process rather than new active stopping process. It should be discussed further and potentially highlighted that a reduction of an ongoing motor initiation process (e.g. reducing β ERD) is different from a specific, distinct and active "stopping" process.

Your current discussion was at times still felt to be implying that β was mechanistically a stopping signal (which would have been supported by an early increase in β with stopping) rather the possibility that β desync reflects motor initiation which was arrested (as implied by a slowing or plateauing or the ERD). An analogy might be the different between reducing pressure on an accelerator pedal, versus actively pressing a brake. In general this should be addressed throughout and the following are particular example areas of text which could be clarified to take account of this β discussion above:

"We suggest that these differences then result in the observed net β-band power increase before SSRT". – This implies β increases, whereas the data thus far presented doesn't support this.

"Inhibitory role of β oscillations" – This isn't definitively shown here.

"Our results strongly support the interpretation that the rIFG acts as the initiator of response inhibition and that β-band oscillations play an important role for its neural implementation." – This implies β is mechanistic for response inhibition – but again this isn't shown here definitively, and β changes could secondarily result from a pausing of action initiation.

We agree that there are different theories on the nature of response inhibition, i. e., either as a pause of an already initiated go response, or as a distinct active stopping process (interpreted as a “brake”), as suggested by Aron et al. (2014). We now highlight this in the revised discussion. More importantly, we now clarify better that we did not provide evidence for a causal role of β-band changes during response inhibition. Hence, we considered our conclusions about β-band oscillations more carefully in the Discussion section. Yet, we would like to point your attention to the fact that we found significant Granger causality in time windows relevant for stopping, but not earlier, suggesting that active signaling from rIFG to pre-SMA specifically in the β band is an important part of the neural dynamics underlying stopping. In sum, the two findings, i. e., rIFG as source of stopping relevant information transfer in the β band and a termination of β-band desynchronization specifically in rIFG already before SSRT, suggest that response inhibition is implemented in part by β-band modulations. We hope that the revised interpretation in the Discussion section about the role of β-band oscillations has now gained in clarity with respect to how it is supported by our results.

Finally – whilst Figure 3 is a welcome addition – the curved greyed out area suggests a strong windowing effect from Time Frequency (TF) decomposition on short time window / epochs causing distortion at the edges which are worse for low frequencies (which have been masked out). This is unconventional to analyse and present like this, with normal practice being to perform the TF decomposition on the continuous data and to epoch power estimates afterwards, or to perform TF decomposition on a much wider epoch of data per trial and then re-epoch / chop down to the (same) time period of interest (e.g. -0.2 to 0.5). It is advised to repeat this analysis with better epoching – currently the windowing is artificially constraining the time and frequencies analysed.

We are aware that to avoid these masked-out areas, we would have to analyze longer time epochs. However, since our analysis was highly hypothesis driven and focused on β-band oscillations after the stop stimulus, we a priori defined the time period of interest between 100–350 ms after stimulus onset (see “tROI” in the manuscript) and pre-processed the raw data for this purpose. We additionally reserved 300 ms before and after this time period to perform a TF decomposition using at least three β-band cycles. More importantly, the LCMV beamformer used for source construction was most sensitive to source activity within the limited time window of interest. Hence, given these constraints, we indeed had to mask out the low frequency areas concerned to avoid boundary effects. Please note that for replying to questions 10 and 11 of the reviewers (before the first revision), we extended the time-frequency window from the original manuscript (0–500 ms, 8–44 Hz) to –200–500 ms, 2–44 Hz. We attach the figure with the extended time-frequency window (and masked-out areas) to this reply (Author response image 6) and use the figure with the originally analyzed shorter time windows in the resubmitted manuscript. We don’t see any crucial additional information for the reader by presenting wider epochs, and the results described in the manuscript are not referring to these extended time frequency windows. However, Figure 3 in the resubmitted manuscript still contains the additional plots with spectral activity per condition, which had definitely improved data interpretation compared to the original manuscript.

Author response image 6
Β-band time-frequency representations of virtual channels at sources identified.

First and second solid line indicate SSRT and SSRTmax, first and second dotted line indicate 10% and 50%-percentiles of RTAC-GO for selected AC-GO trials with RTAC-GO > SSRT (RT AC-GO is the duration between AC-GO signal and button, see Figure 1 of the manuscript). Time axis locked to STOP and AC-GO signal (0s). All plots shows the results of a cluster-based permutation test, two-tailed, αcluster = 0.05, n=59, significant cluster are outlined. (A) Task-versus-baseline power for successful stop (sSTOP) trials. (B) Task-versus-baseline power for correct attentional capture go (cAC-GO trials). (C) Contrast of z-transformed successful stop (sSTOP) versus correct attentional capture go (cAC-GO) trials.

https://doi.org/10.7554/eLife.61679.sa2

Article and author information

Author details

  1. Michael Schaum

    Systemic Mechanisms of Resilience, Leibniz Institute for Resilience Research, Mainz, Germany
    Contribution
    Investigation, Formal analysis, Software, Data curation, Resources, Visualization, Writing - original draft, Writing - review and editing
    For correspondence
    Michael.Schaum@lir-mainz.de
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0002-6589-4530
  2. Edoardo Pinzuti

    Systemic Mechanisms of Resilience, Leibniz Institute for Resilience Research, Mainz, Germany
    Contribution
    Data curation, Software, Formal analysis, Writing - original draft, Writing - review and editing
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0002-5113-723X
  3. Alexandra Sebastian

    Department of Psychiatry and Psychotherapy, University Medical Center of the Johannes Gutenberg University, Mainz, Germany
    Contribution
    Conceptualization, Software, Formal analysis, Investigation, Visualization, Writing - original draft, Writing - review and editing
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0002-8381-8312
  4. Klaus Lieb

    Department of Psychiatry and Psychotherapy, University Medical Center of the Johannes Gutenberg University, Mainz, Germany
    Contribution
    Supervision, Funding acquisition, Writing - review and editing
    Competing interests
    No competing interests declared
  5. Pascal Fries

    1. Ernst Strüngmann Institute (ESI) for Neuroscience in Cooperation with Max Planck Society, Frankfurt, Germany
    2. Donders Institute for Brain, Cognition and Behaviour, Radboud University Nijmegen, Nijmegen, Netherlands
    Contribution
    Conceptualization, Resources, Supervision, Writing - review and editing
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0002-4270-1468
  6. Arian Mobascher

    Department of Psychiatry and Psychotherapy, University Medical Center of the Johannes Gutenberg University, Mainz, Germany
    Contribution
    Conceptualization, Writing - review and editing
    Competing interests
    No competing interests declared
  7. Patrick Jung

    Department of Psychiatry and Psychotherapy, University Medical Center of the Johannes Gutenberg University, Mainz, Germany
    Contribution
    Conceptualization, Investigation, Writing - review and editing, Supervision
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0003-1242-2844
  8. Michael Wibral

    Campus Institute Dynamics of Biological Networks, Georg-August University, Göttingen, Germany
    Contribution
    Supervision, Funding acquisition, Writing - review and editing
    Contributed equally with
    Oliver Tüscher
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0001-8010-5862
  9. Oliver Tüscher

    Department of Psychiatry and Psychotherapy, University Medical Center of the Johannes Gutenberg University, Mainz, Germany
    Contribution
    Conceptualization, Supervision, Writing - original draft, Writing - review and editing, Funding acquisition
    Contributed equally with
    Michael Wibral
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0002-4023-5301

Funding

Deutsche Forschungsgemeinschaft (SFB 1193)

  • Michael Schaum
  • Edoardo Pinzuti
  • Alexandra Sebastian

The funders had no role in study design, data collection and interpretation, or the decision to submit the work for publication.

Acknowledgements

This project was funded by German Research Foundation (CRC 1193, project C04, MW and OT). Parts of this research were conducted using the supercomputer Mogon and advisory services offered by Johannes Gutenberg University Mainz (hpc.uni-mainz.de), which is a member of the AHRP (Alliance for High Performance Computing in Rhineland Palatinate, http://www.ahrp.info) and the Gauss Alliance e.V. The authors gratefully acknowledge the computing time granted on the supercomputer Mogon at Johannes Gutenberg University Mainz.

Ethics

Human subjects: All individual participants included in the study provided written informed consent before participation and consent to publish any research findings based on their provided data in anonymized form. The study was approved by the local ethics committees (Johann Wolfgang Goethe University, Frankfurt, Germany, and Medical Board of Rhineland-Palatinate, Mainz, Germany, IRB Protocol no. 837.128.11), and participants were financially compensated for their time.

Senior Editor

  1. Richard B Ivry, University of California, Berkeley, United States

Reviewing Editor

  1. Simon Little, UCSF, United States

Reviewer

  1. Simon Little, UCSF, United States

Publication history

  1. Received: July 31, 2020
  2. Accepted: March 23, 2021
  3. Accepted Manuscript published: March 23, 2021 (version 1)
  4. Version of Record published: May 4, 2021 (version 2)

Copyright

© 2021, Schaum et al.

This article is distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use and redistribution provided that the original author and source are credited.

Metrics

  • 883
    Page views
  • 171
    Downloads
  • 2
    Citations

Article citation count generated by polling the highest count across the following sources: Crossref, PubMed Central, Scopus.

Download links

A two-part list of links to download the article, or parts of the article, in various formats.

Downloads (link to download the article as PDF)

Download citations (links to download the citations from this article in formats compatible with various reference manager tools)

Open citations (links to open the citations from this article in various online reference manager services)

Further reading

    1. Neuroscience
    Sahana Sitaraman et al.
    Research Article

    Gap junctions between neurons serve as electrical synapses, in addition to conducting metabolites and signaling molecules. During development, early-appearing gap junctions are thought to prefigure chemical synapses, which appear much later. We present evidence for this idea at a central, glutamatergic synapse and provide some mechanistic insights. Loss or reduction in the levels of the gap junction protein Gjd2b decreased the frequency of glutamatergic miniature excitatory postsynaptic currents (mEPSCs) in cerebellar Purkinje neurons (PNs) in larval zebrafish. Ultrastructural analysis in the molecular layer showed decreased synapse density. Further, mEPSCs had faster kinetics and larger amplitudes in mutant PNs, consistent with their stunted dendritic arbors. Time-lapse microscopy in wild type and mutant PNs reveals that Gjd2b puncta promote the elongation of branches and that CaMKII may be a critical mediator of this process. These results demonstrate that Gjd2b-mediated gap junctions regulate glutamatergic synapse formation and dendritic elaboration in PNs.

    1. Neuroscience
    Amicia D Elliott et al.
    Research Article Updated

    Identifying neural substrates of behavior requires defining actions in terms that map onto brain activity. Brain and muscle activity naturally correlate via the output of motor neurons, but apart from simple movements it has been difficult to define behavior in terms of muscle contractions. By mapping the musculature of the pupal fruit fly and comprehensively imaging muscle activation at single-cell resolution, we here describe a multiphasic behavioral sequence in Drosophila. Our characterization identifies a previously undescribed behavioral phase and permits extraction of major movements by a convolutional neural network. We deconstruct movements into a syllabary of co-active muscles and identify specific syllables that are sensitive to neuromodulatory manipulations. We find that muscle activity shows considerable variability, with sequential increases in stereotypy dependent upon neuromodulation. Our work provides a platform for studying whole-animal behavior, quantifying its variability across multiple spatiotemporal scales, and analyzing its neuromodulatory regulation at cellular resolution.