High precision magnetoencephalography reveals increased right-inferior frontal gyrus beta power during response conflict

Flexibility of behavior and the ability to rapidly switch actions is critical for adaptive living in humans. It is well established that the right-inferior frontal gyrus (R-IFG) is recruited during outright action-stopping, relating to increased beta (12-30 Hz) power. Additionally, pre-supplementary motor area (pre-SMA) is plausibly recruited during response conflict/switching, relating to increased theta (4-8 Hz) power. It has been posited that inhibiting incorrect response tendencies is central to motor flexibility. However, it is not known if the commonly reported R-IFG beta signature of response inhibition in action-stopping is also recruited during response conflict, which would suggest overlapping networks for stopping and switching. In the current study, we analyzed high precision magnetoencephalography (hpMEG) data recorded with very high trial numbers (total n > 10,000) from 8 subjects during different levels of response conflict. We hypothesized that a R-IFG-triggered network for response inhibition is domain general and also involved in mediating response conflict. We therefore tested whether R-IFG showed increased beta power dependent on the level of response conflict. We also hypothesized that pre-SMA is an important node in response conflict processing, and tested whether pre-SMA theta power increased for response conflict trials. Using event-related spectral perturbations and linear mixed modeling, we found that both R-IFG beta and pre-SMA theta increased for response conflict trials, with the R-IFG beta increase specific to trials with strong response conflict. This result supports a more generalized role for R-IFG beta in response inhibition, beyond simple stopping behavior towards response switching. Significance Statement Response inhibition is a core component of cognitive control. Neural mechanisms of response inhibition are typically studied using stopping paradigms. However, there is an unresolved debate regarding whether the response inhibition network is specific to stopping or generalizes to switching between tasks and overcoming conflict between competing response tendencies. Increased beta (12-30 Hz) in R-IFG has historically been interpreted as a marker of successful response inhibition in the stop-signal task. Here, we investigated the presence of this electrophysiological marker of response inhibition specifically during response conflict (switching). We found R-IFG beta power increased for trials with strong response conflict, and not for weak or no response conflict, thereby supporting a generalized role for R-IFG beta in response inhibition and switching.

supplementary motor area (pre-SMA) is plausibly recruited during response conflict/switching, 23 relating to increased theta (4-8 Hz) power. It has been posited that inhibiting incorrect response 24 tendencies is central to motor flexibility. However, it is not known if the commonly reported R-25 IFG beta signature of response inhibition in action-stopping is also recruited during response 26 conflict, which would suggest overlapping networks for stopping and switching. In the current 27 study, we analyzed high precision magnetoencephalography (hpMEG) data recorded with very 28 high trial numbers (total n > 10,000) from 8 subjects during different levels of response conflict. 29 We hypothesized that a R-IFG-triggered network for response inhibition is domain general and 30 also involved in mediating response conflict. We therefore tested whether R-IFG showed 31 increased beta power dependent on the level of response conflict. We also hypothesized that 32 pre-SMA is an important node in response conflict processing, and tested whether pre-SMA theta 33 power increased for response conflict trials. Using event-related spectral perturbations and linear 34 mixed modeling, we found that both R-IFG beta and pre-SMA theta increased for response 35 conflict trials, with the R-IFG beta increase specific to trials with strong response conflict. This 36 result supports a more generalized role for R-IFG beta in response inhibition, beyond simple 37 stopping behavior towards response switching. 38 Introduction 50 Sometimes we plan or start to execute an action, but then need to suddenly execute a 51 different action instead. In experiments, this has been called switching, response overriding, or 52 overcoming response conflict; here we use the term response conflict. One notable theory 53 suggests that during motoric response conflict, inhibition of the incorrect response tendency is 54 necessary (see Wiecki & Frank, 2013). This inhibitory control mechanism may then be the same 55 as that of outright action-stopping (e.g., Wessel

Time-frequency decomposition 143
We computed time-frequency transforms of the source-level time series using Morlet 144 wavelets (3 cycles at low frequencies, linearly increasing by 0.5 at higher frequencies), with a 145 range from 4-30 Hz (see Jana et al., 2020). We performed time-frequency transforms for the time 146 series for every vertex in each ROI cluster, to obtain a time x frequency x trial x vertex, 4-147 dimensional matrix of power for each ROI in each session. Then, we averaged across the cluster 148 vertex dimension to create a time x frequency x trial matrix. 149

Event-related spectral perturbation (ERSP) 150
For group level visualizations, we computed ERSPs for trials with and without response 151 conflict for each ROI (i.e., pre-SMA, R-IFG, L-IFG). We converted spectral power to decibels (dB) 152 using a 500 ms baseline prior to the RDK presentation (i.e., during fixation) (see Cohen, 2014). 153 We averaged across all sessions that each subject completed (total N = 24), and then computed 154 a grand average across subjects (N = 8). We subtracted the group average ERSP for trials with no 155 response conflict from the group average ERSP for trials with response conflict to specifically 156 visualize the difference. We plotted these ERSPs with a range from 4-30 Hz on the y-axis and 0-157 500 ms relative to the imperative cue on the x-axis, with a vertical line at 300 ms denoting average 158 RT (Figure 2). We followed the same procedure for R-IFG ERSPs split by levels of coherence 159 (Figure 3). 160

Data preparation for linear mixed modeling 161
To prepare the spectral power data for single-trial level statistical analyses, we computed 162 a simple linear baseline subtraction using a 500 ms window prior to the RDK presentation on that 163 trial (see Grandchamp & Delorme, 2011;Cohen, 2014). We a priori (prior to visualization of the 164 contrast spectrograms) defined time and frequency ranges to average across to obtain single-165 trial power estimates. We used 0-300 ms relative to the imperative cue (i.e., time between 166 imperative cue presentation and average RT). We used 4-8 Hz for theta, and subject-specific and 167 conventional partitions of 12-30 Hz for beta (explained in detail below). We z-scored each of 168 these single-trial power estimates within-subject and -ROI. primary test an evaluation of whether R-IFG low beta specifically was recruited during response 175 conflict. We used an individualized (peak-centered) frequency band of task-relevant low beta to 176 test our primary hypothesis about R-IFG low beta in response conflict. We computed average 177 baseline-corrected spectral power (dB) for all trials that the subject completed, and then 11 averaged across our 0-300 ms time window of interest to obtain a single spectral power estimate 179 for each frequency value in the 12-20 Hz range. Then we extracted the value for which low beta 180 power was greatest, and used that as the center of a narrow subject-specific range (peak low 181 beta +/-1 Hz). In addition to defining subject-specific low beta, we also used a broader band 182 conventional range of low beta (12-20 Hz) and used the same procedure described here to define 183 subject-specific (peak high beta +/-1 Hz) and conventional (21-30 Hz) high beta for secondary 184 comparison. 185

Experimental design and statistical analyses 186
Linear mixed modeling 187 To take advantage of the high trial count (n = 10,496) in this dataset, and model within-188 subject data as well as small cohort between-subject data (N = 8), we used a linear mixed 189 modeling framework using R (v3.6.1, R Core Team, 2019) and the lme4 package (v1.1-21, Bates 190 et al., 2015). All of our models had session nested within subject-specific intercepts as random 191 effects, and the models exploring relationships between neural activity and behavioral 192 performance also had session/subject-specific slopes by response conflict. We used Type III Wald 193 Chi Square tests for our models. We used z-tests, Tukey corrected for multiple comparisons, for 194 pairwise comparisons. 195 To assess the impact of response conflict and RDK coherence on RT, we used a linear 196 mixed model with RT (log10 transformed) as the dependent variable, and response conflict, RDK 197 coherence, and their interaction as fixed effects. To assess the impact of response conflict and 12 RDK coherence on error, we used a generalized linear mixed model with a binomial distribution, 199 response (0 = incorrect, 1 = correct) as the dependent variable, and response conflict, RDK 200 coherence, and their interaction as fixed effects. To assess the impact of response conflict and 201 RDK coherence on neural activity, we used linear mixed models with z-scored power as the 202 dependent variable, and response conflict, RDK coherence, and their interaction as fixed effects. 203 Lastly, to explore relationships between neural activity and behavioral performance, we used 204 linear mixed models with RT (log10 transformed) as the dependent variable and z-scored power 205 as a fixed effect, and generalized linear mixed models with binomial distributions, response (0 = 206 incorrect, 1 = correct) as the dependent variable, and z-scored power as fixed effects. 207 Results 208

Behavioral results 209
On average, subjects responded faster and more accurately on trials with no response 210 conflict compared to trials with response conflict (Figure 1). Linear mixed models revealed a 211 significant interaction between response conflict and coherence for log-transformed RT (X 2 (2) = 212 54.71, p < .0001) and for error (X 2 (2) = 264.19, p < .0001). Pairwise comparisons revealed 213 significant differences in behavioral performance for trials with no response conflict compared 214 to trials with response conflict. Subjects had significantly longer RTs on trials with response 215 conflict for high (Z = -19.89, p < .0001), medium (Z = -16.07, p < .0001) and low (Z = -9.71, p < 216 .0001) coherence trials, and significantly more errors on trials with response conflict for high (Z 217 = 26.94, p < .0001), medium (Z = 20.64, p < .0001), and low (Z = 4.81, p < .0001) coherence trials. 218 To test for previously described theta power increases in pre-SMA during response 231 conflict, we analyzed pre-SMA theta power during response conflict trials compared to no 232 response conflict trials. We also analyzed theta power in R-IFG and L-IFG to test for regional 233 specificity of any theta increases. In keeping with the previous literature, theta power in pre-SMA 234 after the presentation of the imperative cue was higher during response conflict trials compared 235 to no response conflict trials, though the effect size of this increase was modest ( Figure 2B). Our 236 linear mixed model revealed a trend of a main effect of response conflict on pre-SMA theta power 237 (X 2 (1) = 2.82, p = .093). Additionally, theta power in R-IFG and L-IFG after the imperative cue was 238 higher on average during response conflict trials compared to no response conflict trials (Figures 239 2C and 2D, respectively). Our linear mixed models revealed a significant main effect of response 240 conflict on R-IFG theta power (X 2 (1) = 12.07, p = <.0001) and a trend of a main effect of response 241 conflict on L-IFG theta power (X 2 (1) = 3.16, p = .075). 242

Increased beta power during response conflict in R-IFG, not pre-SMA or L-IFG 243
To test whether there was increased low beta power in R-IFG during response conflict, 244 we analyzed subject-specific low beta power during response conflict trials compared to no 245 response conflict trials in R-IFG. We also analyzed conventional low beta, and both subject-246 15 specific and conventional high beta for comparison. We then analyzed these definitions of beta 247 power in pre-SMA and L-IFG to test for regional specificity of any beta increases. On average, low 248 beta (not high beta) power in R-IFG was higher during response conflict trials compared to no 249 response conflict trials (Figure 2C). Our linear mixed models revealed a significant main effect of 250 response conflict on subject-specific low beta power in R-IFG (X 2 (1) = 4.25, p = .039). Our linear 251 mixed models also revealed a significant main effect of response conflict on conventional low 252 beta power in R-IFG (X 2 (1) = 16.17, p < .0001), and no significant main effects of response conflict 253 on conventional nor subject-specific definitions of high beta power in R-IFG. 254 On average, low beta power in pre-SMA also appeared to be slightly higher during 255 response conflict trials compared to no response conflict trials ( Figure 2B) Single-trial analyses revealed significant main effects of response conflict on theta, conventional 273 low beta, and subject-specific low beta power. D) L-IFG shows increased theta power (not beta 274 power) prior to the average RT for response conflict trials compared to no response conflict trials. 275 Single-trial analyses revealed a significant main effect of response conflict on theta power, and 276 no significant main effects of response conflict on conventional nor subject-specific low beta 277 power. 278

R-IFG beta power increased for stronger response conflict trials 279
To test whether the recruitment of R-IFG low beta during response conflict depended on 280 the strength of the response conflict, we looked at the interaction between response conflict and 281 RDK coherence (operationalized as modulating the strength of response conflict on incongruent 282 trials) on subject-specific low beta power, and on conventional low beta power as well. On 283 average, low beta power in R-IFG was higher during response conflict trials compared to no 284 response conflict trials for high coherence trials ( Figure 3B) and for medium coherence trials 285 ( Figure 3C), and not for low coherence trials ( Figure 3D). Our linear mixed models revealed a 286 significant interaction between response conflict and coherence on subject-specific low beta 287 power in R-IFG (X 2 (2) = 6.76, p = .034), and not on conventional low beta power. Pairwise 288 comparisons revealed a significant difference in R-IFG subject-specific low beta power for high 289 coherence (i.e., strong) response conflict trials compared to high coherence no response conflict 290 trials (Z = -4.02, p = .0001), and not for within medium nor low coherence trials. were performed at the single-trial level; these group-level plots are solely for visualization. A) 296 Example source localization for R-IFG from one subject. B-D) R-IFG shows increased low beta 297 power prior to the average RT for response conflict trials compared to no response conflict 298 trials for high coherence (B) and medium coherence (C) trials, and not for low coherence (D) 299 trials. Single-trial analyses revealed a significant interaction between response conflict and 300 coherence on subject-specific low beta power. Pairwise comparisons revealed a significant 301 difference in subject-specific low beta power for response conflict trials compared to no 302 response conflict trials for high coherence trials only, not for medium nor low coherence trials. 303

Impact of pre-SMA theta power on RT and error 305
To test whether the increase in pre-SMA theta during response conflict trials was 306 correlated with behavior, we looked at the main effect of pre-SMA theta power on log-307 transformed RT and response error. We also looked at whether the increases in R-IFG and L-IFG 308 theta during response conflict related to behavioral performance. Our linear mixed models 309 revealed a significant main effect of pre-SMA theta power on RT (X 2 (1) = 5.64, p = .018), and no 310 significant main effects of R-IFG nor L-IFG theta power on RT. Additionally, our generalized linear 311 mixed models revealed a significant main effect of pre-SMA theta power on error (X 2 (1) = 8.12, p 312 = .0044), as well as significant main effects of R-IFG (X 2 (1) = 12.39, p = .00043) and L-IFG (X 2 (1) = 313 11.89, p = .00057) theta power on error. 314

Impact of R-IFG low beta power on error for response conflict trials only 315
To test whether the increase in R-IFG low beta during response conflict trials was 316 correlated with behavior, we first looked at the main effect of R-IFG low beta power on log-317 transformed RT and on response error, and did the same for L-IFG for comparison. Our linear 318 mixed models revealed no significant main effects of R-IFG nor L-IFG subject-specific nor 319 conventional definitions of low beta power on RT nor error. It is plausible that the relationship 320 between R-IFG low beta power and behavior may only be meaningful on trials with response 321 conflict, if mechanisms of response inhibition are only supporting behavioral performance when 322 there is a need to inhibit an incorrect response tendency. Therefore, we conducted an 323 20 exploratory analysis restricted to trials with response conflict. Our linear mixed models revealed 324 a trend of a main effect of R-IFG subject-specific low beta power on error (X 2 (1) = 3.72, p = .054), 325 and no significant main effects of L-IFG subject-specific nor conventional beta power on error. 326

Discussion 327
We found support for two theoretically driven, a priori hypotheses about the recruitment 328 of pre-SMA theta and R-IFG low beta during response conflict. We hypothesized that pre-SMA is 329 a critical region for overcoming response conflict, supported by findings in humans and non- aligns well with a framework implicating theta communication from pre-SMA to STN to M1 to 335 pause motor output until sufficient evidence accumulates to a decision threshold during conflict 336 (see Aron et al., 2016). In the current study we found that, after the imperative cue and prior to 337 the average RT, pre-SMA theta power was increased for trials with response conflict compared 338 to trials with no response conflict, and this increase related to behavioral performance in terms 339 of RT and accuracy in responding. However, the effect size of the theta power increase was 340 modest, with the linear mixed models revealing a trend toward a significant result. We also found 341 that theta increased in R-IFG (significant) and L-IFG (trend towards significance) during response 342 conflict. These latter findings indicate a lack of regional specificity for theta activity which we did 343 not initially predict. This could reflect the high SNR and very high trial number in our data, and 344 21 could also suggest a broader recruitment of cortical theta during response conflict than has 345 previously been reported. 346 We also hypothesized that response switching is a generalized form of stopping and 347 therefore would recruit the previously defined R-IFG beta-triggered inhibitory control network 348 during response conflict. A small number of studies have directly implicated R-IFG activity using Importantly, we found here that low beta power in R-IFG was significantly increased on 364 trials with response conflict compared to trials without response conflict. Notably, we did not see 365 22 any significant increases in beta power in L-IFG. This is a strong comparison region for R-IFG 366 because it is an anatomically matched prefrontal region, but one that is not hypothesized to be 367 a node in the putative inhibitory control network. Another compelling component of this result 368 is that this R-IFG low beta increase during response conflict was specific to higher coherence 369 trials: R-IFG beta power significantly increased for high coherence response conflict trials, 370 quantitatively increased for medium coherence response conflict trials but was not significant, 371 and did not increase for low coherence response conflict trials. We predicted that punctate 372 response inhibition (plausibly via R-IFG beta) might be necessary when the incorrect response 373 tendency is most prepotent (i.e., during strong conflict trials), and our results support this idea. 374 Additionally, we further explored the potential role of R-IFG beta in overcoming response conflict 375 by conducting an exploratory analysis of the main effect of R-IFG low beta power on error for 376 response conflict trials only, and found a trend towards significance. Taken together, these 377 results support a role for R-IFG beta in overcoming response conflict. 378 Some limitations are worth addressing. We analyzed hpMEG data collected from a small 379 cohort of 8 subjects. Although this is a relatively low number of human subjects, the hpMEG data 380 had a very high SNR and was recorded across a very large number of trials (n = 10,496), similar to 381 primate electrophysiology. Therefore, we were able to perform statistically powerful within-382 subject analyses that accounted for across subject factors using mixed modeling. This approach 383 was also supported by strong a priori hypotheses that were directly tested here. We also want to 384 be cautious of reverse inference in our interpretation that the presence of R-IFG beta plausibly 385 reflects response inhibition during response conflict. We conclude here that the recruitment of 386 23 This is also supported by the established role of R-IFG beta during action stopping and 388 computational and empirical work suggesting that response inhibition is important for 389 overcoming response conflict (Forstmann et  R-IFG beta as a marker of successful response inhibition in stopping above and beyond 400 differences in attention by using the stop-signal task with attentional-capture trials (see Schaum 401 et al., 2021). 402 In conclusion, we analyzed hpMEG data with high spatial and temporal resolution 403 recorded during a response conflict paradigm, and found support for two theoretically driven 404 hypotheses about the recruitment of pre-SMA theta and R-IFG beta during response conflict. In 405 a novel result, we showed that R-IFG low beta power was significantly increased for response 406 conflict trials, and specifically strong response conflict trials which plausibly require mechanisms 407 of punctate response inhibition for correct responding. It is plausible that beta-band 408 communication from R-IFG to STN to M1 is recruited to inhibit incorrect response tendencies 409 24 during response conflict. Future work using causal methods such as TMS or neurofeedback can 410 further establish the role of R-IFG beta power in response conflict resolution. Additionally, other 411 response conflict paradigms, particularly those with high ecological validity, can further 412 investigate the relationship between R-IFG beta activity and behavior, as our marginally 413 significant exploratory result suggests that increased R-IFG beta during response conflict trials 414 may relate to more accurate responding. Overall, our results support R-IFG beta as a neural 415 mechanism of overcoming response conflict, in addition to action-stopping. This broadens the 416 role for R-IFG beta as a domain general inhibitory control signal, which may have clinical 417 implications for populations with inhibitory control deficits. 418