Alpha oscillations protect working memory against distracters in a modality-specific way

Alpha oscillations are thought to be involved in suppressing distracting input in working-memory tasks. Yet, the spatial-temporal dynamics of such suppression remain unclear. Key questions are whether such suppression reflects a domain-general inattentiveness mechanism, or occurs in a stimulus-or modality-specific manner within cortical areas most responsive to the distracters; and whether the suppression is proactive (i.e., preparatory) or reactive. Here, we addressed these questions using a working-memory task where participants had to memorize an array of visually presented digits and reproduce one of them upon being probed. We manipulated the presence of distracters and the sensory modality in which distracters were presented during memory maintenance. Our results show that sensory areas most responsive to visual and auditory distracters exhibited stronger alpha power increase after visual and auditory distracter presentation respectively. These results suggest that alpha oscillations underlie distracter suppression in a reactive, modality-specific manner.


Introduction
Our sensory systems are constantly bombarded with information.
Being able to protect relevant information from distraction is a cornerstone for adaptive behavior. At the implementation level, neural oscillations are thought to reflect low-level operations, providing scaffolds for information processing (X.-J. Wang, 2010).
In a seminal paper by Bonnefond & Jensen (2012), human participants had to remember a set of four consonants presented sequentially on a screen, and indicate whether a probe consonant was part of the memory set. In different blocks, while participants maintained the consonant set in memory, a distracting symbol or consonant (corresponding to a weak or strong distracter, respectively) was shown. The authors found that occipital alpha activity increased before the anticipated distracter onset, and that this increase was stronger in anticipation of stronger distracters, suggesting that alpha activity is involved in distracter suppression, consistent with the inhibition account of alpha oscillations (Haegens et al., 2011;Jensen & Mazaheri, 2010;Klimesch, 2012;Klimesch et al., 2007). Yet, the spatial and temporal specificity of such alpha activity modulation in anticipation of task-irrelevant input (i. e., distracters) in a working-memory context remains uncharted. One may hypothesize that the observed alpha power increase in anticipation of distracters is domain-general, reflecting an increase of overall inattentiveness or disengagement of bottom-up stimulation. In other words, the observed alpha activity modulation is not spatially-specific to brain areas specialized in representing the distracters. Fluctuations of alpha activity can then be seen as reflecting the push-pull between "outer-directed" attention and "inner-directed" attention (Cooper et al., 2006;Shaw, 1996). It is thus conceivable that alpha power increase due to the presence of distracters during working-memory maintenance may simply reflect an endogenous shift from an outer-directed state to an inner-directed state. Irrespective of what the external stimulation is (e. g., whether it is an auditory or visual distracter), such a shift is then expected to be accompanied by overall alpha activity increase.
In contrast, the observed alpha power increase in anticipation of distracters may reflect a stimulus-specific/modality-specific inhibition of distracting information. This hypothesis echoes findings from multisensory attention-orienting studies, where participants were cued to attend to stimuli delivered via one sensory modality while ignoring those delivered via other sensory modalities. It has been shown that attention-related alpha activity (i.e., activity during the cue-to-target period) operates in a modality-specific manner-alpha power increase is observed specifically within the task-irrelevant sensory regions (Friese et al., 2016;Keller et al., 2017;Mazaheri et al., 2014), which consequently facilitates selection and prioritization of task-relevant stimuli. Yet, it is unclear whether the suppression of distracting information during working-memory maintenance follows a similar mechanism.
In addition to questions regarding spatial specificity of alpha modulation related to distraction, its temporal specificity is also largely unclear. The critical question in this line of research is whether participants endogenously employ such a mechanism in a proactive, anticipatory way. Previous work has consistently demonstrated that the brain uses predictive information during information processing (Nobre & van Ede, 2018;Summerfield & de Lange, 2014), mostly in a proactive manner (e.g., Ekman et al., 2017;Kok et al., 2017;Mayer et al., 2016). Hence, it is intuitive to expect distracter suppression based on predictive information to be proactive. Curiously, however, recent work on anticipatory distracter suppression has produced mixed results: while some authors reported that posterior alpha activity reflected proactive and preparatory inhibition of predictable visual distracters (van Moorselaar et al., 2020;B. Wang et al., 2019), others reported null results (Noonan et al., 2016;van Moorselaar & Slagter, 2019).
Here, we set out to characterize the role of alpha activity in a working-memory task with predictable distracters presented via visual and auditory stimulation. Using magnetoencephalography (MEG), we are able to track brain activity of interest with sufficient spatio-temporal resolution and test whether alpha activity modulations in response to and/or in anticipation of distracters are domain-general or modalityspecific (Fig. 1B). To preview, our results show that, compared to the no-distracter condition, distracters (both visual and auditory) presented during working-memory maintenance interfere with the memorized items, and result in modality-specific alpha activity increase in sensory regions most responsive to the distracters during memory retrieval.

Data availability
All data and code for stimulus presentation and analysis are available online at the Donders Repository (https://data.donders.ru.nl) at https ://doi.org/10.34973/f1ct-ax21.

Participants
Twenty-six healthy volunteers completed the full experiment. One of them was excluded from analysis because of poor behavioral performance (i.e., overall accuracy lower than 30%), resulting in 25 participants in the reported analysis (mean age = 25.8, SD = 5.61; 17 female). The study was approved by the local ethics committee (CMO Arnhem-Nijmegen). All participants gave informed consent prior to the experiment and received monetary compensation for their participation. Participants who did not already have their T1-weighted anatomical scans available in our institute's database, were invited for a second session, during which we obtained their MRI scans.

Experimental Task
Participants' task was to memorize a visually presented array of six digits, and report the digit at the probed spatial location in each trial. They were instructed to be as accurate as possible while also being fast, as the response window was limited (max 4 s). We used three trial conditions: the auditory distracters trial, where distracting digits were consecutively played binaurally during memory retention; the visual distracters trial, where distracting digits were consecutively displayed during memory retention; and the no distracters trial, where no distracting digits were presented during the memory retention window (matched in length with the other conditions). Different trials were intermixed within each block, and a cue at the start of each trial informed the subject about the trial condition. Each subject completed nine blocks of 40 trials, during which their brain activity was recorded Fig. 1. (A) Trial schematic. This example shows a trial with visual distracters. Stimuli are rescaled for illustration purposes. Time points denote event onsets relative to the cue. (B) Expected alpha power modulation for the domain-general (left) and the modality-specific (right) hypotheses. If the modulation is domain-general, alpha power in the respective sensory regions of interest (ROI) should not differ between the auditory (green) and visual (purple) distracter conditions. If the modulation is modality-specific, alpha power in the auditory ROI should be higher under the auditory distracter condition, while alpha power in the visual ROI should be higher under the visual distracter condition. (C) Behavioral results for auditory distracter (Aud), visual distracter (Vis), and no distracter conditions (No). Left panel: overall accuracy; right panel: percentage of trials with a swapping error. Dots denote individual data points. Shaded areas around the mean denote withinsubject standard errors. non-invasively using magnetoencephalography (MEG).

Stimuli
A bull's eye (outer black ring = 0.5 × 0.5 degree of visual angle (dva), innermost black dot = 0.25 × 0.25 dva) was presented at the center of the screen as the fixation point. Participants were instructed to always maintain fixation, and not to blink during the presentation of the stimuli. Each trial (Fig. 1A) started with a 400-ms cue, informing the subject about the type of upcoming distracters in the current trial (letters "A", "V", and "O" for auditory, visual, and no distracters, respectively). After a cue-to-target delay of 500 ms, the target array consisting of six digits was presented for 800 ms. Each digit was drawn in font size 72 (approximately 1.6 dva), at an eccentricity of 7 dva around central fixation. For trials of the visual and auditory distracter conditions, six visual or auditory distracters respectively were presented after a 800-ms target-to-distracters delay. Among the six distracter digits in each trial, four of them were not shown in the target array, with the remaining two present in the target array. For example, if digits 2, 8, 5, 7, 4, and 9 were used to make the target array, then non-target digits 1, 3, 6, and 0, together with a random selection of two target digits (e.g., 2 and 8), were used as distracters. Each visual distracter (font size 172, approximately 3.8 dva) was centrally displayed for 350 ms with a 150-ms inter-stimulus interval (ISI), and each auditory distracter was binaurally displayed for 350 ms with an ISI of 150 ms, during which the central fixation point remained on the screen. The consecutive presentation of distracters was then followed by a 800-ms distracters-to-probe delay, after which the probe array highlighting the spatial location of interest (marked with a circle) was shown. At the end of each trial, the subject reported the digit at the probed spatial location by rotating the hand of a number dial. The trial ended once the subject had locked their response, or a maximum of 4-s response time was reached.

Data acquisition
Stimuli were displayed on a semi-translucent screen (1920 × 1080pixel resolution, 120-Hz refresh rate) back-projected by a PROpixx projector (VPixx Technologies) during MEG recordings. The experiment was programmed with Psychtoolbox (Brainard, 1997) in Matlab (The Mathworks, Inc.) and ran in a Windows environment. Brain activity was recorded using a 275-channel axial gradiometer MEG system (CTF MEG Systems, VSM MedTech Ltd) at 1200 Hz in a magnetically shielded room. Six permanently faulty channels were disabled during the recordings, leaving 269 recorded MEG channels. Three fiducial coils were placed at the subject's nasion and both ear canals, to provide online monitoring of the subject's head position (Stolk et al., 2013), and to serve as anatomical landmarks for offline co-registration with structural MRI scans. An infrared eye tracker (EyeLink, SR Research Ltd., Mississauga, Ontario, Canada) was used to monitor the subject's eye position during the MEG recordings.
We used a Polhemus 3D tracking device (Polhemus, Colchester, Vermont, United States) to record and digitize the subject's head shape and the location of the three fiducial coils. T1-weighted anatomical scans were acquired with a 3T MRI system (Siemens, Erlangen, Germany). We placed earplugs with a drop of vitamin E into the subject's ear canals during MRI acquisition, to facilitate co-registration with MEG data.

Behavioral data analysis
The influence of condition (three levels: auditory, visual, and no distracters) on median reaction times (RT; including correct response only), accuracy (percentage of correct responses), and swapping error (percentage of trials where participants reported one of the four nontarget digits) was tested using repeated-measures ANOVA. In addition, we computed partial eta squared (partial η 2 ) following Equation 10 by (Lakens, 2013) as effect size measures.

MEG data analysis 2.7.1. MEG preprocessing
MEG data were preprocessed offline and analyzed using the Field-Trip toolbox (Oostenveld et al., 2011) and custom-built Matlab scripts. The MEG signal was epoched based on the onset of the cue (referred to as time zero hereafter). All data were down-sampled to 300 Hz, after applying a notch filter to remove line noise and harmonics (at 50, 100, and 150 Hz). Trials with excessive noise were rejected via visual inspection before independent component analysis (ICA). ICA components were visually inspected and those representing eye and heart artifacts were then projected out of the data (Jung et al., 2000). Finally, outlier trials with extreme variance were removed via visual inspection, leaving on average 350 trials (out of 360 trials, between-subjects SD = 8 trials) per subject for the reported analyses.

MRI processing
MRI data were co-registered to the CTF coordinate system using the fiducial coils and the digitized scalp surface. Volume conduction models were constructed based on single-shell models (Nolte, 2003) of individual participant's anatomical MRIs or of the template brain provided by the FieldTrip toolbox when the subject-specific MRIs were unavailable (for 5 participants). Dipole positions were defined using a cortical surface-based mesh with 15,784 vertices created using Freesurfer v6.0 (RRID: SCR_001847) and HCP workbench v1.3.2 (RRID: SCR_008750). The vertices were grouped into 374 parcels based on a refined version of the Conte69 atlas (Van Essen et al., 2012), allowing us to reduce the dimensionality of the data (similar to Schoffelen et al., 2017). For each dipole position, lead fields were computed with a reduced rank, which accommodates the fact that MEG is less sensitive to radial sources.

Sensor-level spectral analysis
Time-frequency representations (TFRs) of power were calculated for each trial by applying a fast Fourier transform to short sliding time windows. We applied Hanning tapers of 5-cycles length in time steps of 50 ms to single-trial data, prior to computing the spectra (4-30 Hz). Spectral decomposition was applied to synthetic planar gradient data, and combined into single spectra per sensor.

Source reconstruction of time series data
The linearly constrained maximum variance (LCMV) beamformer approach (Van Veen et al., 1997) was used to obtain the source reconstruction of the event-related response to the stimuli. The data covariance matrix was computed over a window of [-1.0, 6.95] s for all trials, time-locked to cue onset, and was subsequently used to construct common spatial filters. We projected the trial data through the resulting spatial filters to estimate the amplitude of the single-trial stimulus evoked response. Neural response time series of each anatomical parcel were computed by taking the average time series across all dipole positions within the parcel, and they were used as input to the Fourier analysis to obtain parcel-level power spectra.
To estimate the spatial distribution of the evoked response to the distracters, we took the sensor-level time series data between distracters onset and offset (time window = [2.5, 5.35] s) and projected these through the constructed common spatial filters. This was done for each condition separately, and the resulting condition-specific spatial distribution of the evoked response was parcellated by taking the mean across all dipole positions within a parcel. To obtain sensory responses specific to the visual (and auditory) distracters, we further normalized the resulting (parcellated) evoked responses corresponding to the visual and auditory distracter conditions with that of the no-distracter condition, where Evoked X denotes the evoked response of the distracter condition of interest.
Y.J. Zhou et al. Sensory regions of interest most responsive to visual and auditory distracters were identified by searching for parcels exhibiting the highest evoked response to the auditory and visual distracters respectively. This was done by contrasting the response amplitude to the first visual (or auditory) distracter with that of the pre-cue baseline window, for all visual (or auditory) distracter trials. One parcel was identified per hemisphere as the visual (or auditory) ROI for each individual participant. While this sometimes resulted in different parcels selected from left and right hemispheres (i.e., not perfectly symmetrical parcels across hemispheres), we verified that enforcing symmetrical parcel selection yields highly similar results (not reported).

Source reconstruction of spectral data
We used the partial canonical coherence beamformer approach (Schoffelen et al., 2008) to localize the sources of oscillatory alpha-band activity in response to the main task trials. To estimate the spatial distribution of alpha-band power, we extracted data segments centered at the time point of interest, then computed cross-spectral density (CSD) matrices using Slepian tapers (Mitra & Pesaran, 1999) centered at a frequency of 10 (±4) Hz. With the CSD matrices and the lead fields, a common spatial filter (for all trials across conditions) was constructed for each dipole position for each subject. By projecting sensor-level CSD through the common spatial filter, the spatial distribution of power was then estimated for each trial. To reduce data dimensionality, we took the mean estimated power of all dipoles within each parcel as the single-trial power estimate of the parcel.
To account for intra-and inter-individual variability in alpha peak frequency (Haegens et al., 2014), we defined the individual alpha peak frequency for each brain area (i.e., anatomically-defined parcel, see MRI preprocessing) and each subject separately using the FOOOF toolbox (Donoghue et al., 2020). Specifically, we applied the FOOOF algorithm to the pre-stimulus 1-s time series computed for each parcel. This time window was chosen as we did not expect any activity modulation by experimental condition before cue onset. Settings for the algorithm were set as: peakwidth limits: [0.2, 12]; max number of peaks: 3; minimum peak height: 0.3; peak threshold: 2; and aperiodic mode: 'fixed'. Power spectra were parameterized across the frequency range of 2 to 30 Hz. This time window was chosen as we would not expect any pre-stimulus modulation by experimental conditions. Based on the algorithm's output (i.e., the fitted aperiodic and periodic/oscillatory components), we defined the individual alpha peak as the oscillatory component's extracted peak frequency that falls within the 8-13 Hz band, or set it to 10 Hz when no clear oscillatory component was identified within the above-mentioned band. In rare cases where more than one extracted peak frequencies fell within the 8-13 Hz band, the lower frequency was used. Unless otherwise specified, the reported region of interest (ROI) based analyses comparing alpha power across different conditions all used individualized alpha peak frequencies. To compute the alpha activity time course for a parcel of interest, we computed the alpha power within each parcel by averaging the above-mentioned power estimate within a ±1 Hz band centered at the individual alpha peak frequency.

Cluster-based permutation tests
Statistical significance was evaluated using cluster-based permutation tests (Maris & Oostenveld, 2007). Briefly, time series of alpha power (e.g., alpha power over time within the selected ROIs) were first compared univariately at each time point. Neighboring time points for which a two-tailed paired t-test resulted in a nominal p-value smaller than 0.05 (uncorrected) were clustered. The sum of the T-values (or F-values) within a cluster was then computed as cluster-level statistics. The cluster with the maximum sum was subsequently used as test statistic. By randomly permuting the condition labels (or by shuffling the predictor variable of interest across trials) and recalculating the test statistic 10,000 times, we obtained a reference distribution of maximum cluster T-values (or F-values) to evaluate the statistic of the actual data (alpha = 0.05). For reference, we also report the smallest p-value obtained for cluster-based permutation tests that failed to reject the null hypothesis.

Distracters lead to interference in memory recall
We first asked whether presenting distracters during the workingmemory maintenance window resulted in worse performance by comparing participants' accuracy across the three conditions of interest. Our repeated-measures ANOVA showed that neither the overall accuracy (F(2, 48) = 1.89, p = 0.162, partial η 2 = 0.006; Fig. 1C, left panel) nor the median reaction times of the correct trials (F(2, 48) < 1, partial η 2 < 0.001; group median reaction times in seconds: auditory, M = 1.608, SE = 0.016; no distracter, M = 1.611, SE = 0.017; visual, M = 1.588, SE = 0.015) significantly differed across conditions. We then asked whether presenting distracters resulted in any interference during recall by focusing specifically on the swapping errors that participants made in their responses. Note that six of the ten digits were used in the target array, and the remaining four were either not shown at all (in the no distracter condition) or were presented as distracters. Swapping errors here refers to errors in which participants reported one of these four nontarget digits.

Alpha oscillations protect working memory in a modality-specific way: ROI-based analysis
Do alpha oscillations protect working memory in a modality-specific manner? Driven by our a priori hypotheses, we started our investigation by focusing on the sensory ROIs. We selected these sensory ROIs in a data-driven way, capitalizing on the evoked responses to the auditory and visual distracters respectively (Fig. 2AB). We verified that (i) the selected sensory ROIs exhibited prevalent alpha oscillatory activity across participants (Fig. 2C), and that (ii) the sensory ROIs showed modality-specific responses to visual and auditory stimuli (Fig. 2D). For the first, we checked that the FOOOF approach provided reasonable fits to the individual power spectrums (group-level goodness-of-fits indexed by median (interquartile) R-squared for left visual, right visual, left auditory, right auditory ROIs: 0.906 (0.081); 0.915 (0.065); 0.913 (0.061); and 0.906 (0.051)) and that the analysis returned an oscillatory component within the 8-13 Hz range for more than 85% of participants for all ROIs (the number of participants exhibiting an alpha component in left visual, right visual, left auditory, right auditory ROIs: 22, 23, 23, and 23, respectively). For the second, we used a two (condition: visual vs. auditory) by two (ROI: visual vs. auditory) repeated-measures ANOVA on the normalized evoked-response amplitude, which showed a significant interaction effect (F(1,24) = 11.76, p = 0.002; post-hoc comparisons: for auditory ROI, auditory vs. visual: t(24) = 3.66, p = 0.001; for visual ROI, auditory vs. visual: t(24) = -1.89, p = 0.071).
Using these ROIs, we then computed the alpha activity time courses for each of the three conditions of interest. Note that the alpha frequency of interest was defined for each ROI within each individual, and hence a normalization procedure is necessary before individual data enter group-level statistical comparisons. Here, the normalized alpha activity time course (per ROI) was computed as α x = αx− αNo αNo * 100%, with α x representing alpha power in the condition of interest, and α No representing alpha power in the no distracter condition.
If alpha oscillations protect working memory in a modality-specific manner (Fig. 1B, right panel), we should expect alpha activity increases in the auditory (visual) regions most responsive to the distracters when the distracters are auditory (visual). In other words, we expect an interaction effect on alpha activity in the two sensory ROIs and distracter conditions, which can be revealed by testing whether the differences between the two conditions have different signs in the different ROIs. We tested such an interaction effect using the full time course (i.e., time window used for statistics = [0, 6.95] s, starting at cue onset and ending at response onset) of alpha activity and corrected for multiple comparisons using a cluster-based permutation approach. Our results showed significant interaction effects on alpha activity, indicated by two distinct significant clusters (Fig. 3AB). One of them (p < 0.001) spans the time window of [3.2, 5.65] s, which overlaps mostly with the distracter presentation (i.e., [2.5, 5.35] s) and is primarily driven by modality-specific evoked responses to the different types of distracters. Another cluster (p = 0.009) spans the time window of [6.35, 6.95] s, which overlaps mostly with the probe window (i.e., [6.15, 6.95] s). To characterize the exact pattern of this interaction effect, we then zoomed in to the time window of [6.4, 6.7] s and performed post-hoc comparisons on the averaged alpha activity in this window (Fig. 3C). Note we only focused on this 300-ms window during the probe presentation to avoid confounding other task aspects due to the use of a 500-ms smoothing window for estimating alpha activity. Our results show that while alpha activity in the auditory ROI is significantly stronger under the auditory distracter condition (t(24) = 2.62, p = 0.015), alpha activity in the visual ROI is significantly stronger under the visual distracter condition (t(24) = -2.80, p = 0.010). The observed interaction suggests that alpha activity increases in a modality-specific manner during memory retrieval, following distracter presentation.
Does the extent of modality-specific alpha power increase relate to the participant's task performance? To answer this question, we compared the modality-specific alpha power, i.e., alpha activity of the auditory (or visual) ROI in response to the auditory (or visual) distracters, between (i) correct vs. incorrect trials and (ii) fast and slow trials (median-split based on reaction times), using a two (modality: auditory vs. visual) by two (response: correct vs. incorrect; or fast vs. slow) repeated-measures ANOVA. While we did not observe a difference between correct vs. incorrect trials, we found that trials with a faster response were associated with higher modality-specific alpha power ( Fig. 3D; main effect of response: F(1,24) = 10.47, p = 0.0004; post-hoc comparisons suggested that this effect was mainly driven by the auditory modality (auditory: p = 0.014; visual: 0.093).

Alpha modulation most pronounced in sensory cortices: whole-brain analysis
While the ROI-based methods provided maximum power in adjudicating between different hypotheses, it may have precluded us from observing other task-related alpha modulations. We therefore complemented our ROI-based results with more exploratory whole-brain analyses, where we specifically investigated the spatiotemporal distribution of the alpha effect. This was done by performing clusterbased permutation tests comparing the three conditions for each frequency within the alpha range at the sensor level. We first focused on the time window from cue onset to distracter onset ([0, 2.5] s time-locked to cue onset, in steps of 50 ms), and found that the three conditions do not significantly differ from each other in this time window (p = 0.683), suggesting little evidence for proactive, preparatory alpha-band modulation when anticipating distracters. Next, we focused on the time window from distracter offset to response onset ([5.35, 6.95] s, in steps of 50 ms), and found that the three experimental conditions significantly differ from each other (p < 0.001), with the difference most pronounced in the sensors covering the occipital and temporal cortices (Fig. 4AB). We then source-localized the alpha modulation (using partial canonical coherence beamformer approach, see Source reconstruction of spectral data for details) by zooming in on the two 800-ms time windows corresponding to the distracters-to-probe delay and probe window. We computed the T-values for pairwise comparisons with the resulting source-level spectral data, to better understand the spatial distributions of the observed alpha power differences.
During the distracters-to-probe delay window, our results (Fig. 4A) showed differences in alpha power between the three experimental conditions in the right auditory area and regions around the left intraparietal sulcus (IPS). Post-hoc pairwise comparisons (Fig. 4C) indicated that alpha power was the highest in auditory areas under the auditory condition compared to the no distracter ( Fig. 4C top panel) and visual distracter condition (Fig. 4C bottom panel); and that it was the lowest in parietal areas under the visual distracter condition compared to the no distracter condition (Fig. 4C middle panel). To understand the dynamics of alpha activity in the above-mentioned parietal regions clustering around the left IPS, we selected one parcel within and zoomed in for further exploratory analyses. First, we examined the time course of the alpha activity within that parcel (Fig. 4E), and found that the modulation (by experimental conditions) was present only during the distracters-to-probe delay window, and seemed to fade out over time from distracter offset to probe onset. Second, we asked if alpha power during the distracters-to-probe delay window had any effect on participants' behavioral responses. We did this by comparing alpha power between (i) correct and incorrect trials and (ii) fast and slow trials (median-split based on reaction times) of the three experimental conditions, using repeated-measures ANOVA with a within-subject three (distracter condition: auditory vs. no distracters vs. visual) by two (response: correct vs. incorrect; or fast vs. slow) design. We did not observe a significant interaction (Fs(2, 48) < 1) nor main effect of response (Fs(1,24) < 1) in both analyses, suggesting that alpha activity in the selected parietal parcel has little impact on behavioral responses.
During the probe window (from 6.15 to 6.95 s), source localization of Average alpha power within the probe window in the auditory ROI in response to fast vs. slow auditory distracter trials, and that in the visual ROI in response to fast vs. slow visual distracter trials. Error bars denote within-subject standard error. Black asterisk denotes statistically significant (p < 0.05) comparison. For all panels, the auditory distracter condition is plotted in green, the visual condition in purple.
the alpha activity (10 Hz, with a frequency smoothing window of ±3 Hz) demonstrated that our experimental manipulation mostly caused alpha-band modulations in sensory cortices (Fig. 4B). Post-hoc pairwise comparisons (Fig. 4D) indicated that alpha power was strongest in bilateral auditory regions under the auditory distracter condition (Fig. 4D top and bottom panels); and it was strongest in bilateral visual regions under the visual distracter condition (Fig. 4D middle and bottom panels). These results validate our a priori hypothesis-driven methods of zooming in on auditory and visual regions most responsive to distracters, and together suggest that alpha oscillations protect memory from distracters in a modality-specific manner during memory retrieval.

Discussion
The current study aimed to characterize the spatial and temporal specificity of alpha activity modulation in a working-memory task with predictable distracters. As expected, we show that presenting distracters during working-memory maintenance causes interference. Furthermore, capitalizing on an ROI-based approach at the source level, we find a significant interaction on alpha power by distracter type and sensory region, consistent with the modality-specific hypothesis of alpha inhibition (Fig. 1B, right panel). Moreover, we find that such a modalityspecific inhibitory mechanism is only present during the probe window when participants retrieve the encoded information from memory, suggesting that the mechanism is reactive, rather than proactive (or preparatory), in nature.
Our task required participants to memorize an array of numbers as well as the spatial location at which they were presented. Since this task relies on visual (spatial) working memory, we did not expect that both types of distracters, irrespective of the sensory modality via which they were presented, would cause interference to a similar extent. Additionally, we speculated such interference might depend on the participant's strategy (Baddeley, 2012). For example, one strategy to memorize the target is to reiterate the numbers of the array (i.e., phonological coding); another strategy is to take a visual snapshot of the target array during its presentation (i.e., visual coding). Indeed, via post-experiment (non-procedural) debriefing, we learned that some participants found the auditory distracter condition to be harder, whereas others found the visual distracter condition harder. As we did not explicitly instruct our participants to use any specific strategies, this factor was not included in the analyses. Yet, as the experimental conditions were manipulated in a within-subject manner, our key findings regarding the spatial and temporal specificity of alpha activity hold irrespective of the specific strategy participants used to maintain the target arrays. Additionally, because we identified our ROIs in a data-driven manner, alpha activity in the identified ROIs likely primarily reflects distracter-related processing, rather than working-memory maintenance related processing.
A crucial aspect of our ROI-based finding is that the extent of modality-specific alpha power increase predicted the participant's reaction times. When the participant was presented with auditory (visual) distracters, stronger alpha power in the auditory (visual) ROI during the probe window led to faster responses. Because we equated temporal predictability of events in all trials, it is unlikely that temporal anticipation of the probe drives the observed alpha effect. Moreover, as the strongest modality-specific alpha modulation was present more than 800 ms after distracter offset (during the probe window), the observed modality-specific alpha modulation is not likely to be contaminated by a spill-over of distracter-related desynchronization of alpha activity. Together, our ROI-based analyses provide robust support that stronger alpha activity results in stronger suppression of the distracter-related information encoded in these regions of interest.
We complemented our ROI-based analysis with a whole-brain analysis, to explore if alpha activity in other brain areas is involved in distracter suppression. Results of the whole-brain analysis showed that the experimental manipulations modulated alpha activity mostly in temporal and occipital areas during the probe window, confirming our ROIbased approach of a priori focusing on brain areas most responsive to the visual and auditory distracters. Curiously, we also observed that during the distracters-to-probe window alpha power in parietal areas around IPS was lower under the visual distracter condition, when compared to the no distracter condition. Such a difference was not present when comparing the auditory and no distracter conditions, nor when comparing the visual and auditory distracter conditions. As the workingmemory task was visual in nature, one may speculate that lower alpha power in parietal areas under the visual distracter condition reflects less inhibition of these areas, which have been shown to play a key role in working memory storage (Bettencourt & Xu, 2016;Christophel et al., 2018;c.f., Rademaker et al., 2019). This explanation is consistent with recent findings that distinct neural networks operating within the alpha frequency range play different functional roles in a working-memory task (Rodriguez-Larios et al., 2022). However, it should be noted that the observed parietal alpha power decrease under the visual distracter condition may reflect the spillover of event-related desynchronization (ERD) of alpha activity caused by the presentation of visual distracters (Klimesch, 2012). Additionally, without a clear link to participants' behavioral responses, caution should be taken inferring the functional role of parietal alpha dynamics in this context.
The temporal window during which we observed alpha activity increase due to the presence of distracters is in sharp contrast to that reported by the seminal paper of Bonnefond & Jensen (2012). While both the current study and Bonnefond & Jensen (2012) used temporally predictable distracters, a crucial difference is how predictive information was provided. Specifically, we used a trial-based design and present predictive information at the start of each trial, whereas Bonnefond & Jensen (2012) used a blocked design to cue the predictive aspect of the distracters (i.e., the modality in the current work, and the extent of similarity between distracter and memory set in Bonnefond & Jensen (2012)). Hence, the lack of anticipatory (or proactive) alpha modulation may be due to the brain's inflexibility in endogenously employing distracter suppression mechanisms (but see van Zoest et al., 2021 for proactive alpha modulation caused by cues). In fact, previous work has shown that, unlike predictive cueing of task-relevant information, predictive cueing of task-irrelevant (or distracting) information only facilitates behavior when the cues are blocked (Noonan et al., 2016). This line of reasoning has called into question the nature of inhibition indexed by alpha activity (Foster & Awh, 2019;Noonan et al., 2018), and has led to a recent proposal suggesting an alternative mechanism (Jensen, 2023). In line with this, Noonan et al. (2018) proposed there are (at least) three types of plausible inhibitory mechanisms: (i) direct inhibition, hypothesized to be flexible and stimulus-specific, and has a proactive temporal profile; (ii) secondary inhibition, hypothesized to be a consequence (thus "secondary") of the selection of task-relevant information, is expected to be non-specific and more reactive in nature; and (iii) expectation suppression, which often requires substantial learning of statistical regularities. The current findings are unlikely to reflect the "expectation suppression" type of inhibition (Zhou et al., 2020), and it is yet unclear whether they should be considered evidence in support of direct inhibition, as we did not observe a proactive temporal profile. While the absence of evidence should be treated with caution, the additional positive findings in support of a reactive temporal profile suggest that the observed activity modulation is most likely reactive, rather than proactive, in nature.
We presented distracters in a rhythmic, temporally predictable way, in order to maximize the neural modulation of interest, as predictability can facilitate the employment of top-down control of distracter suppression. Yet, one should note that distracters are often temporally unpredictable in everyday life, and in this regard, the current task is limited in its ecological validity. Caution should be taken generalizing the current findings to everyday scenarios with unpredictable timing.
Taken together, our current study shows that oscillatory activity in the alpha band protects working-memory traces from predictable distracters in a modality-specific way. These results are consistent with the inhibition account of alpha oscillations. Combining a balanced experimental design (see Wöstmann et al., 2022 for recommendations on designing an experiment on distracters) and high-resolution neuroimaging techniques, our findings shed light on neural mechanisms underlying distracter suppression, and further our understanding on how the brain (de)selects information flexibly.

Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Data availability
All data and code for stimulus presentation and analysis are available online at the Donders Repository at https://data.donders.ru.nl