Balancing excitation and inhibition: The role of neural network dynamics in working memory gating

In everyday life, we constantly need to flexibly adapt to an ever-changing complex environment. While in some situations one needs to be focused and resist distracting inputs, at other times it might be beneficial to let external inputs interfere with current mental representations. While reading these lines, for example, you need to concentrate and maintain current working memory (WM) contents despite possible distractions. Yet, you might also need to shift your attention if someone nearby mentions an urgent matter. A cognitive system enabling this adaptive coordination between distractor-resistant maintenance and updating of mental representations is WM gating (Chatham & Badre, 2015; Durstewitz et al., 2000). Prior studies, utilizing computational models and neurophysiological data, have successfully demonstrated the involvement of dopaminergic frontostriatal loops in these adaptive processes (Fallon & Cools, 2014; Fallon et al., 2017b; Landau et al., 2009). While the prefrontal cortex (PFC) plays a crucial role in sustaining representations in a distractor-resistant manner, the process of updating these is mediated by dopaminergic signaling in basal ganglia go/no-go pathways (Cools & D’Esposito, 2011; Landau et al., 2009). Rises in dopamine promote “go”-pathway activation via D1 receptors, facilitating updating. Lower dopamine levels activate the “no-go”-pathway via D2 receptors, supporting maintenance. Importantly, the effectiveness of phasic rises in dopamine to override tonic dopamine signals in the PFC has been shown to depend on initial baseline dopamine levels in the striatum (Cools & D’Esposito, 2011; Ranganath & Jacob, 2016; van Schouwenburg et al., 2010).

In addition to these dopaminergic mechanisms, the functional organization of neural networks in terms of excitatory and inhibitory (E/I) activity is also crucial for regulating working memory gating. As such, persistent maintenance of working memory contents in the PFC is governed by recurrent excitation within the network and inhibitory feedback regulation thereof (Durstewitz et al., 2000). Within frontostriatal loops, various excitatory and inhibitory connections balance adequate go-activation (triggering updating) or no-go activation (supporting maintenance; see O’Reilly & Frank, 2006). Consequently, the precise interplay of excitatory and inhibitory signals is crucial for effective WM gating. However, the exact mechanisms by which these signals navigate distractor-resistant maintenance and selective updating remain elusive (see also Dipoppa et al., 2016). One approach to examine excitation/inhibition (E/I) network dynamics is to analyze long-range temporal correlations (LRTCs) in ongoing oscillations from resting-state EEG. LRTCs describe the statistical dependency of signal fluctuations across different time scales, reflecting that the network may operate near the critical state—at the sweet spot between order and chaos (Hardstone et al., 2012; Linkenkaer‐Hansen et al., 2001). These critical‐state dynamics in neural oscillations have been linked to optimal information processing (Beggs & Plenz, 2004; Shew et al., 2011; Shew & Plenz, 2013), and deviations from this state have been associated with various cognitive dysfunctions, including depression (Linkenkaer-Hansen et al., 2005), schizophrenia (Moran et al., 2019; Nikulin et al., 2012), and epilepsy (Monto et al., 2007). LRTCs, particularly those derived from spontaneous oscillations during rest, exhibit high heritability (Linkenkaer-Hansen et al., 2007) and test–retest reliability (Nikulin & Brismar, 2004), making them a valuable marker for trait-like network characteristics. Importantly, these correlations emerge only when the excitatory and inhibitory activity across the underlying circuit is well balanced (Poil et al., 2012)—a relationship that has most comprehensively been characterized for the alpha frequency range (Fedele et al., 2016; Pfeffer et al., 2018; Poil et al., 2012). LRTCs from resting-state alpha oscillations hence pose an ideal candidate for inferring the intrinsic E/I network balance. Furthermore, previous research suggests that stronger resting-state alpha LRTCs are associated with better performance on tasks that probe higher-order brain functions such as attention (Irrmischer et al., 2017), sensorimotor processing (Samek et al., 2016), and adaptive decision making (Colosio et al., 2017), and, most importantly, phasic adaptive working memory (Mahjoory et al., 2019).

In addition to LRTCs, also the 1/f slope of the power spectral density (PSD) has been shown to be indicative of E/I network dynamics. Specifically, it reflects the ratio of excitatory and inhibitory activity across the circuit—the network’s E/I ratio (Gao et al., 2017). Animal studies have demonstrated that an increase in excitatory connections is associated with a flatter PSD slope, while an increase in inhibitory connections results in a steeper PSD slope (Gao et al., 2017; see Fig. 3A), determined by the relative abundance of glutamatergic and GABAergic connections, respectively. These findings have been extended to human EEG and MEG studies (Donoghue et al., 2020), show high reliability (Pathania et al., 2021), and were further supported by pharmacological interventions that manipulate the E/I ratio (Colombo et al., 2019; Lendner et al., 2020; Medel et al., 2023; Zsido et al., 2022).

In the present study, we utilize these two markers to investigate how E/I network dynamics relate to WM gating. We first examine their association with performance on a WM gating task, specifically probing ignoring versus updating of WM representations (see Fallon & Cools, 2014; Hartmann et al., 2023; N. Herzog et al., 2024a, 2024b). Next, because WM gating heavily relies on dopaminergic signaling, we also explore the impact of blood amino acid ratio, that is the ratio of tyrosine and phenylalanine (precursors of dopamine) to other large neutral amino acids, on the observed relationships. This ratio is indicative of peripheral blood dopamine precursor availability and has been shown to relate to dopamine signaling in the striatum (Leyton et al., 2004; Montgomery et al., 2003; see Fig. 2A), and working memory gating performance (N. Herzog et al., 2024b). Lastly, we also probe whether effects of LRTCs or 1/f slope on WM gating may be explained by a more direct neural signature of WM gating, that is P300 amplitude at the moment of ignoring distractors versus updating. The P300 component is an event-related potential (ERP) that is commonly used as a neural marker for the allocation of attentional resources as well as updating of working memory representations (Gray et al., 2004; Polich, 2007).

Given previous reports, we hypothesized that (i) stronger resting-state alpha LRTCs will be associated with better overall WM performance, with more pronounced effects on updating, and that (ii) a steeper PSD slope—reflecting more inhibition—should foster better maintenance at the expense of updating. We further expect blood amino acid ratio to impact these relationships, and that both, LRTC and PSD slope, will influence P300 amplitude depending on the condition.

In total, 78 participants were recruited via the internal participant database of the Max Planck Institute for Human Cognitive and Brain Sciences (Leipzig, Germany) and via advertisements in public places and university facilities. Participants were specifically screened to exclude those with a history of clinical drug or alcohol abuse, neurological or psychiatric disorders, or first-degree relative history of neurological or psychiatric disorders. None of the participants had symptoms of depression at the time of the experiment, as assessed via a screening interview using the Structured Clinical Interview for DSM-IV (SCID; Wittchen et al., 1997). Five subjects had very noisy EEG data (see methods) and were excluded from the analysis. The final sample size thus was 73 participants (39 female; 34 male, 0 diverse). The mean age was 26.75 years (SD = 3.79, min = 20.10, max = 36.29). Mean estimated IQ was 108 (SD = 9.52, min = 74, max = 118) and mean BMI was 23.22 kg/m² (SD = 2.68, min = 18.92, max = 29.89). The study was conducted in compliance with the principles of the Declaration of Helsinki and was authorized by the Ethics Committee of the Medical Faculty at the University of Leipzig (172/19-ek). All participants provided written informed consent before participation and were financially compensated for their time.

This study was pre-registered at https://osf.io/zdmkx (see transparency statement in the supplements for an explanation of deviations). After an initial screening via phone, participants were asked to come to the laboratory for two separate sessions (Supplementary Fig. S1). On the first day, they underwent the screening session where further inclusion and exclusion criteria were checked. After inclusion, serum blood samples were taken to extract information on amino acid profiles indicative of central blood amino acid ratio. Participants, therefore, came overnight-fasted. After the blood draw, participants filled in the Dietary Fat and free Sugar Questionnaire (DFS; Francis & Stevenson, 2013; Fromm & Horstmann, 2019), assessing their eating behavior and the BIS/BAS questionnaire (Behavioral inhibition and activation system scales; Carver & White, 1994; Strobel et al., 2006), assessing trait impulsivity. On the second test day, participants first did two neuropsychological tests: the digit span task (assessing baseline WM; Wechsler, 2008) and an IQ Test (“Wiener Matrizen Test-2”; Formann, Waldherr & Piswanger, 2011). After that, participants were prepared for the EEG. They first did a 10-min eyes-closed rest EEG recording. Participants were instructed to sit still, close their eyes, and stay awake. Participants then did a short training on two working memory tasks—only one of which is subject to this study (see pre-registration). Task order was counterbalanced. Upon completion of the training, participants were asked to explain the task in their own words. The experimenter had a checklist that entailed three major points participants had to mention (see Supplementary Table S1). Participants could only proceed the actual experiment if they named all three points and had a training accuracy of >80%. All participants fulfilled these criteria on the first attempt. EEG activity was recorded during the working memory task. Upon completion of the tasks, participants were asked to indicate their level of tiredness and concentration felt during the task on a 10-point Likert scale.

The task used is the same as in previous studies conducted in our laboratory (Hartmann et al., 2023; N. Herzog et al., 2024a, 2024b). It is a variation of a delayed match-to-sample task originally designed by Fallon and Cools (2014), and comprises four conditions as shown in Figure 1A. In the ignore condition, participants first had to memorize a pair of patterns marked with a “T” for “target.” Subsequently, they are shown another pair of distraction patterns labeled with an “N” for “non-target,” which they must ignore. After that, participants are presented with a probe stimulus and had to determine whether it matches one of the first two target stimuli (“T”). In the update condition, participants again memorize a pair of “target” patterns (marked with a “T”). Then, they are presented with a new set of “target” patterns, which replace the previously presented stimuli as the target. Thus, at the probe phase after an update event, participants need to respond according to whether the probe image matches one of the second (updated) pair of stimuli. The two control conditions, which feature only one pair of stimuli, are matched to the temporal delay between encoding the to-be-matched targets and probe in the ignore and update conditions. The probe is presented for 2,000 msec. Each trial is separated by a fixed inter-trial interval of 2,000 msec. The task is separated into 4 blocks, each consisting of 32 trials spread evenly across the 4 conditions, totaling 128 trials. Feedback on average accuracy is presented after each block, serving as a small break for the participant. The stimuli are randomly computer-generated, monochromatic RGB “spirographs.” Outcome measure is accuracy. The total duration of the task is approximately 30 min.

EEG data were recorded using 61-channel Ag/AgCl scalp electrodes (Brain Products GmbH, Gilching, Germany) in an EEG booth. The electrodes were mounted in an elastic cap (easyCAP, Herrsching, Germany) according to the international standard 10–20 extended localization system. The signal was amplified with a QuickAmp amplifier (Brain Products GmbH, Gilching, Germany). Additionally, two electrodes were placed above and beneath the right eye and recorded vertical (vEOG) and horizontal (hEOG) eye movements. One bipolar electrode attached to the right and left forearm recorded electrocardiogram (ECG). The electrodes were referenced to the common average reference with FCz being a ground electrode. The electrodes’ impedances were kept below 10 kΩ, the sampling rate was 1,000 Hz. EEG data were initially pre-processed using the MATLAB-based (MathWorks, Inc., Natick, MA, United States) EEGLAB toolbox (version 14.1.1b) and custom-written scripts. After downsampling the data to 500 Hz, it was band-pass filtered between 0.5 and 45 Hz (4th order Butterworth filter applied back and forth). In addition, a notch filter at 50 Hz was applied to remove any remaining power line artifacts. Flat channels and channels with correlation <0.8 with neighboring channels were removed using the clean_artifacts() function, and then interpolated. Bad epochs were detected using an adjusted version of the trimOutlier() function: We set different amplitude threshold levels for the noise detection at low-frequency (1–15 Hz) and high-frequency (15–45 Hz) ranges. For the low-frequency range, the individual noise threshold was defined as three SD above the mean amplitude of the filtered signal. We set a constant amplitude threshold of 40 μV for the high-frequency range given that noise in the high-frequency range is smaller in amplitude. Recordings for which the total bad segment length exceeded 60 s were inspected visually to confirm that the marked segments were indeed contaminated by noise. We excluded vEOG, hEOG, and ECG channels from the dataset and then applied independent component analysis (ICA, extended Infomax; Bell & Sejnowski, 1995). Components that reflect artifacts, such as muscle activity, non-biological noise, or eye-blinks, were marked and removed. Finally, we visually inspected the PSD of the multi-channel data of all subjects to determine whether data were sufficiently cleaned.

LRTCs were calculated using detrended fluctuation analysis (DFA; Hardstone et al., 2012). First, EEG signals were filtered into the alpha frequency range between 8 and 14 Hz and the amplitude envelope was extracted using the Hilbert transform. Then, the cumulative sum of the signal envelope was calculated. Finally, the root-mean-square fluctuation of the linearly detrended signal was calculated as a function of window size and plotted on double logarithmic axes. Time windows ranged between 2 and 25 s. Fitting a least-squares line eventually gives the DFA exponent, as reflected by the slope of the line. Notably, scaling exponents (α) in the range of 0.5–1.0 are generally considered indicative of persistent correlations indicating that the past of the signal affects its future, while a value of 0.5 indicates uncorrelated white noise (Cannon et al., 1997; Hardstone et al., 2012; Stam & de Bruin, 2004). The PSD of each channel’s data was calculated from the cleaned data using 4-second Hamming windows overlapping by 50% using Welch’s method. We used the Python (version 3.9) implementation of the FOOOF algorithm (Donoghue et al., 2020) on the PSD to estimate the slope of the 1/f decay for each channel separately: Here, broad-band power spectra between 3 and 40 Hz were modeled as $P(f)∼1/fγ$⁠, where γ is the spectral slope. Additionally, we extract relative alpha power from the resting-state EEG to account for sleepiness. Relative alpha power was defined as the proportion of the alpha (8 to 13 Hz) wave activity compared with the total power of all frequencies.

To determine the peripheral amino acid ratio, blood serum samples were collected from participants who had fasted overnight. These samples were analyzed at the Institut für Laboratoriumsmedizin, Klinische Chemie und Molekulare Diagnostik, University Clinic Leipzig. The blood amino acid ratio was calculated as:

All behavioral analyses were done in RStudio v4.0.2 (R Core Team, 2015; RStudio Team, 2016). To assess whether intrinsic E/I network dynamics would relate to condition-dependent task accuracy, we ran linear mixed models using lmer() function from lme4 package. The dependent variable in our models was accuracy. Before calculating average accuracy per condition, trials with a reaction time <200 ms were excluded, as those trials can be considered false alarms. Also, trials with a reaction time >2,000 ms were excluded, as those are considered misses. The predictor variables were the within-subject factor condition (ignore vs. update vs. control long vs. control short), and its interaction with the between-subject factor DFA exponents or PSD slope, respectively. To account for the within-subject nature of our data, we also included the random factor subject in all our models. Due to model convergence problems, we were not able to maximize the random structure of our model as recommended for linear mixed models (Barr, 2013). As we did not have any a priori hypothesis about the location of the effects, we ran separate models per electrode and did cluster-based permutation testing to correct for multiple comparisons (see below). We ran separate models for each network measure. Our models were:

Additionally, because WM gating depends on dopaminergic signaling, we further probed the influence of blood amino acid ratio on the relationship between intrinsic E/I network dynamics and WM gating in an exploratory manner. We, therefore, added blood amino acid ratio as factor of interest to the models:

To mitigate the issue of multiple comparisons in EEG analysis, we employed cluster-based permutation testing (Maris & Oostenveld, 2007). Specifically, as we ran one model per electrode, we identified electrodes in which DFA or PSD slope exhibited a significant condition-dependent association with task performance and grouped them into clusters. A group of electrodes was considered a cluster if it comprised at least two spatially adjacent electrodes that both demonstrated significant interaction effects at p < 0.05. The statistical significance of each identified cluster was then assessed by comparing its cumulative t-value against a null distribution of cumulative t-values generated by randomly resampling the data and recalculating the models 1,000 times. For every permutation, we determined and recorded the cumulative t-value of the largest spatially adjacent cluster. The cumulative t-value of our original cluster was then evaluated against this permutation-based null distribution. A cluster was deemed statistically significant if its cumulative t-value exceeded the 95th percentile of the permuted cumulative t-values, indicating that the observed clustering is unlikely to have arisen by chance under the null hypothesis.

Because DFA exponent, PSD slope, as well as task accuracy might partially be explainable by other confounding factors, we ran further control analyses. For this, we calculated a mean DFA and PSD slope exponent score across all electrodes within a cluster. These cluster averages were then used to re-run our models, incorporating possible confounders such as age, BMI, DFS, and BIS/BAS scores, task-related tiredness and concentration, sex, and IQ. Notably, since both the alpha DFA exponent and PSD slope may be affected by alpha power and SNR (e.g., Linkenkaer-Hansen et al., 2007; Meisel et al., 2017; Pathania et al., 2021), we also included relative alpha power in the control model. By standardizing alpha power against total EEG power, we emphasize the effects of alpha-specific activity, rather than broadband difference in overall power, accounting for the individual contribution of alpha power. Additionally, this approach partially addresses variations in the data’s signal-to-noise ratio. Due to issues with model convergence, all continuous covariates were z-transformed.

Given that LRTC solely indicates whether the E/I balance is “optimal”, but not clearly discern whether the network leans more toward excitation or inhibition, we opted for an additional exploratory analysis. For this, we chose to look at how LRTCs and PSD slope would interactively predict condition-dependent task performance. This enabled an exploration of whether criticality in a more “excitable” sample (flatter PSD slope) would predict task performance differently than criticality in a more “inhibitory” sample (steeper PSD slope). For this, we build another average PSD slope value, based on those electrodes where LRTC significantly predicted condition-dependent performance (Fig. 2C, white circled electrodes). The final model included the dependent variable accuracy and the between-subject factors average DFA and average PSD slope, and the within-subject factor condition. For this analysis, we chose to look at our main conditions of interest only (ignore vs. update). The model was:

In order to gain a more detailed understanding of the origin of our effects, we repeated our sensor-level analysis (reported above; models 1, 2, 1a, and 2a) at the source level. We performed the analysis in Matlab R2022b, using the lead field matrix of the New York head model (Huang et al., 2016). A total of 4,502 sources with fixed orientations perpendicular to the cortical surface were considered. To match the preprocessing of the EEG data, a common average reference transform was applied to the lead field matrix (see Kapralov et al., 2023). The time series of activity for all sources were reconstructed using the eLORETA algorithm (exact low-resolution brain electromagnetic tomography), as implemented in the M/EEG Toolbox of Hamburg (METH, Guido Nolte; RRID: SCR_016104) for inverse modeling. We set the regularization parameter of eLORETA to 0.05 and used an identity matrix as the noise covariance matrix. The source-projected data were then filtered and transformed to compute the DFA and PSD slope exponent per vertex with exactly the same parameters used at the sensor level (Section 2.5). Subsequently, we used these values to perform vertex-wise analysis utilizing linear mixed models (lmer() in R v4.2.2). The models included the dependent variable average accuracy, and the within-subject predictor condition, and its interaction with the between-subject predictor source DFA exponent (model 4a) or source PSD slope and amino acid ratio (model 4b). For this analysis, we decided to refrain from correction methods, as it was a complementary analysis.

Preprocessing of the task-EEG was done similar to resting-state EEG preprocessing. After that, task-EEG analysis was done using custom-written scripts in MNEmne Python (version 3.9). Continuous data were epoched based on task condition (ignore, update). Epochs extended from 500 ms pre-stimulus to 1,800 ms post-stimulus. Baseline correction was applied to each epoch using a pre-stimulus interval from -500 to -50 ms. After that, custom artifact rejection criteria were implemented to mark and reject remaining epochs with abnormal signal fluctuations: Epochs were marked if (i) more than 30% of the samples in a single channel exceeded an amplitude threshold of 15 µV, considering this as a sign of potentially artifactual activity, and if (ii) such activity was observed in more than two channels. After that, marked epochs were inspected visually and manually removed if they were indeed too noisy. On average, 4.93 epochs were removed per subject. Five subjects had too noisy task-EEG (>60% of epochs removed) and were, therefore, excluded from the analyses. By plotting the grand average ERP time series, we found that P300 amplitude (measured at Pz) peaked at around 320 msec. To capture this component, we thus extracted the average amplitude from 310 to 500 msec (Fig. 5A, gray area) per trial.

To check whether P300 amplitude would be a valid indicator of WM gating in our task, we ran trial-by-trial analysis using a logistic regression glmer() function from the lme4 package in R. We built the model using trial accuracy (binary; correct vs. incorrect) as outcome variable, and the within-subject factors P300 amplitude and condition (ignore vs. update) as interacting predictors. Trials with a reaction time <200 and >2,000 ms were excluded for this analysis. The model was

To delve deeper into the influence of LRTCs or PSD slope on P300 amplitude, we performed further analysis utilizing trial-based linear mixed models (lmer(), lme4 package). Condition-dependent P300 amplitude was predicted by either cluster DFA (model 6a) or the combined effects of cluster PSD slope and amino acid ratio (model 6b).

None of the participants performed below chance level (50% correct trials per condition, see Fig. 1B). Accuracy across conditions differed significantly (p_uncorrected < 0.001). Participants performed best in the short control condition (mean accuracy = 95.5%, SD = 4.2%), followed by update (mean = 94.5%, SD = 7.6%), control long (mean = 93.3%, SD = 6.4%), and ignore (mean = 89.5%, SD = 7.8%).

Next, to test whether WM gating performance would depend on intrinsic network dynamics in terms of E/I balance, we correlated resting-state DFA exponents to condition-dependent accuracy (see model 1 in Methods section). The grand average DFA exponent was 0.768 (SD = 0.014; min = 0.738; max = 0.796). Mean LRTC exponents averaged over subjects are presented in Figure 2B. We found that DFA exponents in 10 electrodes predicted condition-dependent task performance significantly (see Supplementary Table S2, column DFA*condition). These electrodes formed two spatially adjacent clusters (Fig. 2C). Using cluster-based permutation testing, we then checked for significance of these clusters. The bigger, right frontoparietal cluster (C4, FC4, C2, CP4, CP6, P6, P8, PO10) was significant at p < 0.05. The smaller, frontocentral cluster (FC1, C1) was not significant (p > 0.05). See Supplementary Figure S2A. Control analyses indicated no influence of confounding variables such as age, sex, and BMI. (see Methods section) on this relationship. Notably, although relative alpha power had a significant main effect on overall task accuracy (p = 0.028), it did not diminish the primary effect of interest; the two-way interaction between mean DFA and condition remained significant. See Supplementary Table S3 for the full model output.

To further investigate the observed DFA–condition interaction, we built a cluster-average DFA value and looked at its effect on accuracy per condition separately (Fig. 2D). These post hoc tests revealed that the initial interaction was likely due to there being a negative relationship between DFA exponents and accuracy in the ignore (β = -0.213, SE = 0.091, p_uncorrected = 0.023) as well as in the two control conditions (β_{control short} = -0.111, SE = 0.050, p_uncorrected = 0.030; β_{control long} = -0.221, SE = 0.074, p_uncorrected = 0.003), but not in the update condition (β = 0.01; SE = 0.09; p_uncorrected = 0.840).

In order to gain a more detailed understanding of the origin of DFA–condition interaction, we repeated the analysis done at the sensor level at the source level (see Methods section). Results reveal widespread effects (Fig. 2E). Sources significant at p < 0.01 (uncorrected) belong to the right frontal cortex, supramarginal gyrus, precentral gyrus, supplementary motor cortex, and parietal cortex, as well as left insular and orbitofrontal cortex (Fig. 2F). See Supplementary Table S4 for a detailed list of all significant vertex clusters. Post hoc analyses looking at the effects of source DFA exponents (averaged over all vertices with p < 0.01) per condition show a pattern similar to that observed at the sensor level (Supplementary Fig. S3A).

Given the reliance of WM gating on dopaminergic signaling, we further probed the influence of blood amino acid ratio (as proxy for dopamine synthesis capacity) on the relationship observed above. Results show no significant three-way interaction in any of the electrodes (Supplementary Table S2, column DFA*condition*AA), suggesting that dopamine synthesis capacity does not moderate DFA effects on condition-dependent performance. Additional control analyses revealed that there was no direct correlation between blood amino acid ratio and resting-state DFA exponent (ß = 0.017; SE = 0.029; p = 0.541; see Supplementary Fig. S5A).

Next, we investigated how E/I ratio, that is resting-state PSD slope, relates to WM gating. Grand average PSD slope exponent was 1.779 (SD = 0.121; min = 1.604; max = 2.019). Mean PSD slope averaged over subjects is presented in Figure 3B. Correlating PSD exponents to condition-dependent accuracy (see model 2 in Methods section), we find no significant two-way interactions (Supplementary Table S2, column PSD*condition), indicating that PSD slope does not relate to condition-dependent task performance.

Interestingly, though, when adding amino acid ratio as a factor of interest, condition-dependent effects of PSD slope became apparent. We found significant three-way interactions in 22 electrodes (Supplementary Table S2, column PSD*condition*AA). These 22 electrodes could be divided into two spatially adjacent clusters (left frontocentral cluster: F7, FT7, FC5, FC3, FC1, C1; right frontoparietal cluster: F4, FC2, C4, P4, P8, O2, PO10, FC4, FT8, C2, C6, CP4, TP8, P6, PO4, PO8; see Fig. 3C). Permutation testing results showed that both clusters were significant at p < 0.05 (Supplementary Fig. S2B). Adding control variables such as age, sex, IQ, BMI, DFS, or BIS/BAS scores, or task-related tiredness and concentration to the models did not change these results. Importantly, while relative alpha power did show a significant main effect on overall task accuracy (p = 0.021), it did not take away from the main effect of interest, that is the three-way interaction of condition, PSD slope, and amino acid ratio stayed significant. See Supplementary Table S5 for the full model output. Additional exploratory analyses revealed that there was no direct correlation between blood amino acid ratio and resting-state PSD slope exponent (ß = 0.007; SE = 0.028; p = 0.502; see Supplementary Fig. S5B).

To further elaborate on the observed PSD—amino acid ratio—condition interaction, we built a cluster average PSD score per cluster and ran additional post hoc analyses, testing the interaction of cluster PSD and amino acid ratio per condition separately (Fig. 3D). These analyses show that the initial interaction was driven by a blood amino acid ratio—PSD slope interaction in ignore (β = -3.51, p_uncorrected = 0.021) and control short (β = -1.94, p_uncorrected = 0.020) conditions, but no such interaction in the update condition (β = 1.30, p_uncorrected = 0.406). For the long control condition, the interaction was not significant (p_uncorrected = 0.179), but the direction of effect aligned with ignore and control short conditions (β = -1.730). This effect was similar for both clusters.

Again, to get a better grasp of the origin of this effect, we repeated the analysis on the source level (see Methods section). Similarly to LRTCs, results show that the origin of the PSD slope, amino acid ratio, and condition interaction effect was widespread (Fig. 3E). Sources with p < 0.01 (uncorrected) were found in the right pre- and post-central gyri, frontal cortex, supramarginal gyrus, precuneous and parietal cortex, and cingulate gyrus, as well as left supplementary motor cortex, subcallosal cortex, and precuneous parahippocampal gyrus (cortex, cingulate gyrus; Fig. 3F). See Supplementary Table S4 for a detailed list of all significant vertex clusters. Post hoc analyses of the interaction between PSD slope (averaged over all voxels with p < 0.01) and amino acid ratio per condition reveal a pattern similar to that observed at the sensor level (Supplementary Fig. S3B).

Given that LRTC solely indicates whether the E/I balance is “optimal”, but does not clearly discern whether the network leans more toward excitation or inhibition, we opted for an additional exploratory analysis. For this, we chose to look at how LRTCs and PSD slope would interactively predict condition-dependent task performance. This way we could check whether criticality in a more “excitatory” sample (flatter PSD slope) would predict task performance differently than criticality in a more “inhibitory” sample (steeper PSD slope). Results reveal that indeed PSD slope and DFA predicted accuracy interactively (β = -1.096, SE = 0.4574, p = 0.016). This did not differ per condition, however (p_{condition*DFA*PSD} = 0.406). Please refer to Supplementary Table S6 for the full Anova Output. To dissect this DFA–PSD slope interaction further, we split the sample into steep versus flat PSD slope, based on median split, and looked at how DFA relates to accuracy in each of these subsamples. These analyses showed that DFA was negatively correlated with accuracy when PSD slope was steeper (β = -0.328; SE = 0.121, p = 0.008), while it did not correlate with accuracy when PSD slope was flatter (β = -0.106; SE = 0.164; p = 0.522; Fig. 4A). Additionally, in the sample with a flatter PSD slope (higher excitation), there was a trend significant interaction effect of condition and DFA (p = 0.071). The correlation between DFA and accuracy was numerically positive but insignificant in the update condition (β = 0.181; SE = 0.172, p = 0.299), while it was numerically negative and insignificant in the ignore condition (β = -0.106; SE = 0.156, p = 0.503). To further explore whether the interaction between DFA and PSD slope on accuracy might be associated with underlying differences in LRTCs across PSD groups, we checked whether DFA differs systematically between the two subgroups (Fig. 4B). Indeed, we found that in the subgroup with a steeper PSD slope, the mean DFA is significantly higher (mean = 0.812, SD = 0.095) than in the flat-slope PSD subgroup (mean = 0.737, SD = 0.0838, t(69.31) = -3.5784, p < 0.001), as shown in Figure 4B. To check whether this dependency leads to multicollinearity in our initial model (Model 3), we conducted a Variance Inflation Factor (VIF) analysis (see, e.g., Thompson et al., 2017). The results showed a VIF of 1.10 for both PSD slope and DFA, indicating that there was no concerning collinearity among the predictors.

To gain an insight into how LRTCs and 1/f slope influence behavioral outcomes, we examined their effects on brain activity during moments of ignoring or updating stimuli. For this, we extracted the P300 event-related potential (ERP) in response to the to-be-ignored or updated stimulus and first investigated whether its amplitude could predict condition-dependent task performance. Since higher P300 amplitude is generally associated with cognitive processes such as attention allocation (Kok, 2001; Polich, 2007) and working memory updating (Donchin & Coles, 1988; Verleger, 1988), we hypothesized that in our task, P300 amplitude would be higher in response to updating a stimulus compared with ignoring one. We observed the opposite pattern, however: average P300 amplitude was significantly higher in the ignore (mean = 1.651 μV, SD = 3.242 μV) than in the update condition (mean = 1.207 μV, SD = 3.356 μV, p < 0.001; Fig. 5A). Please refer to Supplementary Table S7 for the full model output. Next, we tested whether this difference in amplitude could explain task performance. Results showed a trend-significant effect of P300 amplitude on accuracy, which varied by condition (condition*P300: β = -0.063, SE = 0.037, p = 0.086). After controlling for covariates (see Methods section), this effect became significant (β = -0.081, SE = 0.039, p = 0.039). See Supplementary Table S8 for the full model output. Post hoc analyses show that P300 amplitude was negatively associated with update performance (β = -0.071, SE = 0.031, p = 0.025), while it was not associated with performance in ignore trials (β = 0.003, SE = 0.024, p = 0.884; Fig. 5B). As we conducted a trial-based analysis, we noticed that some P300 amplitudes were negative. To check whether these values indicated a shift in peak latency rather than true negative amplitudes, we split all trials into high- and low-amplitude groups, based on whether their amplitudes were above or below 0 μV, respectively. This follow-up analysis confirmed our suspicion (see Supplementary Fig. S4), suggesting that our findings might be due to a combination of changes in amplitude and peak latency shifts. Finally, we investigated whether LRTCs or the PSD slope–amino acid ratio interaction (previously found to be relevant at the behavioral level) could predict condition-dependent P300 amplitude. Our analysis revealed that neither LRTCs nor the PSD slope–amino acid ratio interaction significantly predicted condition-dependent P300 amplitude (p_{DFA * condition} = 0.146, p_{PSD * amino acid ratio * condition} = 0.289). See Supplementary Tables S9 and S10 for full model outputs, respectively.

Balancing distracter-resistant maintenance and flexible updating of working memory representations is a key challenge for the brain and something that commonly goes awry in neuropsychiatric disorders (e.g., Fallon et al., 2017a; Tanaka, 2006; Uitvlugt et al., 2016). Previous investigations have mainly focused on the neural locus of this gating process, demonstrating the involvement of prefrontal cortex and striatum (D’Ardenne et al., 2012; Fallon et al., 2017b; Nir-Cohen et al., 2020). The current study significantly advances these previous findings by exposing the E/I dynamics involved in working memory gating sub-processes. Specifically, we show that these processes can be predicted by two resting-state markers of E/I balance (LRTC) and E/I ratio (PSD slope), respectively.

Interestingly, the effects of both LRTCs and PSD slope were particularly evident in the ignore, control long, and control short conditions, but not in the updating condition. This pattern is noteworthy as it highlights a shared characteristic among the affected conditions: despite interference in the ignore condition, all of them generally require the stable maintenance of WM representations over time. It seems, therefore, that network E/I dynamics primarily affect maintenance-related subprocesses of WM gating. This finding may be explained by the fact that maintenance processes are predominantly managed by the PFC (Frank et al., 2001; Hazy et al., 2006; O’Reilly & Frank, 2006). Since our study uses scalp EEG to capture LRTCs and PSD slope, these measures most likely reflect cortical dynamics, including those generated in the PFC, which is a cortical structure. In contrast, updating primarily relies on signaling of subcortical structures, specifically of the striatum (Hazy et al., 2006; O’Reilly & Frank, 2006; van Schouwenburg et al., 2010). The lack of significant effects of E/I dynamics on updating could, therefore, stem from methodological constraints, particularly the challenge of detecting striatal activation using scalp EEG. The condition-specific pattern observed in our data thus aligns well with literature pointing toward the engagement of different neural circuits underlying the two working memory gating processes. Our data provide further evidence for the dissociation of these processes at the level of cortical network dynamics.

Alternatively, the observed maintenance-specific pattern might suggest that cortical E/I dynamics affect overarching working memory processes rather than gating per se. Supporting this interpretation, we observed that neither LRTCs nor PSD slope influenced condition-dependent P300 amplitude—a more immediate marker of the gating process in our task.

Given its condition-specific relationship to task performance, we conclude that P300 amplitude may serve as a valid marker for working memory gating in our task. Contrary to our expectations, however, the average P300 amplitude was higher in response to stimuli that should be ignored compared with those that required updating. Moreover, an increase in amplitude correlated with declining performance in update trials. At first glance, these findings seem to contradict established theories about P300, reflecting attention allocation and context updating (Donchin & Coles, 1988; Kok, 2001; Polich, 2007; Verleger, 1988). From a different perspective, however, P300 amplitude increases can also be interpreted as reflecting the number of cognitive resources utilized (Ghani et al., 2020). Our results hence suggest that ignoring irrelevant information may demand more cognitive resources, possibly because suppressing incoming information is more challenging. This interpretation gains support by our behavioral data, which indicate that participants generally performed worst in the ignore condition, implying a higher level of difficulty. In our paradigm, elevated P300 amplitude might, therefore, signify controlled inhibition, which would be detrimental in updating trials where participants should encode rather than inhibit stimuli. However, our data lack empirical support for this interpretation, preventing us from drawing definitive conclusions. Utilizing concurrent EEG-fMRI could help link P300 amplitude changes to specific brain regions involved in ignore and updating processes, potentially validating the proposed relationship between P300 and gating.

Nonetheless, the fact that neither LRTCs nor PSD slope affects this paradigm-specific marker of working memory gating suggests that cortical E/I network dynamics may influence task outcomes through mechanisms distinct from those reflected by P300-related gating. The precise pathways by which these cortical network properties modulate performance in our task thus remain an open and intriguing question, warranting further investigation.

Stronger LRTCs are frequently observed to correlate with improved cognitive performance (Nakao et al., 2019; Simola et al., 2017). Most important to our study, Mahjoory et al. (2019) found that stronger LRTCs were linked to faster and more accurate adaptive cognition in a working memory task—a finding we anticipated to extend to our study. Our results contradict these expectations, however, showing that LRTCs generally exhibited a negative relationship with WM task performance. Effect sizes for the significant negative relationships ranged from β = -0.11 to β = -0.22, indicating that even small increases in DFA exponents are associated with noticeable decreases in accuracy. Importantly, our control analysis, assessing the contribution of relative alpha power, suggests that the observed effects are unlikely to be confounded by individual alpha power differences, thereby highlighting the unique role of LRTCs.

The unexpected negative association between LRTCs and task performance might be explainable by theories suggesting that stronger LRTCs, which imply greater autocorrelation of the signal, go along with an increased signal “memory” (see Simola et al., 2017). This concept proposes that higher LRTC scaling relations might lead to a prolonged influence of past neural states on current and future states. Providing empirical support for this concept, Smit et al. (2013) demonstrated that individuals with higher LRTC scaling showed a greater tendency to carry forward timing errors when attempting to maintain a steady rhythm in a fixed tapping task. Stronger temporal dependencies hence seem to lead to the carryover of past information. In the current study, participants are required to flexibly adapt to the demands of the task on a trial-by-trial basis, with no cues at the beginning of each trial to indicate whether it will be an ignore, update, or control trial. Only during the interference phase do participants become aware of the specific condition. In this adaptive task environment, stronger LRTC—reflecting greater signal “memory”—may hinder overall performance by causing interference between trials (also see Euler et al., 2016). For example, features of an item seen in a previous trial may be carried over to the next trial, interfering with the mental representation of the current target item. In line with this reasoning, the negative effect of stronger LRTCs becomes particularly evident in maintenance-heavy conditions (ignore and controls) where discarding redundant information, such as presented distractors or (more generally) items seen in previous trials, is essential. Greater carryover of past information (i.e., greater signal memory) due to stronger LRTCs might thus disproportionately impact accuracy in these conditions. Updating, on the other hand, might not be affected by LRTCs, as it relies more heavily on rapid adaptation to changing information. The inherently adaptive nature of our task might prime an overly adaptive state, potentially rendering the influence of intrinsic signal memory (as represented by LRTC) less relevant for updating processes. Conversely, or in addition, the phasic increase of dopamine associated with updating may facilitate the selection of relevant information while suppressing outdated stimuli. This might counteract the potential interference caused by heightened signal memory, effectively reducing the negative impact of LRTCs on updating performance.

Interestingly, additional exploratory analyses examining the interaction between LRTCs and PSD slope suggest that the negative impact of strong LRTCs might be particularly detrimental to task accuracy in individuals with a steep PSD slope. Follow-up analyses revealed that this group generally exhibited higher DFA exponents, reflecting stronger LRTCs. The latter observation aligns with the network models proposed by Poil et al. (2012), which suggest that a lower E/I ratio generally results in temporal correlations closer to criticality (see Fig. 4A in Poil et al., 2012). In the context of our adaptive task, the combination of an inhibition-dominated network and stronger signal memory may lead to increased rigidity, which could explain the detrimental impact on task performance observed in this group. In contrast, participants with a flatter PSD slope exhibited lower baseline DFA exponents, indicating weaker LRTCs. The combination of weaker LRTCs (less signal memory) and a more excitable network (flatter PSD slope) may enable better adaptability by reducing the carryover of information from previous trials and enhancing responsiveness to new stimuli. In this more flexible neural state, the strength of LRTCs did not significantly affect task accuracy, suggesting that greater intrinsic network flexibility reduces the relevance of long-range dependencies.

Overall, it hence seems that a specific range of LRTC values can be either favorable or detrimental system dynamics, depending on the particular network predispositions and context (see Euler et al., 2016; N. D. Herzog et al., 2021). Such interpretations should be taken with caution, however, as this was an exploratory analysis. Nevertheless, our results reinforce the notion that, despite potential state-dependent variations, resting-state LRTCs possess trait-like qualities, providing meaningful predictions of cognitive performance. Future research could build on these findings by manipulating the excitation–inhibition balance, potentially through pharmacological interventions, to investigate how changes in E/I balance affect the impact of LRTCs on cognition. Understanding these mechanisms in greater detail might inform therapeutic approaches to improve cognitive function in clinical populations.

Contrary to our initial expectations, we found no significant direct association between PSD slope and condition-dependent task performance. However, when considering blood amino acid ratio, condition-specific effects of the PSD slope emerged. Specifically, for individuals with high amino acid ratios, there was a negative correlation between PSD slope and task performance in the three maintenance-related conditions (ignore, control short, and control long). There was no such relationship for low amino acid ratio individuals. The effect sizes for the significant interactions ranged from β = -1.73 to β = -3.51, indicating that even modest changes in amino acid ratio and PSD slope are associated with substantial decreases in task performance. Again, control analysis indicated no significant contribution of alpha power to these effects. Notably, the interaction effect was again absent in the update condition, reaffirming the maintenance specificity of cortical E/I dynamics. We speculate that our findings can be understood in light of the inverted U-shaped relationship between dopamine and PFC-dependent cognition (e.g., Cools & D’Esposito, 2011): while intermediate dopamine levels foster good cognitive performance, extreme levels—both high and low—are associated with performance deterioration. A high amino acid ratio suggests an increased presence of dopamine precursors phenylalanine and tyrosine in the bloodstream (Moja et al., 1996), indicating heightened dopamine synthesis capacity and, consequently, elevated dopamine levels in the brain (Barrett & Leyton, 2004; Coull et al., 2012; Leyton et al., 2004; Montgomery et al., 2003). Comparing amino acid ratio ranges in our data with those from studies utilizing similar calculation methods (e.g., from our laboratory: Hartmann et al., 2020, 2023; or externally: Montgomery et al., 2003), the high amino acid group in our sample appears to fall further right on the inverted U-shaped curve (Fig. 6). When dopamine levels are high, the PFC tends to operate in a D1-dominant state, which promotes heightened network stability (Durstewitz et al., 2000). This enhanced stability makes the neural circuits better equipped for maintaining working memory more resistant to distractions. However, this can come at a cost if inhibition becomes excessive, as indicated by a steeper PSD slope. The combination of a steep PSD slope and high amino acid ratios may, therefore, contribute to an “over-inhibition” state, resulting in an excessive suppression of neural activity. Given the adaptive nature of our task, as mentioned earlier, such a state could be detrimental to performance. At more middle range dopamine levels, however, variations in network E/I ratio might not have such a significant impact, as the system may be less prone to circuit stability.

Taken together, our findings hence indicate that the high amino acid ratio individuals in our sample are more sensitive to changes in network E/I ratio. This interpretation aligns with recent research indicating that excessive dopamine might disrupt the delicate balance between excitation and inhibition in the PFC, potentially via GABAergic signaling (see Domenico & Mapelli, 2023).

Given the specificity of our LRTCs and PSD slope effects to maintenance-related conditions, we anticipated that they might primarily be attributable to the dorsolateral prefrontal cortex (dlPFC). Indeed, source-level analysis partially confirmed the involvement of the dlPFC, with particularly pronounced effects in the right hemisphere. However, effects also extended to parietal areas, indicating the potential involvement of the frontoparietal attention network (FPN; e.g., Scolari et al., 2015), a key system in cognitive control and working memory processes. Our findings are hence in line with previous studies suggesting the involvement of FPN in a variety of tasks that require phasic adaptive cognitive control (Eriksson et al., 2015; Mahjoory et al., 2019; T. P. Zanto & Gazzaley, 2013). Strikingly, the topography observed in our data largely resembles that reported by Mahjoory et al. (2019), who also investigated the effects of resting-state LRTCs on working memory. Both, our study and that of Mahjoory et al. (2019), found a dominance of the right hemisphere, suggesting a consistent pattern of LRTC involvement in working memory processes across different experimental paradigms. This consistency strengthens the reliability of our findings and indicates a robust relationship between LRTCs in the frontoparietal network and working memory processes.

For PSD slope, the effects also spread to the left supplementary motor area (SMA), and some effects seemed to originate from the posterior cingulate cortex (PCC). The potential involvement of the PCC is particularly noteworthy, given its extensive interconnections with subcortical regions, including the basal ganglia (Vogt, 2019). This finding is intriguing as it aligns well with our observation of an interaction between PSD slope and amino acid ratio, that is dopamine synthesis capacity. As previously discussed, dopamine signaling in the basal ganglia is crucially involved in WM gating (e.g., Cools & D’Esposito, 2011; Frank et al., 2001). The PCC’s involvement may thus represent a potential bridge between cortical dynamics (as measured by PSD slope) and subcortical dopaminergic processes influencing WM performance. The exploratory nature of our source-level analysis and the absence of multiple testing correction necessitate a conservative approach to drawing conclusions, however. Yet, the consistency between our sensor-level and source-level results lends credibility to these exploratory findings.

Our findings significantly advance our understanding of the neural mechanisms underlying working memory gating processes. By demonstrating the role of resting-state cortical E/I dynamics in maintenance-related working memory processes, we provide new insights into how the brain manages specific aspects of information processing within the broader context of working memory gating. Our findings suggest that cortical E/I dynamics particularly influences the maintenance, rather than the updating component of gating. Moreover, our results challenge the broad assumption that stronger LRTCs would always be beneficial for cognitive functioning (also see Euler et al., 2016; Fusca et al., 2023; Hardstone et al., 2012; Jäger et al., 2024), implying that they are either favorable or detrimental system dynamics, depending on the particular network predispositions and/or cognitive demands. Furthermore, our PSD slope findings underscore the importance of a delicate network E/I ratio for optimal cognition, complementing the established role of dopamine levels in working memory. These results have potential implications for understanding cognitive deficits in neuropsychiatric disorders characterized by working memory dysfunction. They suggest that alterations in cortical E/I dynamics could contribute to these deficits, opening new avenues for targeted manipulation studies and, consequently, interventions.

The data and code used to produce the presented results will be available at https://github.com/O-BRAIN/WORMCRI upon publication.

N.H., L.K.J., E.C., and A.H. designed the study. N.H. and E.C. collected the data. N.H. analyzed the data. N.H. drafted the manuscript. Code review was done by P.S. Source-level analysis was done by N.K. E.C., P.S., N.K., S.J.F., V.N., A.V., L.K.J., and A.H. critically revised and approved the final manuscript.

This research was supported by the FAZIT Foundation, the “Wiedereinstiegsstipendium” from the University of Leipzig, and funding from the Max Planck Society. The author gratefully acknowledges this financial support.

No potential conflicts of interest relevant to this article were reported.

The authors thank Sylvia Stasch and Lisa Okhof for helping with recruitment of participants and data collection. Furthermore, we thank the “Institut für Laboratoriumsmedizin, Klinische Chemie und Molekulare Diagnostik” at the University Clinic Leipzig for the analyses of our blood samples.

Supplementary material for this article is available with the online version here: https://doi.org/10.1162/imag_a_00380.