Low frequency oscillations – neural correlates of stability and flexibility in cognition

The study addresses how the brain balances stability (maintenance of information) with flexibility (encoding and updating) in visuospatial working memory (vsWM). The authors propose that large-scale, low-frequency synchronization in theta (4–8 Hz) and alpha (8–14 Hz) bands could underpin distinct functional brain states that alternate during cognitive processing. Prior work shows low-dimensional spatiotemporal patterns across behaviors in animals and widespread theta/alpha activity in humans across perception and cognition. Low-frequency oscillations may regulate high-frequency spiking via phase–amplitude coupling (PAC), suggesting a mechanism for controlling information flow. The research aims to identify such synchronized networks, define brain states, relate state switching to cognitive performance, and test mechanistic hypotheses using an in-silico whole-brain model.

Prior literature indicates that large-scale low-frequency activities (theta/alpha) contribute to diverse cognitive processes, including eye-movement control and complex tasks. PAC between low-frequency phase (theta/alpha) and gamma amplitude has been observed in M/EEG and intracranial recordings, supporting the idea that slow networks gate high-frequency information flow. Low-dimensional cortex-wide activation motifs have been reported in mice, and human studies link oscillatory synchronization to perception and working memory capacity. Dorsal alpha synchronization has been associated with vsWM maintenance and top-down visuospatial attention, pre-saccade activity, and motor planning—functions requiring stable spatial representations. Occipital theta responses early after stimulus onset relate to visual processing and may modulate detection thresholds and reaction times. Interhemispheric connectivity has been linked to vsWM manipulation, suggesting potential functional distinctions between dorsal theta and alpha. Frontal theta is often observed in verbal WM with M/EEG, and primate recordings show frontal theta during vsWM; however, sensor limitations in human MEG may hinder detection. These lines of evidence motivate examining theta/alpha networks as state mechanisms balancing stability and flexibility.

Data: Two primary MEG datasets and one additional distractor dataset were analyzed. The HCP dataset included 83 adults (22–35 years) performing an N-back task (0-back, 2-back) during MEG; each stimulus was presented for 2000 ms followed by a 500 ms fixation, in blocks of faces or tools. A 4-subject dataset comprised repeated measures over seven sessions (days 1–39) with two vsWM tasks—WM-Grid (4×4 grid, sequences of 5 or 6 positions; 300 ms stimuli, 1000 ms delays) and Odd One Out (sequences of 5 stimuli with three shapes; remember odd-shape positions; 300 ms stimuli, 1000 ms delays)—plus a verbal recognition control task (listen for letter Q; matched timing). A third distractor dataset included 13 analyzable participants performing a sequential vsWM task with distractor vs no-distractor trials (four 500 ms bars with 500 ms delays; bars 2–3 distractors in distractor trials). MEG processing: Anatomical reconstructions (FreeSurfer) were parcellated into 200 Schaefer atlas regions. Source reconstruction used dSPM-based MNE with fixed-orientation dipoles (5 mm spacing), regularized noise covariance (estimated per dataset), and optimized collapsing to parcels. Morlet wavelets at 6 Hz (theta) and 10 Hz (alpha) were applied (5 cycles), downsampled, and edge windows removed. Synchronization was quantified using the imaginary part of phase-locking values. Network identification: Within each subject and frequency band (theta, alpha), independent signals were separated via ICA on filtered trial data (HCP: whole trial 0–2.5 s; 4-subject: 0–6.5 s), extracting 2–5 components. Linear ICA weight vectors (200 parcels) defined network spatial maps; higher weights indicated stronger regional contributions. Four consistent networks were identified: posterior theta, dorsal theta, posterior alpha, and dorsal alpha; a late-emerging frontal theta component in the 4-subject dataset was excluded from main analyses. Reliability was assessed via cosine similarity across sessions, tasks, and subjects with one-sided t-tests. State definition: Timepoints were clustered into four states using k-means on ICA timeseries: State 1 (posterior theta-dominant), State 2 (posterior alpha-dominant), State 3 (dorsal alpha-dominant), State 4 (dorsal theta-dominant). The proportion of subjects sharing states over time was quantified. Behavioral associations: In HCP, reaction time in the 2-back task and performance in a cognitive battery (Flanker, Card Sort, Processing Speed; PCA-derived PC1) were related to a state-switching rate (number of transitions per trial) using quadratic regression. Task-phase and load effects: In the 4-subject dataset, state activity during stimulus (encoding) vs delay (maintenance) periods was analyzed, with a verbal recognition control task to test specificity of state switching. Load-dependency of time in States 1 and 3 was regressed against WM load (WM-Grid and Odd One Out). Distractor control: A third dataset assessed internal regulation of State 1 (posterior theta) by comparing time in State 1 during identical stimuli in distractor vs no-distractor trials (time windows for stimuli 2–4), using within-subject differences and one-sided paired t-tests. Computational model: A whole-brain in-silico model used structural connectivity and distances from MICA-MICS (200 cortical nodes plus one basal ganglia–thalamus node per hemisphere, combined into a single BG-thalamus node per hemisphere for output). Each cortical node had an oscillatory layer (Kuramoto model; ω=0.04·2π, k=0.22, delays τ=2·D, dt=0.1 ms, white noise var=0.0025) and a spiking layer (rate model; du/dt = Σ C u(t-τ) − a·u + I, with a=0.25; dt=0.1 ms). PAC modulated spiking by oscillatory phase: duPAC/dt = u·(−0.5 sinθ + 0.5). Synchronized networks were generated by increasing thalamic oscillatory input selectively to posterior or dorsal regions, and by modulating subcortical oscillation frequency to target theta or alpha. Information flow was quantified via transfer entropy (TE) computed from normalized, binned time series over delays 0.1–40 ms. Conditions tested: stimulation of V1 or IPS under posterior vs dorsal synchronization (10 Hz), and frequency effects (theta 6 Hz vs alpha 10 Hz) within posterior and dorsal networks. Statistical tests used one-sided or two-sided t-tests as appropriate, with FDR correction for multiple comparisons in node-wise analyses.

- Four large-scale synchronized networks were identified in theta and alpha bands: posterior theta, dorsal theta, posterior alpha, dorsal alpha (Fig. 3). - Reliability of network spatial maps (cosine similarity) was significant across sessions (t=13.2, df=77, p<1e−21), tasks (t=37.1, df=154, p<1e−78), and subjects (t=23.7, df=58, p<1e−31); median similarities (SD) included posterior alpha across tasks 0.87 (0.21) and dorsal alpha across sessions 0.75 (0.21) (Table 1). - State dynamics: Up to ~90% of HCP subjects were in the same state at a given time; on average, ~65% were in the dominant state (Fig. 4). - Cognitive performance association: A quadratic relationship between state-switching rate and 2-back reaction time (β_SD=0.24, 95% CI [0.08, 0.41], t=2.88, df=80, p=0.0051) with optimal performance around nine switches; similarly for PC1 from cognitive tests (β_SD=0.36, 95% CI [0.13, 0.58], t=3.18, df=80, p=0.0021), optimal at nine switches (Fig. 5). - Encoding vs maintenance states: During stimulus presentation (encoding), State 1 (posterior theta) was most active; during delays (maintenance), State 3 (dorsal alpha) dominated. The verbal recognition control task did not show systematic switching between States 1 and 3 (Fig. 6). - Load effects: In WM-Grid, time in State 1 during stimuli decreased linearly with load (β_SD=0.79, 95% CI [0.57, 1.01], t=7.33, df=33, p<1e−7); in Odd One Out, a non-significant similar trend (β_SD=0.22, 95% CI [−0.16, 0.59], t=1.17, df=28, p=0.25). During delays, State 3 showed a quadratic relationship peaking at load 3 for both WM-Grid (β_SD=1.00, 95% CI [0.80, 1.19], t=10.33, df=39, p<1e−11) and Odd One Out (β_SD=0.86, 95% CI [0.58, 1.14], t=6.19, df=33, p<1e−8) (Fig. 6). - Distractor manipulation: Time in State 1 differed between distractor and no-distractor trials for identical stimuli (mean difference across windows=128.2 ms, 95% CI [43.8, ∞], t=2.73, df=12, p=0.009, d=0.79), consistent with internal regulation and load-dependent gating (Fig. 7). - Modeling—network generation: Increasing thalamic input selectively synchronized target networks; within-network synchronization exceeded outside by ~63% (posterior: mean diff=0.08, 95% CI [0.06, ∞], t=10.20, df=9, p<1e−5, d=3.40; dorsal: mean diff=0.09, 95% CI [0.06, ∞], t=14.81, df=9, p<1e−7, d=4.94) (Fig. 8). - Modeling—information routing: TE from V1 was 201% (SD=80%) larger under posterior vs dorsal synchronization; TE from IPS was 207% (SD=28%) larger under dorsal vs posterior synchronization (posterior: mean TE diff=0.018, 95% CI [0.013, ∞], t=7.83, df=9, p<1e−4, d=2.61; dorsal: mean TE diff=0.021, 95% CI [0.019, ∞], t=26.88, df=9, p<1e−9, d=8.88). Increases were strongest within synchronized networks vs rest (posterior: mean diff=0.018, 95% CI [0.013, 0.023], t=8.60, p<1e−4; dorsal: mean diff=0.019, 95% CI [0.016, 0.021], t=15.95, p<1e−7) (Fig. 9). - Modeling—frequency effects: Posterior theta vs alpha showed no TE difference (p=0.85); dorsal theta increased TE relative to dorsal alpha (mean diff=0.011, 95% CI [0.008, 0.013], t=11.64, df=9, p<1e−6, d=3.88). Dorsal alpha reduced interhemispheric transfer but increased TE to specific regions (e.g., frontal eye field and ipsilateral posterior cortex) (Fig. 9).

The findings support that cognition during vsWM involves alternating brain states characterized by large-scale synchronization in theta and alpha bands. Optimal control of state transitions relates to performance, indicating that balancing flexibility (encoding) and stability (maintenance) is behaviorally advantageous. State 1 (posterior theta) aligns with early visual processing and flexible sampling, dominating during stimulus presentation and showing load-dependent downregulation consistent with input gating; distractor trials further demonstrate internal regulation independent of stimulus identity. State 3 (dorsal alpha) reflects stable visuospatial representations, strongest during delays and peaking at load three, consistent with classical vsWM capacity limits. The computational model provides mechanistic support: selective synchronization via thalamic input facilitates information routing through PAC, with spatially targeted networks optimizing transfer from relevant sources (posterior synchronization boosts V1 outputs; dorsal synchronization boosts IPS outputs). Frequency modulates routing in dorsal networks (theta favoring broader, including interhemispheric transfer; alpha prioritizing intrahemispheric frontoparietal pathways), suggesting differential roles in manipulation vs maintenance. Posterior theta vs alpha distinctions remain less clear experimentally and computationally, highlighting an area for further investigation. The U-shaped performance relationship with state switching resembles dopamine’s inverted-U effects on cognitive control, pointing toward potential neuromodulatory influences and fronto–basal ganglia–thalamic mechanisms in orchestrating state transitions.

This work identifies four low-frequency synchronized networks that define functional brain states during vsWM and demonstrates that optimal switching among these states predicts better cognitive performance. Posterior theta emerges as an encoding state, and dorsal alpha as a maintenance state, both modulated by task demands and load. An in-silico model with PAC shows that synchronized networks can causally route information based on spatial topology and frequency, providing a mechanistic account of how flexibility and stability are achieved. Future research should clarify distinctions between posterior theta and alpha functions, investigate neuromodulatory and subcortical (pulvinar/basal ganglia–thalamus) control of state transitions, refine computational models to test alternative coupling and network configurations, and examine generalizability to other cognitive domains and larger, more diverse samples.

- MEG signal-to-noise limitations, particularly for frontal sensors, may hinder detection of frontal theta components relevant to vsWM. - Transfer entropy analyses require long continuous data segments; practical constraints and SNR issues in real MEG limit direct application to fast state changes, motivating reliance on simulations. - The 4-subject repeated-measures dataset is small; although reliability was assessed, generalization warrants larger cohorts. - Network identification within subsets was incomplete (63%), indicating potential variability in detectability across sessions/tasks. - Computational model simplifications (e.g., combined basal ganglia–thalamus node, parameter choices, PAC implementation) may not capture all biological complexities; posterior theta vs alpha effects were not differentiated in TE. - Frontal theta network emerged with practice in the 4-subject dataset but was excluded, leaving questions about training-related signals and their relevance to core states.