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

A central question in cognitive neuroscience is how the brain balances stability and flexibility, for example during eye movements and visuospatial working memory (vsWM) where encoding/updating requires flexibility and maintenance requires stability. The authors posit that alternating large-scale functional states, implemented through low-frequency synchronization, could support these demands. Prior work shows low-dimensional, large-scale spatiotemporal patterns in animals and humans, with prominent theta (4–8 Hz) and alpha (8–14 Hz) rhythms. Low-frequency oscillations can modulate high-frequency spiking via phase-amplitude coupling (PAC), potentially controlling information flow. The study aimed to characterize low-frequency synchronized networks underlying cognitive states, relate their temporal transitions to behavior, map states to encoding vs maintenance phases, and test how spatial distribution and frequency of synchronization impact information routing using an in-silico model with realistic connectivity and PAC.

Previous studies have reported large-scale low-frequency activity linked to perception and cognition, including theta/alpha-band synchronization across cortex. M/EEG and intracranial data demonstrate PAC between theta/alpha phase and gamma amplitude, suggesting low-frequency rhythms can gate information in higher frequencies. Alpha-band synchronization has been linked to vsWM capacity and top-down visuospatial attention; occipital theta/alpha activity has been tied to visual processing and detection thresholds. Interhemispheric connectivity relates to vsWM manipulation. Frontal theta often appears in verbal WM maintenance and in primate vsWM recordings, though its presence in human MEG vsWM is less consistent, potentially due to SNR and training effects. These findings motivated the hypothesis that low-frequency networks define functional states governing flexibility (encoding) and stability (maintenance) via dynamic routing mechanisms.

Datasets: (1) HCP MEG: 83 adults performing 0-back/2-back WM task (stimuli 2000 ms, 500 ms fixation; blocks by category). (2) 4-subject longitudinal MEG: four participants scanned over seven sessions across eight weeks performing two vsWM tasks (WM-Grid and Odd One Out; 300 ms stimulus, 1000 ms delay) and a control verbal recognition task matched in timing. (3) Distractor vsWM MEG: 17 participants (13 analyzed) performing sequential memory with distractor trials (four bars; in distractor trials, items 2 and 3 not to be remembered). MEG acquisition/preprocessing: Whole-head Elekta Neuromag TRIUX (NatMEG); tSSS (MaxFilter), ICA for artifact removal, notch filters at 50 Hz and harmonics, downsampling (HCP to 508.63 Hz; 4-subject to 250 Hz), epoching with buffer windows to minimize edge effects. Source reconstruction: FreeSurfer cortical reconstructions, Schaefer 200-region parcellation, dSPM MNE with fixed-orientation dipoles (5 mm spacing), optimized collapse operator to parcel level, noise covariance estimated per dataset/task and regularized. Time-frequency: Morlet wavelets at 6 Hz (theta) and 10 Hz (alpha), 5 cycles, downsampled to 5× center frequency; edge windows removed. Synchronization measures: global synchronization computed via imaginary part of phase-locking value; network-specific signals assessed post-ICA. Network identification: For each subject and frequency band, independent components (2–5 per subject/frequency) were extracted (scikit-learn ICA) from filtered trial data (HCP: 0–2.5 s; 4-subject: 0–6.5 s). Linear ICA weights (200 cortical regions) defined spatial maps of synchronized networks. Four robust networks emerged: posterior theta, dorsal theta, posterior alpha, dorsal alpha. Reliability: In the 4-subject dataset, cosine similarity of ICA weight vectors was compared across sessions, tasks, and subjects (controlling the other factors), identifying 63% of networks when partitioned by task and session. Clustering into states: Concatenated ICA time series were clustered using k-means (k=4) within each dataset; each time point assigned to one of four states dominated by a single network: posterior theta (state 1), posterior alpha (state 2), dorsal alpha (state 3), dorsal theta (state 4). Behavioral analyses: In HCP, reaction time (2-back) and a cognitive battery (flanker inhibitory control, Dimensional Change Card Sort, Pattern Comparison Processing Speed) were analyzed; PCA on the three tests yielded PC1 as a general performance measure. State-switching rate was defined as the number of transitions between states per trial; quadratic relationships with RT and PC1 were modeled using standardized coefficients. Encoding vs maintenance: In the 4-subject dataset, time spent in states during stimulus presentations (encoding) vs delays (maintenance) was quantified; load-dependency analyzed via regressions across sessions. Control verbal task tested specificity. Distractor dataset: Differences in time spent in the encoding state (state 1) for identical stimuli between distractor vs no-distractor trials were tested with one-sided paired t-tests, aggregating within-subject across stimuli 2–4. In-silico model: A 202-node network (200 cortical Schaefer regions + bilateral basal-ganglia–thalamus nodes merged per hemisphere) with subject-specific structural connectivity and distances from 10 MICA-MICS subjects. Two layers per node: (i) oscillatory Kuramoto layer (dθ/dt = ω + k Σ Cnp sin(θp(t−τnp)−θn), with delays τnp=2Dnp, noise variance 0.0025, dt=0.1 ms) where thalamic drive to targeted cortical nodes increased coupling and set synchronization frequency (theta or alpha); (ii) spiking layer with rate dynamics (du/dt = Σ Cnp up(t−τnp) − a un + In, a=0.25) and global strength (w=0.8), influenced by oscillations via PAC: duPAC/dt = un(−0.5 sin θn + 0.5), enhancing spiking near oscillatory troughs. Information flow quantified by transfer entropy (TE) from stimulated nodes (V1 or IPS) to others, computed from normalized, binned time series, integrating delays 0.1–40 ms. Simulations run per structural subject and averaged; significance via paired t-tests (one- or two-sided as hypothesized) with FDR where appropriate.

- Four large-scale networks in theta and alpha bands were identified across datasets: posterior theta, dorsal theta, posterior alpha, and dorsal alpha (MEG ICA-derived spatial maps; Fig. 3). - Reliability (4-subject dataset) of network spatial maps was significant across: tasks (t=37.1, df=154, p<1e−78), sessions (t=13.2, df=77, p<1e−21), and subjects (t=23.7, df=58, p<1e−31); median cosine similarities reported (e.g., posterior alpha across tasks: 0.87 (SD 0.21)). - Temporal clustering revealed four states dominated by each network; up to ~90% of HCP subjects were in the same state at certain times; average ~65% in the dominant state at a given time (Fig. 4). - State switching rate exhibited a U-shaped relationship with performance (HCP): quadratic effect on 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 at ~9 switches; quadratic relation with PC1 (β_SD=−0.36, 95% CI [−0.58, −0.13], t=−3.18, df=80, p=0.0021), again optimal at ~9 switches (Fig. 5b–c). - Encoding vs maintenance mapping (4-subject dataset): during encoding, state 1 (posterior theta) dominated; during maintenance, state 3 (dorsal alpha) dominated (Fig. 6a). The control verbal recognition task did not show systematic switching between these states (Fig. 6b–c). - Load effects: During stimulus presentation, time in state 1 decreased linearly with load in WM-Grid (β_SD=0.79, 95% CI [−1.01, −0.57], t=7.33, df=33, p<1e−7); no significant effect in Odd One Out (β_SD=0.22, 95% CI [−0.59, 0.16], t=−1.17, df=28, p=0.25). During delay, time in state 3 showed a quadratic relationship peaking at load 3 in WM-Grid (β_SD=−1.00, 95% CI [−1.19, −0.80], t=−10.33, df=39, p<1e−11) and Odd One Out (β_SD=−0.86, 95% CI [−1.14, −0.58], t=−6.19, df=33, p<1e−6), with alpha synchronization ceiling beyond load 3. - Distractor dataset: Despite identical stimuli, time in state 1 differed between distractor and no-distractor trials as predicted (shorter for distractors on stimuli 2–3, longer on stimulus 4), mean within-subject difference across three periods=128.2 ms (95% CI [43.8, ∞]), t=2.73, df=12, p=0.009, d=0.79 (one-sided). - In-silico model: Thalamic drive selectively synchronized targeted cortical networks (posterior or dorsal). Synchronization inside vs outside network increased by ~63% (posterior SD=15.7%; dorsal SD=9.3%): posterior mean difference=0.08 (95% CI [0.06, ∞]), t=10.20, df=9, p<1e−5, d=3.40; dorsal mean difference=0.09 (95% CI [0.06, ∞]), t=14.81, df=9, p<1e−7, d=4.94. - Synchronization gated information flow (TE): With 10 Hz synchronization, TE from V1 to brain was 201% (SD=80%) larger under posterior vs dorsal synchronization; TE from IPS was 207% (SD=28%) larger under dorsal vs posterior synchronization. Statistics: posterior condition mean TE difference=0.018 (95% CI [0.013, ∞]), t=7.83, df=9, p<1e−4, d=2.61; dorsal mean TE difference=0.021 (95% CI [0.019, ∞]), t=26.88, df=9, p<1e−9, d=8.88. Increases were strongest within the synchronized network (paired comparisons p<1e−4 and p<1e−7, respectively; Fig. 9b). - Frequency effects: Posterior network TE did not differ between theta (6 Hz) and alpha (10 Hz) (p=0.85). Dorsal network TE was higher under theta than alpha (mean difference=0.011, 95% CI [0.008, 0.013], t=11.64, df=9, p<1e−6, d=3.88), with dorsal alpha showing reduced interhemispheric transfer but increases to specific areas (e.g., frontal eye field, ipsilateral posterior cortex) (Fig. 9c–d).

The study demonstrates that large-scale synchronized networks in the theta and alpha bands define functional brain states during vsWM. Optimal control of transitions among these states correlates with better performance, supporting the hypothesis that dynamic state selection balances cognitive flexibility and stability. Mapping states to task phases showed that posterior theta supports encoding (flexibility), while dorsal alpha supports maintenance (stability), and both are modulated by working memory load. The distractor dataset indicates internal regulation of the encoding-related posterior theta state beyond stimulus-driven effects. Computational modeling provides a mechanistic account: thalamically-driven cortical synchronization creates spatially selective oscillatory networks that, via PAC, bias routing of high-frequency spiking information within and between regions. Information flow depends on both where synchronization occurs and its frequency, with dorsal theta facilitating broader (including interhemispheric) transfer than dorsal alpha, consistent with distinct roles in maintenance vs manipulation. Together, these results suggest a cognitive control mechanism where selective, frequency-specific large-scale networks dynamically gate information flow to meet task demands.

This work identifies four reproducible low-frequency cortical networks whose temporal dominance defines brain states during visuospatial working memory. State transition rates exhibit a U-shaped relationship with performance, and specific states map onto encoding (posterior theta) and maintenance (dorsal alpha), each modulated by cognitive load. An in-silico model grounded in realistic connectivity shows how thalamically-driven synchronization and PAC can route information selectively by space and frequency, offering a mechanistic explanation for flexible vs stable processing. Future research should clarify distinctions between posterior theta and alpha functions in visual processing, investigate the neural control of state transitions (e.g., fronto-basal ganglia and pulvinar contributions), examine conditions eliciting frontal theta in human vsWM, and extend models to incorporate richer subcortical dynamics and task-dependent neuromodulation.

- MEG-based detection of high-frequency spiking-related activity is limited by signal-to-noise, necessitating reliance on low-frequency measures and simulations for information flow estimates. - Not all networks were identifiable in every subject/session/task partition (63% identified in fully partitioned subsets), potentially reducing generalizability of spatial maps. - Frontal theta synchronization commonly reported in verbal WM and primate vsWM was not robustly observed here, possibly due to MEG SNR and task domain differences; a frontal theta component emerged only after repeated sessions in the 4-subject dataset. - The posterior theta vs alpha functional distinction remains unresolved experimentally; the model did not show frequency-dependent transfer differences in posterior regions. - The computational model simplifies subcortical interactions (basal ganglia–thalamus combined and treated as a control input) and assumes unidirectional PAC influence; parameters may not capture all biological complexities. - Sample size for longitudinal within-subject analyses was small (n=4), although repeated sessions mitigate some concerns; the distractor analysis included 13 subjects. - State-switching performance relationships, though robust, are correlational; causal manipulation of states was not performed in human data.