Shared spatial and temporal principles govern connectome dynamics across timescales

Abstract

While the brain processes information at various speeds, little is known about how the functional connectome can concurrently support multiple speeds in parallel. FMRI and electrophysiological modalities have been used to study connectome dynamics at slow and fast speeds, respectively. But it is often implicitly assumed that these modalities capture the same underlying neural processes, with fMRI doing so through a low-pass temporal filter. However, recent work suggests the alternative possibility that connectome dynamics comprise distinct processes operating at multiple timescales. If such multiscale connectivity processes indeed coexist, a key question is whether their patterns and sequences are organized on the basis of shared regularities, i.e., common spatial and temporal principles.

In simultaneous human fMRI and source-localized EEG, we investigated the connectome’s foundational constituents—the instantaneous co-activation patterns—across six timescales of neural activity, from infraslow through γ-band. We found streams of recurrent co-activation patterns, or states, that operate in parallel and asynchronously across timescales, thereby forming a timescale-overarching spatial principle. These states also occurred in highly similar sequences at all timescales, revealing a timescale-overarching temporal principle.

Together, these findings indicate that the connectome comprises multiple dynamic streams operating in parallel at distinct speeds, from tens of milliseconds to seconds, rather than a single stream filtered by each modality’s temporal resolution. Spatial and temporal principles that span these streams enable their integration into a unified system. Consequently, research on human behavior and mental disorders should account for the full range of the connectome’s timescales.

Introduction

The large-scale functional connectome is the whole-brain pattern of cross-region connectivity that provides the intrinsic architecture for distributed neural processing (1, 2). Early investigations of this architecture in humans marked a watershed moment in neuroscience history, catalyzing a line of research that remains central to date (3–5). Using fMRI, these studies offered an unprecedented glimpse into the intricate network of time-varying connections within the human brain in both health (6, 7) and disease (8–10). However, fMRI’s temporal resolution limits investigation of functional connectivity at the rapid timescales over which neural systems process information (11, 12).

Consider the example of speech comprehension. The brain processes phonemes within tens of milliseconds, words over seconds, sentences-level themes over even slower timescales (13). Critically, these processes must occur in parallel; without concurrent processing across timescales, speech comprehension—and many other cognitive functions—would fail (14).

Despite the inherently multi-scale nature of cognition, time-varying connectome dynamics are often treated as a single, unified stream of co-activation patterns—a byproduct of fMRI’s exquisite ability and dominance in capturing whole-brain connectivity (15–17). Consequently, far fewer empirical studies have considered connectome dynamics at timescales beyond those accessible to fMRI (18–23) (apart from in-silico generative models 24, 25). While infraslow fMRI has yielded valuable insights into the large-scale brain organization, it offers only a narrow window into the brain’s full temporal repertoire. Even when decomposed into distinct infraslow sub-bands (26–29), fMRI co-activation patterns capture only a fraction of the timescales at which population-level neural activity unfolds (30).

Studies using temporally resolved modalities, including MEG, EEG, and intracranial EEG, have revealed intrinsic connectivity networks (ICNs) that exhibit striking spatial correspondence with those derived from infraslow fMRI. These spatially coherent ICNs emerge not only in time-averaged (static) connectome architectures (e.g., MEG (31–33); concurrent EEG-fMRI (34–36)), but also in transient, short-lived connectome states that recur across time (e.g., MEG (18–20)).

Beyond establishing cross-modal spatial correspondence, concurrent multimodal studies have aimed to characterize when such correspondences arise by bridging the temporal resolution gap between electrophysiology and fMRI. To align fast neural activity with slower BOLD fluctuations, prior work has convolved the canonical hemodynamic response function (HRF) with EEG-derived independent components (35), spectra (35, 37–39), or state time courses (22), as well as used sliding-window approaches (21) (see review in 40). These efforts have substantially advanced understanding of neural dynamics jointly captured by fMRI and electrophysiological modalities. Conceptualizing the common and unique neural processes captured by fMRI and EEG/MEG, respectively, in a Venn diagram (Fig. 1A), these multimodal studies have primarily examined the overlapping region between modalities, emphasizing the shared spatiotemporal patterns across timescales.

Fig. 1. Integrative framework linking fast and infraslow neural processes through a blueprint-based connectome approach.

(A) Conceptual Venn diagram illustrating shared and unique contributions of fast and infraslow neural dynamics to connectome-level measurements. The overlapping region (light grey) denotes temporal co-occurrence of fast and infraslow neural events that become accessible through concurrent EEG-fMRI, such as previously established co-occurrence of individual ICNs in fMRI with spatially diffuse fluctuations (35, 37–39) or covariances (22) of the EEG spectrum. The non-overlapping regions (dark grey for fast electrophysiological; white for infraslow) encompass neural and non-neural processes unique to each timescale. Here, we test whether overarching spatiotemporal regularities (hatched areas) underlies some portion of neural processes spanning fast and infraslow domains. A straightforward example of shared spatial regularity is the observation of ICNs in EEG/MEG (18–20, 32, 33) that resemble those established in fMRI. However, it remains unclear whether some of these rapid ICN events in EEG/MEG reflect distinct neural processes underlying fMRI-defined ICNs (potentially occurring at non-overlapping timepoints), and whether individual ICNs optimally characterize spatial regularities shared across timescales. (B) Combinatorial framework for spatial regularities based on ICN combinations. Left: 68 cortical regions from an independent brain atlas (47), color-coded by seven canonical ICNs (48): visual (VIS), sensorimotor (SMN), dorsal attention (DAN), ventral attention (VAN), limbic (LIM), frontoparietal (FPN), and default mode (DMN). Top: Example blueprints projected onto the cortical surface, representing configurations with one to six co-active ICNs. Bottom: 68×126 matrix enumerating all possible ICN combinations, where rows correspond to ICNs (row length proportional to the number of assigned regions) and columns denoting blueprints, represented as binary vectors marking active (colored) and inactive (blank) regions.

However, such disproportionate focus on cross-modal convergence may obscure equally critical evidence pointing to additional dynamic connectome processes that temporally diverge (non-overlapping areas of the Venn diagram in Fig. 1A). While electrophysiological networks may transiently associate with fMRI ICNs, some evidence suggests that such associations may be non-uniform and temporally decoupled (22, 23, 32). Together, these observations raise an alternative possibility: that connectome dynamics unfold in multiple, asynchronous streams (23).

To probe these divergent processes, we shift focus to timescale-specific neural processes in the connectome dynamics that remain elusive when brain activity is treated as a single stream of co-activations shared across slow and fast data modalities. We hypothesize that these non-overlapping regions of the Venn diagram are not merely artifacts of modality-specific physiological and non-physiological noise or suboptimal transfer functions (e.g., the HRF), but are partly neural in nature (41).

Further, we consider the possibility that some—albeit certainly not all—of these timescale-specific neural processes are governed by spatial and temporal regularities shared across neural timescales (some portion of hatched area in Fig. 1A). This idea builds on a defining property of complex systems, including the brain (42, 43), whereby organizational principles recur across scales (44, 45). Accordingly, we postulate that the large-scale connectome dynamics are constrained by fundamental spatial and temporal principles that transcend the specific temporal resolution of a given modality (46).

To characterize spatial regularities, we identified recurrent connectome states across all timescales. We conceptualized a broad repertoire of dynamic connectome states (Fig. 1B) as a discrete and comprehensive set of 126 state blueprints (i.e., ${\sum }_{\left\{k=1\right\}}^{\left\{6\right\}}\left(\begin{array}{c}7\\ k\end{array}\right)=126$, where 7 represents the total number of ICNs and k indexes all non-empty subsets excluding the full combination $\left(\begin{array}{c}7\\ 7\end{array}\right)$, which covers the entire brain). Each blueprint is a binary vector (Fig. 1B; columns of the matrix), representing one of all possible combinations of the seven canonical ICNs (48), mapped onto 68 cortical regions from the Desikan-Killiany atlas (47). Accordingly, each blueprint corresponds to a unique co-activation pattern arising from a specific combination of ICNs (19, 48).

The rationale for using ICNs as the basis of a combinatorial state space across neural timescales builds on extensive evidence that these networks constitute internally cohesive and functionally stable units. ICNs exhibit stronger within-than between-network connectivity (48, 49), are reliably observed across both slow (fMRI) and faster (EEG/MEG) timescales (18, 19, 22, 50), and persist across diverse cognitive states, including rest (51). In this sense, ICNs represent one basic form of spatial regularity, common to fast and slow neural timescales. Therefore, we postulate that ICNs serve as building blocks of a broader connectome state repertoire underlying their moment-to-moment reconfigurations (52).

A key departure from prior work is our proposition that ICNs do not recur solely in isolation or within a limited set of favored combinations. Instead, we hypothesize that ICNs flexibly co-activate in all possible combinations (53), forming an exhaustive and discrete state space of dynamic whole-brain co-activation patterns. This combinatorial framework is motivated by theories positing that ICNs constitute cognitive architectures that pre-configure the brain for optimal processing of different cognitive demands (2). Through flexible reconfiguration, potentially occurring in parallel across timescales, these ICNs could support diverse mental operations over time (54).

To test this framework across timescales, we tracked moment-to-moment ICN co-activation patterns across six neural timescales, spanning from infraslow-fMRI, to δ-, θ-, α-, β-, and γ-band source-reconstructed EEG. This framework enables a unified investigation of the spatial and temporal principles that govern the connectome organization, independent of any one modality or timescale.

We formalize our hypothesis along three axes. First, we predict that instantaneous whole-brain co-activation patterns reduce to a discrete set of recurrent states that recurs across all neural activity timescales, constituting a timescale-overarching spatial principle. Second, we predict that transitions between these states are structured and non-random. We further propose that these sequences are preserved across timescales, revealing a timescale-overarching temporal principle. Third, we consider the possibility that these spatial states occur asynchronously across timescales. If identical states recur at different speeds, distinct states may be expressed concurrently across timescales, challenging the prevailing view of the connectome as a single unified stream of co-activations and instead supporting a parallel, multi-stream architecture of large-scale connectome dynamics.

If observed, these spatial and temporal regularities would offer strong empirical support for a timescale-overarching theory of connectome dynamics. Such a framework construes the connectome as a parallel, multi-stream process spanning the full range of timescales relevant to human cognition and behavior. Given that spontaneous fluctuations in large-scale co-activation patterns are behaviorally relevant in both humans (55–58) and animals (59, 60), we tested these proposed principles using a concurrent EEG-fMRI dataset acquired during resting state (N=26). We further tested the generalizability of our findings in two independent resting-state samples: an independent concurrent EEG-fMRI dataset (N=24) and a large-scale EEG-only dataset (N=443).

Results

Results with leakage correction applied via signal orthogonalization (63) are provided as a supplemental analysis (see Supplementary Information (SI) 2.1).

Canonical ICN combinations as spatial basis of connectome dynamics

This expository section outlines the conceptual framework linking neural timescales to the speed of connectome dynamics, as operationalized through the blueprint-fitting approach (Fig. 2). As illustrated in Fig. 2A, neural activity at each region (defined by the parcellation scheme) can be decomposed into frequency-specific components, which we refer to as neural activity timescales. For each frequency band, the corresponding power envelope captures slower amplitude fluctuations relative to the underlying neural signals. As connectome states (i.e., whole-brain co-activation patterns) are constructed from these envelopes, their transition rate (i.e., speed of connectome dynamics) is inherently coupled to the speed of the envelope fluctuations.

Fig. 2. Basic Concepts and Approach.

(A) Conceptual illustration linking neural activity timescales to the speed of connectome dynamics. We distinguish neural activity timescale (blue; the canonical frequency band (Hz) in which the timeseries was filtered) from the speed of power envelope (orange; the rate of transitions between distributed co-activation patterns). Amplitude envelope fluctuations evolve more slowly than the underlying signal and determine how long co-activation states between regions (e.g., A and B, bottom) persist or change over time. (B) Blueprint fitting procedure. Spontaneous co-activation patterns (z-scored BOLD or band-limited EEG amplitude across 68 regions) are compared to 126 predefined blueprints using spatial Pearson correlation. At each timeframe, the blueprint with the highest correlation is selected as the best-fitting state. (C) Blueprint state sequences across neural timescales. For each timescale, the upper panel shows correlation timeseries between four sample blueprints and empirical co-activation patterns over a 2.8 s window. Colored boxes indicate when each blueprint was best fitting. The lower panel shows the full sequence of best-fitting blueprints (1–126, color-coded as in Fig. 1B), illustrating progressively slower state transitions at slower neural timescale. (D) Speed of connectome state transitions. Mean lifetime of blueprint states, averaged across 126 states for each subject, increases systematically from ~20 ms to ~3000 ms. These lifetimes closely track the speed of regional amplitude envelopes and relate to the center frequency of canonical bands (orange and blue in 1A).

To identify the blueprint that best captures the spatial organization of brain activity at each time frame, we computed Pearson’s correlation between each blueprint and the instantaneous whole-brain co-activation pattern (Fig. 2B). These co-activation patterns were derived from z-scored BOLD signal for fMRI and from band-limited amplitude envelopes for source-localized EEG. At each time frame, the blueprint with the highest spatial correlation was selected as the best-fitting state, yielding a time-resolved sequence of discrete blueprint labels for each neural activity timescale (Fig. 2C).

As shown in Fig. 2C–D, blueprint state sequences revealed marked difference in transition speed across timescales. State lifetimes increased systematically from fast to slow neural activity timescales, with average durations of ~20 ms at the fastest EEG band and ~3000 ms in fMRI, spanning more than three orders of magnitude. These estimates are consistent with prior reports of state lifetimes in electrophysiological (18, 20, 22, 61) and hemodynamic data (50).

Spatial principles of connectome dynamics are shared across timescales

Fig. 3A illustrates that blueprint states close resemble empirical whole-brain co-activation patterns across timescales. At each timescale, time frames assigned to a given blueprint exhibited average spatial pattern that matched the corresponding blueprint, as would be expected if the best-fitting blueprint in fact explained the co-activations.

Fig. 3. Spatial principles governing co-activation patterns across neural timescales.

(A) Empirical fit of state blueprints to co-activation patterns. Leftmost: Blueprints 1–7, each representing a single ICN (colored as in Fig. 1B). For each neural timescale (left to right), carpet plots display z-transformed activity across 68 Desikan regions (rows) and all timeframes in which the given blueprint was identified as best-fitting across subjects (see Fig. 2C for procedure). Darker color indicates higher activity. Horizontal bands of elevated activity (red arrows) align with the ICNs active in the corresponding blueprint. Right of each carpet plot: Average co-activation pattern across these time frames closely resembles the corresponding blueprint, confirming consistent spatial expression across timescales. (B) Statistical validation of blueprint fit. For each subject, boxplots show the mean spatial correlation between empirical co-activation patterns and their best-fitting and second best-fitting blueprint. Boxplots indicate the median, interquartile range (IQR), and whiskers extending to 1.5×IQR; individual outliers are plotted. Empirical fits (green) are compared against two null models: spatially shuffled blueprints (NullLabelPermute; red) and phase-randomized regional time series (NullPhasePermute; purple). Spatial fit of the best-fitting blueprints exceeds the second best-fitting blueprints as well as both nulls across all timescales. (C-D) Replication across independent datasets. Superior blueprint fit relative to both null models is replicated in the independent concurrent EEG-fMRI dataset (C) and the large-scale EEG-only dataset (D). SI 2.1 for methodological and statistical details.

This spatial correspondence was quantified by comparing blueprint fits (i.e., spatial correlations with empirical co-activation patterns) against two null models (Fig. 3B; see SI): spatially scrambled blueprints (Null_LabelPermute; random permutation of Desikan region labels) and phase-randomized regional timeseries (Null_PhasePermute; independent phase randomization per region (62)). Across all timescales, blueprint fits were significantly higher than both null models (vs. Null_LabelPermute, t₍₂₅₎: 19.32–35.68; vs. Null_PhasePermute, t₍₂₅₎: 18.23–37.36; all p < .05/6). These effects replicated in a separate concurrent EEG-fMRI dataset (t₍₂₃₎ range: 22.47–46.32, all p < .05/6) and a large-scale EEG-only dataset (t₍₄₄₂₎ range: 83.09–109.01, all p < .05/6), both in resting state. All EEG effects remained significant after correcting for source-leakage via signal orthogonalization (63) (see SI 2.1).

We further demonstrated the specificity of blueprint assignments by comparing the spatial fit and temporal properties of the best-fitting and second best-fitting blueprint sequences. Beyond significantly higher spatial fit (all t₍₂₅₎ > 85, p < .05/6), the best-fitting blueprints exhibited substantially greater dissociation from null models than their second-best counterparts (all t₍₂₅₎ > 10, p < .05/6). The magnitude of this separation between the best and second best-fitting blueprints was substantial, with effect sizes reaching Cohen’s d ≥ 16 and remaining large after null correction d ≥ 2, suggesting that blueprint assignment is unlikely to be substantially driven by quantization noise (see SI 3.2.3). This conclusion was further supported by simulating the impact of noise on blueprint assignment (see SI 3.2.4). Notably, some degree of spatial similarity between top candidates is expected given the combinatorial nature of the blueprint space: many blueprints differ by the inclusion or exclusion of only one or two ICNs. Our findings demonstrate that the best-fitting blueprint is not a randomly selected among spatially similar blueprints or noise-driven fluctuations but demonstrate that blueprint assignment reflects structured biological processes rather than random selection among spatially similar alternatives.

Further, to ensure that observations in fMRI were not limited by EEG’s lower spatial resolution, we applied the blueprint fitting procedure to fMRI data parcellated using a higher-resolution atlas (300-region Schaefer atlas (64)). The results remained consistent: blueprint fits were significantly higher than both Null_LabelPermute (t₍₂₂₎ = 34.8) and Null_PhasePermute (t₍₂₂₎ = 35.43; both p < .05/6). We also confirmed that head motion did not inflate or obscure blueprint fits (see SI 3.2).

In summary, a key novelty of our framework lies in modeling co-activations across combinations of ICNs, rather than treating ICNs as isolated units. Using a predefined set of 126 blueprints, each timeframe was assigned the single best-fitting spatial configuration composed of up to six ICNs. Across time frames and subjects, blueprints involving multiple ICNs were selected more frequently, indicating that a wide range of ICN combinations, rather than single-network activations, contributes substantially to the spatial variance observed in co-activation patterns (see SI 2.1.3).

Strikingly, the set of expressed blueprints included co-activations between ICN traditionally considered less likely to co-activate—such as DAN×DMN (see Discussion) (65, 66). These findings challenge the notion of strict functional segregation between such systems and point instead to transient, spontaneous integration. Further, we tested whether ICNs included in a blueprint exhibit temporally coordinated fluctuations, computed by Pearson correlations, during the timepoints when that blueprint best fits the data. The results confirm that blueprint-assigned timepoints capture structured inter-network co-fluctuations rather than mere local amplitude changes, indicating connectivity-related dynamics (see SI 2.1.4).

Together, these findings demonstrate that the brain flexibly recruits a broad repertoire of ICN combinations, drawing from the nearly full combinatorial space afforded by the canonical networks (see Tables S3–4).

Temporal principles of connectome dynamics are shared across timescales

Having shown that co-activation patterns across all six neural timescales reflect combinations of canonical ICNs (i.e., state blueprints), we next asked whether temporal progression of these states follows non-random principles and whether such temporal structure is shared across timescales. We focused on three key properties of blueprint sequences: transition probabilities, fractional occupancies, and mean lifetimes.

Transition probability matrices exhibit structured, non-random organization across timescales (Fig. 4A, left). Relative to null transition matrices (generated by shuffling the temporal ordering of best-fitting blueprint states 100 times while preserving fractional occupancy), observed transition probability matrices were significantly more dissimilar from the null than expected by chance (t₍₂₅₎ range: 22.70–58.22; all p < .05/6), indicating consistent temporal regularity in state transitions.

Fig. 4. Temporal principles of blueprint state sequencing across neural timescales.

(A) State transition probabilities across timescales. Left: Representative transition probability matrices (126 × 126) at each of six neural timescales, showing likelihood of transitioning from one blueprint (row) to another (column). Diagonal-dominant, pointed oval-shaped structure indicates preferential transitions between spatially similar states (see SI 3.1.2). Overall symmetry reflects mirrored spatial structure between the first and last 63 blueprints (cf. Fig. 1B). While absolute transition values vary with state lifetimes (cf. Fig. 1D), relative transition structure is preserved across timescales (see SI 3.1.3). The infraslow matrix uses a different color scale to accommodate its longer lifetimes. Right: Cross-timescale similarity of transition matrices. Upper triangle (black-to-copper) shows pairwise Pearson correlations; lower triangle (black-to-white) shows t-statistics comparing empirical matrices to nulls. (B-C) Cross-timescale regularity in blueprint expression. (B) Fractional occupancy across all 126 states and timescales (colored as in Fig. 1B), revealing conserved “skyline” structure. Six vertical dashed lines mark representative blueprints with maximal occupancy within each ICN-count group (i.e., those with 1, 2, … ICNs active). As exemplified by these dashed lines, when a given blueprint occupies a particularly large proportion of the recording, its complement, i.e. the blueprint with the opposite set of ICNs being active/inactive, tends to do so as well. (C) Mean lifetime of each blueprint (±1 standard deviation across subjects). Both measures exhibit distinct, non-random patterns across blueprints. Rightmost matrices quantify cross-timescale similarity of the “skylines”, with correlations (upper triangles) and t-statistics versus nulls (lower triangles). Although absolute lifetimes differ markedly across timescales, the relative distribution across blueprints is preserved. Notably, correlations between infraslow and faster timescales are weaker, likely due to limited blueprint sampling in fMRI caused by slower switching dynamics.

We then asked whether this non-random transition structure was shared across timescales. Within-subject spatial correlations between transition probability matrices from all pairs of timescales (15 total) revealed high similarity (Fig. 4A, right), with correlations ranging from r = 0.62–0.67 between infraslow and faster timescales and r = 0.90–0.96 across EEG bands. These empirical correlations were significantly greater than those obtained from shuffled sequences (t₍₂₅₎ range: 175–330, all p < .05/15), demonstrating that temporal organization of blueprint transitions reflects a timescale-overarching principle of connectome dynamics.

Importantly, the cross-matrix correlation reflects the relative structure of transitions—how transition probabilities are distributed across blueprint pairs—independent of their absolute values. While this relative structure was highly similar across timescales, absolute transition probabilities differed, particularly between diagonal and off-diagonal values. These differences reflect timescale-dependent differences in state persistence (diagonal) and switching (off-diagonal) (cf. Fig. 2C–D; see also SI 3.1.3). Thus, although the speed of connectome dynamics differs markedly across timescales, the ordering of state transitions remains remarkably stable, pointing to a shared, timescale-overarching temporal organizational principle.

Notably, a prominent feature of the transition probability matrices was a pointed oval-shaped motif (light yellow in Fig. 4A), reflecting a strong tendency for transitions between spatially similar blueprints. Across timescales, approximately 70% of all transitions occurred between blueprints differencing by only one ICN, whereas transitions between more dissimilar blueprints comprised a smaller (~10%) but non-negligible proportion (see SI 3.1.2). This conserved pattern favors gradual reconfiguration while allowing occasional large shifts Together, these results reveal a core temporal regularity of connectome dynamics: transitions follow structured, non-random rules, with consistent proportions of gradual versus large shifts, regardless of neural timescale. State transitions favor spatial continuity but incorporate occasional jumps to highly dissimilar states, which results in efficient coverage of the full combinatorial state space spanned by ICN combinations.

The organizational principle reflected in transition probability matrices extends beyond transitions themselves to broader statistical properties of connectome sequencing. In a Markov framework, fractional occupancy reflects the system’s long-term equilibrium, determined by the transition matrix, while mean lifetime corresponds to self-transition probabilities (i.e., the diagonal of the matrix). Therefore, the observed regularities in transition structure are expected to propagate to these higher-order temporal properties.

Consistent with this expectation, the distribution of fractional occupancy and mean lifetime across blueprints were preserved across timescales (Fig. 4B–C, left). Fractional occupancy “skylines” showed moderate-to-strong cross-timescale correlations (r = 0.42–0.49 between fMRI and EEG; r = 0.85–0.93 within EEG timescales), all exceeding null expectations (t₍₂₅₎ range: 8–24, all p < .05/15; Fig. 4B, right). Similar results were observed for mean lifetimes: r = 0.17–0.19 between fMRI and EEG, and r = 0.83–0.89 within EEG timescales (t₍₂₅₎ range: 13–37, all p < .05/15; Fig. 4C, right).

Crucially, these correlations reflect the relative patterning of fractional occupancy and lifetime across the 126 blueprint states, rather than their absolute durations, which differ markedly across modalities (Fig. 2D). Yet, the preserved “skylines” structure reinforces the existence of a common organizational scaffold spanning neural activity timescales. Notably, blueprints composed of single ICNs (blueprints 1–7) accounted for a small (~12.2%) and consistent fraction of total fractional occupancy (mean ± std: blueprint (#) 1 = 2.92 ± 0.81%, #2 = 1.65 ± 0.60%, #3 = 2.57 ± 0.60%, #4 = 1.05 ± 0.56%, #5 = 1.40 ± 0.57%, #6 = 2.21 ± 0.87%, #7 = 3.87 ± 1.23%).

These observations were replicated in both the secondary EEG-fMRI data and the large-scale EEG-only data (67–69), with all effects robust before and after source-leakage correction via signal orthogonalization (63) (see SI 2.2). Notably, the EEG-fMRI validation dataset included fMRI recorded at fourfold higher temporal resolution (TR = 0.5 s) than the primary fMRI dataset (TR = 2 s), yet estimated blueprint lifetimes remained unchanged.

Together, these findings demonstrate that connectome state sequencing—indexed by transition probabilities, fractional occupancies, and lifetimes—is governed by a timescale-overarching principle that generalizes across recording modalities and temporal resolutions.

Connectome dynamics unfold in parallel, asynchronous streams across timescales

The discordant speeds of connectome state transitions across neural timescales (Fig. 2C–D) carry important implications for the temporal architecture of brain dynamics. State lifetimes span more than three orders of magnitude, from ~20 ms in fast EEG bands to 3000 ms in fMRI. Despite these differences, blueprint transition sequences exhibit similar structure across timescales, suggesting that common co-activation states recur at temporally scaled rates.

However, such temporal scaling implies asynchrony of state sequences across timescales, countering the prevailing view that often treats the connectome as a single stream of co-activations (cf. introduction). In this conventional view, slower fMRI dynamics is interpreted as a cumulative expression of fast, short-lived electrophysiological states (70), which would predict temporal alignment of identical blueprint states across timescales. The absence of such alignment would speak to an alternative scenario—one in which connectome dynamics unfold as multiple, parallel streams operating at distinct intrinsic speeds.

To adjudicate between the two scenarios, we quantified temporal overlap (i.e., synchrony) of identical blueprint states across all pairs of timescales (15 total), accounting for 6-second hemodynamic delay. As a null, we randomized the blueprint sequence at one of the two timescales, with lifetimes preserved. Across all pairs of infraslow and faster timescales, temporal overlap did not exceed the null distribution (t₍₂₅₎ range: 0.04–1.09, all p > 0.05/15), even after HRF-convolution of EEG-derived spatial correlation timeseries (t₍₂₅₎ range: 0.02–0.25, all p > 0.05/15 (62); see SI 2.2.4). Temporal overlap among EEG timescales was similarly minimal, averaging below 1.5% of time frames (see SI 2.2.4).

Together, these findings show that blueprint sequences at different timescales unfold asynchronously, even when scaled to account for the hemodynamic delay. This observation supports the view that the brain implements multiple, parallel streams of connectome dynamics, each operating at its own intrinsic speed. Rather than forming a single sequence of co-activation states, connectome dynamics appear to reflect a temporally layered architecture, enabling concurrent processing across the full spectrum of neural rhythms.

Discussion

Scale-overarching principles of connectome dynamics in spatial and temporal domains

This study challenges the prevailing view of the connectome as a unitary, temporally homogeneous stream of dynamics. Instead, our findings support a multi-stream model in which distinct sequences of co-activation patterns unfold asynchronously at different transition speeds (cf. Figs. 2C and 4C). Beyond the well-established overlap between slow and fast modalities (i.e., the shared region of the Venn diagram, Fig. 1), our results highlight the non-overlapping regions of the diagram, revealing timescale-specific neural processes that evolve independently across infraslow and electrophysiological timescales. Importantly, despite transition speeds differing by orders of magnitude, these parallel streams are governed by shared spatial and temporal regularities.

In terms of spatial regularities, we found that instantaneous co-activation patterns can be captured by a simple, interpretable set of recurrent states defined by combinations of canonical ICNs. While ICNs have long been viewed as isolated components of large-scale brain architecture, our results show that their full potential emerges primarily through combinatorial expression. The brain exploits nearly the full range of ICN combinations (Fig. 4B): any network can co-activate with any other, and single-ICN states occupy only a minor fraction of time. This pattern underscores that connectome dynamics are fundamentally integrative rather than confined to individual networks.

Notably, the combinatorial state space includes ICN pairings traditionally considered unlikely, such as co-activation between the DMN and “task-positive” control networks, such as DAN and FPN, which are often described as anticorrelated during specific cognitive states (65, 66). Indeed, two of the most frequently expressed blueprints involved simultaneous engagement of the DMN with multiple task-positive networks (FPN, VAN, and DAN; #118 and #126; Fig. 4B). The robust recurrence of these integrative states highlights previously underappreciated co-activation patterns and align with dynamic fMRI evidence for flexible alternation between negative and positive correlations among the DMN and task-positive networks (66), suggesting a flexible, context-dependent reconfiguration of network relationships.

In terms of temporal regularities, we identified that blueprint sequences follow structured, non-random transition rules that preserve the relative balance between gradual and large shifts across neural timescales. Although transitions preferentially maintain spatial continuity, they also incorporate infrequent jumps to highly dissimilar states, enabling efficient coverage of the full combinatorial space of ICN configurations. Together, these findings suggest that some dynamics of the human connectome are founded on spatial and temporal principles that are conserved across timescales.

Cognitive significance of ICN-based spatial principles

Our observation suggests that ICNs are not functionally exclusive entities but act as modular building blocks within a flexible repertoire of co-activation states. Each state may instantiate a distinct “cognitive architecture” in which multiple ICNs are transiently integrated to support ongoing processing (2). When freely recombined, the seven canonical ICNs span a comprehensive space of 126 configurations. The spontaneous expression of this full blueprint repertoire during task-free wakefulness indicates that the brain remains primed for flexible engagement across diverse processing contexts, even in the absence of explicit task demands (53).

These ICN-based co-activation states thus serve as a shared spatial “language”, supporting multiscale dynamics. Their consistent expression from infraslow activity to gamma-band oscillations points to a timescale-overarching spatial principle, whereby the same building blocks recombine to shape dynamic brain states across temporal resolution.

Cognitive significance of the principles’ scale-overarching nature

The emergence of connectome dynamics across multiple timescales likely reflects the brain’s intrinsic tuning to the temporal structure of natural environments, where behaviorally relevant information unfolds concurrently across timescales (41). As environmental inputs impact neural systems differently depending on their temporal scale (12), to process these multi-timescale inputs effectively, the brain must generate neural responses with matching statistical properties. Consistent with this view, both natural behavior and population-level neural activity exhibit scale-free spatio-temporal structure that are correlated (71).

As highlighted in the introduction, a similar multiscale scenario is evident in speech comprehension, where rapid phonemic processing and slower syntactic intergration operate in parallel across distinct timescales (13, 72). Importantly, such parallel processes are characterized by non-overlapping timescales (73–75) and have been proposed as a prerequisite for flexible information routing in complex systems (76).

In line with dynamical models, where the co-existence of multiple timescales is essential for realistic simulations of brain activity and cognition (25, 77–80), our findings reveal parallel streams of brain states— or cognitive architectures—evolving at different speeds. We propose that these asynchronous streams constitue complementary pathways, enabling efficient processing of information distributed across a broad range of temporal scales.

Significance of scale-overarching principles in complex systems

The repetition of a structured pattern across scales is a hallmark of complex systems, widely observed in complex networks (81). A classic example is Roman cauliflower (82), whose fractal-like structure recurs across scales and follows power-law distributions between frequency and size. Unlike objects with a characteristic scale (e.g., apples), scale-free structures lack a dominant average and instead exhibit self-similarity across orders of magnitude.

In neuroscience, converging evidence suggests that the brain is a complex system. Power-law distributions have been reported for individual timeseries from both fMRI and electrophysiology (83), indicating preserved signal structure across timescales. At the network level (6, 42), scale-free properties have been observed in the spatial and topological distribution of anatomical and static functional connections (42, 84, 85), as well as in broad sensor-level EEG topographies (86).

Consistent with a complex system view of the brain, our results show that the same set of ICN-based blueprint states recurs consistently across neural activity timescales. Moreover, these states follow structured transition patterns (i.e., sequencing rules) that are conserved, despite large differences in the speed of underlying neural rhythms. Together, these findings point to a multiscale architecture in which both spatial configurations and temporal sequencing principles are preserved across the hierarchy of brain rhythms. Investigating potential power-law characteristics of state sequences within timescales could further strengthen this complex system view in the future.

Anatomical substrate of multiple connectome streams

The observation of parallel streams of connectome dynamics suggests that fast and slow signals captured by different neuroimaging modalities index partially distinct aspects of neural processing. These differences likely reflect biophysical properties of underlying neurons and circuits. For example, electrophysiological signals may preferentially capture neural activity in thick, myelinated, fast-conducting fibers, whereas BOLD signals may be more sensitive to activity in thinner or unmyelinated, slow-conducting fibers (12). Thus, EEG- and fMRI-derived dynamics may emphasize specialized neural circuits operating at different timescales—circuits that are partly distinct in microscale yet converge at the macroscale of ICNs.

A complementary axis of differentiation may arise from cortical laminar structure. Fast oscillations (e.g., gamma band) are predominantly observed in superficial, feedforward layers, whereas slower rhythms (e.g., beta and below bands) are more prominent in deep, feedback layers (87). This laminar gradient suggests that co-activation patterns at different timescales may emerge from partly distinct cortical layers, operating concurrently to support multi-scale processing. This possibility may become testable with technological extension of layer-resolved whole-brain fMRI (88) to concurrent EEG.

Such partially distinct neural substrates underlying signals across different timescales may help explain why blueprint fits, albeit statistically highly significant, is small in effect size (Fig. 3B). Our study deliberately targets signal variance attributable to canonical ICN co-activations, whereas substantial residual variance likely reflects additional circuit-, layer-, or modality-specific processes. Importantly, the small, explained variance does not imply a lack of functional relevance—a point well-established for evoked neural responses, often dwarfed by the variance of ongoing activity yet remain behaviorally significant (89).

Several methodological considerations warrant note. The modest sample size of our concurrent EEG-fMRI data may limit statistical power; however, all major findings were replicated across independent EEG-fMRI datasets, with and without EEG leakage correction, and further corroborated in a high-powered EEG-only cohort. Another limitation concerns the spatial resolution of source-space EEG, which limits the parcellation schemes that can be examined; however, reproducing the fMRI findings with a higher-resolution atlas (300-region Schaefer) mitigates concerns that the conclusions depend on a specific parcellation choice.

Conclusion

Together, the spatial and temporal regularities identified here point to a timescale-independent architecture of connectome dynamics. Rather than unfolding as a single unified stream, the connectome operates as a multi-timescale process spanning at least three orders of magnitude (~20–3000 ms), in which distinct co-activation states evolve concurrently at their own intrinsic speeds. This temporal multiplexing, analogous to principles used in telecommunication systems (23), supports simultaneous engagement of multiple functional networks and integration of information across a broad behaviorally relevant timescales.

This framework opens novel avenues for both basic and translational research. Multi-modal approaches can directly probe the contributions of distinct connectome streams to cognitive processes operating at different timescales, while inter-individual variations in multi-timescale profiles (e.g., state occupancy or transition probabilities) could yield more sensitive markers of cognitive function and transdiagnostic risk than traditional single-timescale approaches (69, 90–92). Characterizing these multi-timescale dynamic connectome features also holds promise for advancing biofeedback-based interventions in cognitive treatment and augmentation (93). Altogether, the uncovering of a connectome multiplex, governed by timescale-overarching principles, motivates a new phase in connectome dynamics research and its real-world applications.

Materials and Methods

Methods are summarized below; full procedural details are provided in the SI Appendix. We analyzed three datasets: the primary dataset is briefly described here, and the two independent validation datasets are detailed in the SI Appendix.

Subjects

26 healthy subjects (7 females, 5 left-handed, mean age 24.39, age range 18–31) with no history of neurological or psychiatric illness. All procedures were approved by the local Research Ethics Committee (CPP Ile-de-France III), and written informed consent was obtained from all participants (38). Each subject completed three runs of 10-minute resting-state concurrent EEG-fMRI. During acquisition, participants wore earplugs to reduce scanner noise and were instructed to remain still, stay awake, and keep their eyes closed. For three participants, one resting run was excluded due to insufficient EEG quality. The resting-state data were drawn from a larger protocol that also included 10-minute naturalistic film stimuli, which were not analyzed here (94).

Concurrent EEG-fMRI data acquisition

Concurrent EEG-fMRI data were acquired on a 3T Siemens Tim-Trio scanner. Functional MRI was collected using a GRE-EPI sequence (TR = 2000 ms, TE = 50 ms, 40 slices, 3 mm isotropic voxels), yielding 150 volumes per run (450 volumes total). High-resolution T1-weighted anatomical images were acquired for each participant. EEG was recorded simultaneously using two 32-channel MR-compatible BrainAmp amplifiers with 62 scalp electrodes, referenced to FCz, and synchronized with the scanner clock.

MRI processing

Structural MRI was processed using FreeSurfer (v6.0.0) for cortical reconstruction and segmentation. The cortex was parcellated into 68 regions using the Desikan-Killiany atlas, consistent with prior work. The regions were further assigned to seven large-scale networks based on the Yeo et al. framework (48) based on geometric distance. fMRI preprocessing included slice-timing correction, realignment, co-registration to anatomy, nuisance regression (motion, CSF, white matter, and global signal), and band-pass filtering (0.009–0.08 Hz).

EEG processing

EEG data were corrected for gradient and pulse artifacts, resampled to 250 Hz, detrended, and low-pass filtered. Cleaned EEG signals were source-localized using Brainstorm with individual head models and minimum-norm estimation, yielding source activity across 15,000 cortical vertices. Source signals were averaged within atlas regions, band-pass filtered into five canonical frequency bands, and amplitude envelopes were extracted using the Hilbert transform. Envelope timeseries were concatenated across runs for subsequent analyses, without applying leakage correction.

Results with leakage correction applied via signal orthogonalization (63) are provided as a supplemental analysis (see SI 2.1). Leakage correction reduced the overall effect size, as expected, because this conservative procedure removes real (non-spurious) zero-lag long-range functional connectivity (95–103). Empirical and theoretical work further showed that these zero-lag interactions constitute a major component of whole-brain functional connectivity and strongly support the large-scale connectome (104, 105). Thus, removing them inevitably diminishes the magnitude of observed effects. Importantly, however, across all validation datasets, the overall model fit remained well above chance regardless of whether source-leakage correction was applied.

Identifying the spatial principles of connectome dynamics

We defined a comprehensive set of foundational state “blueprints” to capture instantaneous whole-cortex co-activation patterns observed in empirical data. The term “blueprint” (cf. “template” in (53)) denotes idealized states reflecting central tendencies or attractors in dynamic brain activity. Each blueprint was constructed by combining subsets of the seven most commonly observed ICNs. Extending prior fMRI studies demonstrating that certain ICN combinations explain recurrent instantaneous co-activation patterns (53), we (i) applied this framework to EEG source-space data and (ii) systematically enumerated all 126 possible ICN combinations, involving 1 to 6 ICNs (see Fig. 1B). The full seven-network combination was excluded, as it spans the entire cortex, thus lacking spatial specificity

This approach was enabled by source-localizing EEG signals to the same cortical parcellation (Desikan atlas) used for fMRI, providing a shared spatial framework across modalities. Unlike data-driven state definitions—such as those obtained from hidden Markov models (HMM) or independent component analysis (ICA)—blueprints provide a common spatial “language” that facilitates direct cross-modal comparison. Moreover, because blueprints are applied at the native temporal resolution of each modality (seconds for fMRI and milliseconds for EEG), they enable investigation of co-activation dynamics from seconds (fMRI) to milliseconds (EEG), bridging slow and fast brain dynamics.

To track states evolution over time, we computed frame-wise spatial correlation between each of the 126 blueprints and empirical whole-brain co-activation patterns, separately for fMRI BOLD signals and for source-space EEG amplitude envelopes across five frequency bands (δ, θ, α, β, and γ). At each time frame, the blueprint with the highest correlation was selected as the best-fitting state, yielding time-resolved sequences of dominant co-activation patterns six timescales: infraslow fMRI dynamics, and five EEG frequency bands. These sequences provided a direct visualization of large-scale brain dynamics across modalities and temporal resolutions (see Fig. 2C).

Null models for testing spatial principles

To assess the effectiveness of blueprints in explaining variance in whole-brain co-activation patterns, we compared their performance against two null models. A region-label permutation null (Null_LabelPermute) disrupted intrinsic ICN structure by randomly reassigning regional labels while preserving regional counts. A phase-permutation null (Null_PhasePermute) (62, 106, 107) eliminated temporal synchrony between regions while preserving regional spectral properties.

For each null model, the correlations between best-fitting blueprint and empirical data were compared against surrogate data (mean of 100 randomizations per subject), with group-level significance tested using paired t-test, Bonferroni-corrected for six timescales (p < .05/6).

Identifying the temporal principles of connectome dynamics

To characterize temporal principles underlying connectome dynamics, we analyzed transition probabilities between all pairs of blueprint-defined brain states. For each timescale, we computed normalized pairwise transition probabilities between successive best-fitting state blueprints, yielding a 126×126 matrix that captured both state persistence (diagonal elements) and state switching (off-diagonal elements).

In addition, we quantified fractional occupancy (the proportion of time spent in each blueprint state) and mean lifetime (the average duration of continuous visits to each blueprint). Together, these metrics provided a compact characterization of temporal dynamics between blueprints across six neural activity timescales.

Non-random structure of temporal dynamics

To assess whether the observed transition patterns exhibited non-random structure, we compared empirical transition probability matrices (Fig. 4A) against null models that preserved state lifetimes but randomized transition order. Null distributions were generated by randomly shuffling the temporal order of best-fitting blueprint Dissimilarity (1-correlation) between empirical and null matrices was quantified for each timescale and evaluated at the group level using paired t-tests with Bonferroni correction across the six timescales (p < 0.05/6).

We next tested whether non-random transition organization was shared across timescales, indicative of a timescale-independent principle of connectome dynamics. We computed correlations between transition probability matrices for all 15 unique pairs of timescales and compared them against null transition matrices derived from shuffled sequences, using group-level paired t-tests with Bonferroni correction for 15 comparisons (p < 0.05/15). Finally, we assessed whether characteristic distributions (“skylines”; Fig. 4B, C) of fractional occupancy and mean lifetime were conserved across timescales by correlating these metrics across timescales and evaluating using nulls derived by permuting blueprint labels (1 through 126).

Cross-Timescale Synchrony of Blueprint Dynamics

To determine whether connectome dynamics unfold synchronously across timescales, we quantified temporal overlap of identical blueprint states between all pairs of timescales. Overlap was defined as concurrent activation of the same blueprint and evaluated against null distributions that preserved state lifetimes but randomized temporal alignment. Significance was tested using group-level paired t-tests with Bonferroni correction across 15 timescale pairs (p < 0.05/15). To account for the hemodynamic delay inherent to fMRI, all analyses were repeated with a 6-s temporal lag.

In a complementary analysis, we tested whether infraslow fMRI blueprints corresponded to low-pass filtered expressions of faster EEG dynamics. For each EEG timescale, the spatial correlation timeseries of each blueprint (see Fig. 2C) were convolved with the HRF and compared to the corresponding fMRI-derived spatial correlation timeseries. Empirical correlations were compared against phase-randomized null distributions, with group-level inference Bonferroni-corrected across blueprints (p < 0.05/126).

ICN-Level Reorganization During Blueprint State Transitions

To characterize ICN-level reorganization during blueprint transitions, we quantified the number of ICNs that changed activation status between consecutive states (see SI Appendix). Each blueprint was represented as a seven-element binary vector encoding ICN activation (1 = active, 0 = inactive) (48), and transition size was defined as the number of ICNs switching between active and inactive states.

Distributions of transition size were computed for all consecutive timepoints and compared across six timescales to test whether the structure of state transitions was conserved across temporal resolutions.

Significance Statement.

This study challenges prevailing assumptions about how the brain’s large-scale functional connectivity fluctuates from moment to moment. Using concurrent EEG-fMRI (electroencephalography-functional magnetic resonance imaging) data, we examined human brain connectivity across timescales from infraslow to fast. We show that connectivity does not operate at a single speed; rather it unfolds simultaneously across multiple timescales at once. Importantly, we found similar spatial patterns of connectivity and similar sequencing of these patterns regardless of the timescale. This observation demonstrates spatiotemporal regularities governing brain connectivity across different speeds of brain activity. Understanding these multiscale dynamics is essential for advancing research on mental health and how the brain functions as an integrated system that supports cognitive processes of various speeds.

Data, Materials, and Software Availability.

We confirm that the derived data for all three datasets (“Primary,” “Validation I,” and “Validation II”), specifically the signal timeseries across all brain parcels included in the study, will be shared in an open repository and accessible without request. Data have been deposited in all derived data will be shared in the Illinois Data Bank (https://databank.illinois.edu/datasets/IDB-1727544?code=is8JfTZSjwuBugJqyKzVljELCx0-PKIris3upH44DfM) (108). Validation (I) dataset is shared via the National Institute of Mental Health Data Archive (NDA) platform (10.15154/dwbj-8074), in accordance with NIMH data-sharing policies (109). Public deposition of the raw data for the “Primary” dataset, acquired in France, is not permitted due to restrictions under the European Union’s GDPR regulations.

The data were collected prior to the implementation of GDPR, and participants did not provide explicit consent for data sharing (even in anonymized form) with external researchers. Similarly, public deposition of the Validation (II) datasets is not possible because the informed consent obtained at the time of collection did not include provisions for sharing raw neuroimaging data in open repositories. As such, unrestricted data sharing would not be consistent with the ethical and regulatory frameworks governing these datasets. All custom code used for data processing and analysis is publicly accessible on GitHub (https://github.com/connectlab/spatiotemporal_principles.git) (110).