Metastability demystified – the foundational past, the pragmatic present and the promising future

Metastability in neuroscience refers to coexisting tendencies among neuronal populations and brain regions to transiently couple for integrated processing while preserving relative independence for functional specialization. Despite widespread use of the concept, its empirical and theoretical roots are often conflated, and heterogeneous methodologies make it difficult to interpret findings across studies. This Review aims to unify understanding by tracing the origins of metastability, surveying practical signatures used in empirical data, outlining dynamical mechanisms that can generate metastable behaviour, and clarifying misconceptions (e.g., that metastability is simply state switching or that it must involve noise). The purpose is to provide a rigorous conceptual and methodological framework for using metastability to study healthy and disordered brain function, and to guide future theoretical and experimental developments.

The Review distinguishes between the physical neural system, its governing attractor (dynamical) landscape, and observed spatiotemporal activity. Metastability is contrasted with monostability and multistability: while multistable systems require noise to switch among stable attractors, metastable systems intrinsically cycle among saddle-like regions that attract in some directions and repel in others. Historically, metastable phenomena were noted in thermodynamics/chemistry (e.g., supersaturated solutions, supercooled water) but modern neuroscientific usage stems from coordination dynamics (e.g., fish fin coordination, human finger–metronome tasks) where ‘almost-in-sync’ states alternate with desynchronization. Parallel computational lines introduced itinerant dynamics and chaotic itinerancy to account for transient neural patterns, and percolation-based approaches offered additional perspectives. The Review consolidates these literatures to frame today’s neuroscientific investigations.

This is a comprehensive narrative Review synthesizing theoretical foundations, computational models, and empirical applications of metastability. The authors (1) formally define metastability within dynamical systems, (2) review and operationalize practical, heuristic signatures estimable from neuroimaging/electrophysiology (e.g., std-KOP, std-SPECT, std-IGNITE, mean-VAR), with a publicly available MATLAB code library, (3) appraise multiple dynamical routes to metastability (fixed-point memory; cycles of saddles via delays or asymmetric stimuli; chimera states; chaotic itinerancy) and contrast them with noise-driven multistability, (4) survey empirical studies linking signatures to ageing, neurodegeneration, psychiatry, anaesthesia, development, and injury, (5) identify neurobiological drivers at micro- and macroscales (e.g., clustering, inhibitory balance, structural topology, heterogeneity, noise), and (6) propose inference frameworks including generative models, stringent null models, and analysis of dwell-time distributions to help disambiguate metastability from multistability.

• Conceptual clarifications: Metastability is a property of a system’s dynamical landscape (involving saddles and unstable attraction), not of data; commonly used ‘signatures’ are heuristic markers of transient dynamics and are necessary but not sufficient to prove metastability; noise is not required for metastable cycling (unlike multistable switching).
• Practical signatures: (1) std-KOP (temporal variance of global synchrony), (2) std-SPECT (variance of the spectral radius of time-varying functional connectivity), (3) std-IGNITE (variance of intrinsic ignition breadth), and (4) mean-VAR (mean temporal variance of the leading eigenvector’s entries from phase-aligned connectivity). These capture integration–segregation variability but do not, on their own, disambiguate metastability from multistability.
• Empirical patterns: Reduced signatures of metastability in ageing (humans and rats), Alzheimer’s disease and mild cognitive impairment; reductions under anaesthesia and after traumatic brain injury; reduced metastability in preterm neonates compared to full-term; increased metastability-related measures in schizophrenia/early psychosis (e.g., elevated mean-VAR; modular increases in std-KOP). Together, results suggest both increases and decreases are associated with pathology, implying a putative ‘optimal range’ or sweet spot for healthy function.
• Dynamical routes to metastability: Identified mechanisms include fixed-point memory (ghost attractors post-bifurcation), cycles of saddles enabled by delays or asymmetric stimuli (winnerless competition), chimera states in phase-lagged oscillator networks, and chaotic itinerancy via Milnor attractors. These differ from noise-driven multistability in which shallow attractors and fluctuations drive switching.
• Neurobiological drivers: Microscale factors include clustered excitatory architecture, synaptic weights, recurrent inhibition, and intrinsic noise; macroscale drivers include structural connectome topology (clustering coefficient, small-worldness, eigenvector centrality), regional heterogeneity (e.g., intrinsic frequencies), and noise level, which can modulate state durations and clustering.
• Inferential advances: Generative model inversion tuned to metastable regimes, stringent null models beyond phase randomization, and state dwell-time distributions offer avenues to discriminate metastability (often yielding gamma-like dwell-time distributions in models) from multistability (often exponential or stretched-exponential dwell times).

By rigorously defining metastability and separating it from multistability and criticality, the Review addresses confusion in the literature and clarifies how transient brain dynamics may arise from distinct mechanisms. It shows that widely used signatures quantify integration–segregation variability but cannot alone prove metastability, guiding researchers to combine signatures with generative modelling and null models. The synthesis of micro- and macroscopic drivers links neurobiology and network topology to dynamic regimes, helping interpret empirical associations with ageing, disease, and cognition. Highlighting dwell-time statistics as a potential discriminator offers a concrete analytical path forward. Collectively, the framework supports the view that healthy cognition may require a balance (an ‘optimal range’) of metastable dynamics enabling flexible coordination, while emphasizing the need to demonstrate advantages of metastability over multistability via model-based comparisons and information-theoretic metrics.

The Review unifies the conceptual, methodological, and empirical landscape of metastability in neuroscience. It establishes metastability as a coherent dynamical construct with multiple mechanistic routes and practical heuristic signatures, cautions against conflating signatures with definitions, and differentiates metastability from criticality and multistability. Future priorities include: (1) model-based inference demonstrating when metastability better explains data than multistability, (2) stringent null models and dwell-time analyses to discriminate mechanisms, (3) mapping neurobiological determinants (cellular, molecular, receptor/transcriptomic architecture) of metastable regimes, and (4) translating insights into biomarkers and targeted interventions (pharmacological, psychotherapeutic, and neuromodulatory) to restore healthy dynamic regimes.

• Heuristic signatures (std-KOP, std-SPECT, std-IGNITE, mean-VAR) are necessary but not sufficient to confirm metastability and can conflate metastability with multistability or randomness.
• Reconstruction of attractor landscapes in large neural systems is often infeasible given current data volume/granularity.
• Normative values for signatures are lacking; interpretations are typically relative across groups/conditions.
• Disambiguating mechanisms (metastability vs multistability vs criticality) from data alone is non-trivial and sensitive to modelling assumptions, control parameters, and state variable choices.
• Dwell-time-based discrimination, while promising, remains partially explored and requires further principled validation.
• Scale dependencies (micro to macro) and model simplifications (e.g., mean-field) can obscure the underlying dynamics, complicating inference and generalizability.