The core methodological innovation is called integrated effective connectivity, or iEC. Effective connectivity refers to the directed influence one brain region exerts over another, a quantity that ordinary functional connectivity, which merely measures statistical co-fluctuation, cannot provide. Estimating directed connections from fMRI has long been a contentious enterprise, with dozens of competing algorithms, from Granger causality and vector autoregressive models to LiNGAM-style structural equation methods and dynamic causal modeling, each resting on different assumptions and each prone to distinctive failure modes. Rather than betting on a single technique, the team integrated nine such algorithms into a single consensus estimate. Each algorithm’s output was weighted by optimized coefficients determined through Bayesian optimization, with the weights tuned so that the combined directed network best reproduced the statistical and dynamic fingerprints of the empirical fMRI data, including static functional connectivity and the temporal variability of connectivity over time.

Crucially, the researchers did not rely on internal consistency alone. They validated the framework against ground truth in two ways. First, using whole-brain simulations built on a Hopf oscillator model, they generated synthetic directed networks, ran the iEC pipeline on the simulated activity, and tested whether the method could recover the known directed structure. The full integrated framework outperformed the best individual algorithm, achieving median correlations of roughly 0.75 with ground truth at finer parcellations and about 0.65 at coarser ones. Second, and more compellingly biologically, they turned to the macaque monkey, where decades of invasive tract-tracing experiments have quantified the laminar origins of interareal projections. The fraction of labeled neurons, or FLN, and the supragranular labeled neurons ratio, or SLN, provide a histological signature of whether a connection is feedforward or feedback. The team applied iEC to resting-state fMRI from macaques in the PRIME-DE repository and found that the directed connectivity estimates aligned with these anatomical markers of directionality, suggesting that the method recovers something real about how signals travel between cortical areas.

With validation in hand, the group applied the framework to resting-state fMRI data from hundreds of participants in the Human Connectome Project S1200 release, using both the Schaefer-100 parcellation and the finer multi-modal MMP-360 atlas. From the consensus directed network, they computed a hierarchy value for each cortical parcel, essentially a measure of how far upstream or downstream a region sits in the global flow of signals. The result was strikingly orderly. Hierarchy levels increased monotonically along the principal axis of cortical organization that runs from sensorimotor cortex through association areas to paralimbic regions, the same trajectory captured by the well-known sensorimotor-to-transmodal gradient. Primary sensory and motor areas, which are densely myelinated and dominated by the granular koniocortex cytoarchitectural type, sat at the bottom of the flow hierarchy, while agranular and dysgranular limbic and paralimbic cortices sat at the top.

This ordering matters because it matches an independent line of biological theory. The structural model of laminar connectivity, developed through the work of Henry Kennedy, Nikolaus Krich and others, predicts the direction of cortical signaling from the laminar distribution of neurons that send and receive interareal projections. Areas whose projections originate predominantly in deep layers and terminate in granular layers are feedforward senders; areas with the opposite pattern are feedback receivers. The iEC-derived hierarchy recovered exactly this predicted organization in the human brain, without any laminar information being supplied, and it also tracked inverted myelination and cortical type annotations derived from histological atlases, including the Campbell cytoarchitectonic map and modern myeloarchitectonic reconstructions. In other words, a purely functional, data-driven measurement of directed signal flow in humans recapitulates a century of microanatomical knowledge about cortical architecture.

The most surprising finding, however, is that this hierarchy is not static. When the team applied the framework to data collected under different conditions, including movie-watching from the StudyForrest dataset and a tonic-pain paradigm, the shape of the cortical hierarchy changed systematically with the participant’s attentional and processing state. During externally oriented processing, such as watching a film, the hierarchy flattened: the distinction between lower and higher levels of the flow hierarchy became compressed, and signal flow became more distributed and less strictly ranked. During internally focused conditions, including rest and the sustained pain state, the hierarchy steepened, with greater engagement of interoceptive regions that monitor the body’s internal milieu. The investigators interpret this steepening as reflecting the increased weight of bodily and self-referential signals in the cortical conversation when attention turns inward.

These state-dependent shifts resonate with a growing body of literature on hierarchical reorganization. Previous work by Morten Kringelbach, Gustavo Deco and colleagues reported a flattening of brain hierarchy during naturalistic movie-watching compared with rest, and studies of cortical gradients have shown that the principal sensorimotor-to-association axis compresses or expands depending on task demands, development and clinical condition. What the new study adds is directional information: it is not merely that the amplitude of regional activity or the strength of correlations changes, but that the very direction in which signals propagate across the cortex reorganizes. This distinction between correlation and directed influence has been a persistent bottleneck for human connectomics, and the iEC framework offers one route past it, at least at the macroscale resolution that fMRI affords.

The technical safeguards the team employed deserve emphasis, because the credibility of any effective connectivity result hinges on them. The researchers showed that most constituent algorithms were robust to hemodynamic response function variability, and that the integrated estimate remained highly stable before and after blind deconvolution of region-specific hemodynamic lag, with correlations of 0.91 in macaque and 0.87 in human data. They demonstrated that including negative, inhibitory-flavored connections improved hierarchy estimation in higher-order cortical areas, where excitation-inhibition balance is thought to shift systematically along the cortical gradient. They also verified that the Hopf whole-brain model used for simulation adequately reproduced both static and dynamic features of empirical brain activity in both species, lending legitimacy to the synthetic ground truths used for training. All code and processed derivatives are publicly available through the authors’ GitHub repository, allowing independent replication of the entire pipeline.

The implications reach well beyond basic systems neuroscience. Because the method yields a directed, individually estimable network, it opens the door to quantifying atypical signal flow in clinical populations. Prior work has already documented compressed cortical hierarchy in autism and in schizophrenia using undirected gradient measures; a directional framework could clarify whether such compression reflects altered feedforward propagation, exaggerated feedback, or both. Similarly, conditions characterized by aberrant interoception and bodily awareness, from chronic pain to anxiety disorders, might be reframed in terms of state-dependent steepening of the cortical hierarchy, providing a mechanistically grounded target for interventions. The methodological benchmarking literature, including recent large-scale comparisons of connectivity mapping approaches, has highlighted how fragile many directed-connectivity estimates are; a validated, integrated estimator that survives adversarial testing represents a meaningful step toward reliability.

Cautions remain. Effective connectivity inferred from fMRI is still an indirect inference, filtered through the vascular hemodynamic response, and the authors themselves note that one macaque-fMRI-tested algorithm, multivariate Granger causality, had to be excluded from human analyses owing to its sensitivity to hemodynamic variability. The tonic-pain dataset is restricted and not publicly available, and the framework’s assumptions, though tested in simulation, ultimately rest on the fidelity of macroscale models. Yet the convergence achieved here, across species, across algorithms, across parcellations and against histological ground truth, is difficult to dismiss. If the findings hold up under independent replication, the picture that emerges is of a human cortex whose direction of information flow obeys the laminar logic discovered in animal anatomy, but whose hierarchical geometry bends to the demands of the moment, flatter when the world commands attention and steeper when the mind turns toward itself.

Subject of Research: Directed signal flow hierarchy and its state-dependent reorganization in the human cerebral cortex

Article Title: State-dependent signal flow hierarchy in the human cerebral cortex

Article References: Oh, Y., Ann, Y., Lee, J.-J., Ito, T., Froudist-Walsh, S., Paquola, C., Milham, M., Spreng, R. N., Margulies, D., Bernhardt, B., Woo, C.-W., & Hong, S.-J. (2026). State-dependent signal flow hierarchy in the human cerebral cortex. Nature Neuroscience. https://doi.org/10.1038/s41593-026-02389-8

Image Credits: AI Generated

DOI: 10.1038/s41593-026-02389-8

Keywords: effective connectivity, cerebral cortex, cortical hierarchy, fMRI, feedforward feedback, signal flow, interoception, connectomics, laminar connectivity, brain states, neuroimaging, macaque validation