Structurally Guided Neural Networks Reveal Cortical Pathways Behind Proactive Task Switching

A novel neuroimaging framework combining cortical structural connectomes with electroencephalography source signals inside recurrent graph neural networks has identified specific anatomical pathways that coordinate proactive cognitive control during task switching. Researchers at Kyoto University and collaborating institutions developed anatomical-connectivity-guided functional connectivity, termed ACG-FC, to track how the human brain prepares to switch between distinct behavioral rules before an anticipated event occurs. Evaluating twenty healthy young adult participants engaged in an intermittent-instruction task-switching paradigm, the research team analyzed brain activity during the cue-to-target interval, an epoch when the cognitive system must proactively update stimulus-response mappings. The computational model achieved a classification accuracy exceeding eighty percent in discriminating between rule repetition and rule switching, substantially outperforming standard deep neural networks and traditional machine learning pipelines lacking structural priors. This demonstrates that embedding physical white-matter pathways as architectural priors helps computational models isolate the precise functional circuits that govern executive flexibility and temporal expectation.

Proactive cognitive control represents an essential prospective mechanism through which individuals voluntarily allocate attention, suppress irrelevant information, and configure internal representations prior to incoming demands. In traditional cognitive neuroscience investigations, researchers examine rule shifting within the cue-to-target interval of task-switching experiments to understand how the brain transitions from habitual execution inertia to deliberate, goal-directed preparation. While functional magnetic resonance imaging provides exceptional spatial localization of cortical networks, its hemodynamic temporal resolution cannot adequately resolve rapid neural reorganizations that unfold within two seconds after an instructive cue. Conversely, electroencephalography delivers millisecond-level temporal precision to capture oscillatory activity, particularly within the low-frequency delta and theta frequency bands spanning two to thirteen hertz, which underpin executive functions. However, canonical functional connectivity analyses infer interactions purely from signal-dependent statistical metrics, such as phase locking values or coherence across electrode pairs, largely disregarding whether underlying axonal pathways physically support those functional links.

To bridge the divide between anatomical wiring and dynamic electrophysiological signaling, the investigators integrated a group-level deterministic structural connectome directly into the computational graph of recurrent graph neural networks. Rather than treating source-reconstructed electroencephalography as a topology-free tensor where spatial relationships lose permutation invariance, the team mapped cortical signals onto thirty-four regions of interest per hemisphere defined by the Desikan-Killiany atlas. The anatomical template, derived from diffusion tensor imaging across eighty healthy adults, established two hundred and forty-eight undirected structural edges connecting regional centroids with an overall network density of approximately eleven percent. By enforcing message passing functions along these biologically grounded pathways, graph convolutional layers aggregated spatial features across connected cortical neighbors, while recurrent units tracked their temporal evolution across cue-locked epochs. Unlike post-hoc masking approaches that merely filter conventional functional connectivity networks after statistical estimation, this framework treats structural pathways as continuous structural priors during end-to-end supervised representation learning.

The experimental protocol evaluated twenty right-handed native Japanese-speaking university students, comprising ten females and ten males aged twenty to twenty-five years, tested under institutional ethical approvals. Participants viewed centrally presented visual stimuli consisting of geometric shapes differing in color and form, responding via index-finger button presses according to alternating categorization rules prompted by preceding Gabor patch cues. Each cue appeared for one hundred milliseconds, followed by a fixed nineteen-hundred-millisecond stimulus onset asynchrony before the target stimulus was displayed. Behavioral performance demonstrated robust cognitive costs, with participants maintaining an overall mean accuracy of over ninety-five percent across target trials while exhibiting significant switch costs and restart costs in reaction times. Responses to the initial target following a switch cue averaged approximately 638.9 milliseconds, compared to 563.6 milliseconds following a repeat cue, reflecting the heightened processing overhead required to dismantle prior stimulus-response associations and proactively establish newly relevant behavioral contingencies.

Continuous electrophysiological recordings obtained through a thirty-two scalp electrode montage were preprocessed using the MNE-Python toolkit to remove ocular and cardiac artifacts before source reconstruction. The researchers applied weighted minimum norm estimation normalized by dynamic statistical parametric mapping to project sensor-level signals onto the FreeSurfer average cortical surface, mitigating depth bias and volume conduction confounds. Global field power analysis across all electrodes identified two distinct, statistically significant late-latency windows where non-phase-locked electrophysiological power differed substantially between switch and repeat trials after false discovery rate correction. The first window, spanning five hundred to eight hundred milliseconds post-cue, corresponds to the endogenous late positive component reflecting context evaluation, memory retrieval, and conflict resolution during rule updating. The second continuous window, spanning nine hundred to eleven hundred milliseconds, corresponds to the contingent negative variation, an established slow negative potential reflecting prospective anticipation and preparatory motor readiness for the impending stimulus.

Within these two functionally differentiated latency windows, the temporal graph convolutional network architecture demonstrated decisive performance advantages over a comprehensive battery of comparison models across tenfold cross-validation folds. In the 500 to 800 millisecond window, the temporal graph convolutional network achieved a mean classification accuracy of 0.827 and an area under the receiver operating characteristic curve of 0.902, outperforming alternative recurrent graph architectures such as EvolveGCN and dynamic gated graph networks. Furthermore, relative to temporal pattern attention-based long short-term memory networks lacking structural topological guidance, the structurally guided network reduced validation loss by approximately 0.097 while boosting classification accuracy and discrimination metrics by more than six percentage points. Similar performance superiorities emerged in the 900 to 1100 millisecond window, where the structurally constrained graph model improved accuracy by 0.072 over convolution-recurrent baselines, verifying that macroscale structural priors regularize neural representations and suppress non-informative feature aggregation.

To reverse-engineer the biological pathways driving model decisions, the authors deployed the Integrated Gradients algorithm across two distinct attribution pipelines to generate normalized edge attribution values. Attribution Process I pooled all cue trials regardless of instruction type, applying non-parametric permutation tests with ten thousand iterations to isolate fifteen statistically significant common functional connections within the 500 to 800 millisecond window. These connections organized into three discrete, functionally complementary subnetworks centered around the frontoparietal multiple demand system. Subnet 1 occupied frontal and parietal cortices, led by an interhemispheric bridge between the left and right dorsolateral superior frontal gyri and an influential connection between the rostral middle frontal gyrus and caudal anterior cingulate cortex. Subnet 2 linked left temporal and parietal areas via the insula and supramarginal gyrus to encode semantic task context, while Subnet 3 centered on the right insular cortex to coordinate inhibitory control over competing sensory features.

During the subsequent 900 to 1100 millisecond window corresponding to the contingent negative variation, the network architecture underwent substantial topological simplification as cognitive demands transitioned from rule updating to target expectation. Common attribution analysis revealed that the prefrontal executive decision circuit of Subnet 1 completely disappeared, indicating that abstract rule formulation and intentional goal selection had concluded. In contrast, Subnets 2 and 3 remained active in the bilateral frontotemporal regions, albeit with reduced edge counts totaling six significant connections. Subnet 2 preserved four functional links centered on the left inferior temporal gyrus and parahippocampal structures to sustain active perceptual representations in working memory. Concurrently, Subnet 3 maintained two connections linking the right insular cortex to orbital and opercular inferior frontal regions, preserving proactive suppression against impulsive responses and sensory distractions as the scheduled target presentation approached.

Attribution Process II segregated trials by cue instruction, uncovering localized differences in functional pathway relevance between switching and repeating cognitive sets. Stepwise linear regression models incorporating these differential connectivity values accounted for 27.7 percent of reaction time variance in repeat trials and 31.7 percent in switch trials during early preparation. In the late preparatory window, differential connections explained 34.5 percent of reaction time variance in repeat trials and 25.2 percent in switch trials. These predictive relationships significantly surpassed models constructed from unconstrained functional connectivity metrics, establishing that task-optimized structural pathways capture variance directly relevant to motor execution speed.

Quantitative graph comparisons demonstrated notable topological convergence between ACG-FC networks and conventional phase-based functional connectivity metrics, while highlighting critical anatomical refinements. Non-parametric permutation tests against five thousand random surrogate networks confirmed that spectral distances between common ACG-FC networks and conventional functional networks were thirty to forty-four percent lower than null distribution averages. Nonetheless, the authors emphasize several methodological boundaries: attribution values reflect statistical feature contributions within an artificial neural network rather than direct causal physiological transmission along axons. Furthermore, the analysis utilized a group-level binarized connectome and a thirty-two-channel electrode montage in a sample of twenty young adults, limiting individual anatomical specificity and subcortical pathway resolution. Future investigations combining high-density electroencephalography with individual-level probabilistic structural connectivity could expand this framework to examine resting-state or other task-free neurophysiological data.

Subject of Research: Anatomical connectivity-guided recurrent graph neural network modeling of proactive task-switching electrophysiology

Article Title: Anatomical-connectivity-guided functional connectivity reveals task-relevant pathways during proactive task-switching via recurrent graph neural networks

Article References: Wang, S., Miyata, A., Okuya, T., Yanagawa, H., Sakaki, A., Ichinose, N., & Kumada, T. (2026). Anatomical-connectivity-guided functional connectivity reveals task-relevant pathways during proactive task-switching via recurrent graph neural networks. Brain Informatics, 13(1), Article 20. https://doi.org/10.1186/s40708-026-00300-6

Image Credits: AI Generated

DOI: 10.1186/s40708-026-00300-6

Keywords: proactive cognitive control, task-switching paradigm, recurrent graph neural networks, anatomical connectome, electroencephalography source reconstruction, late positive component, contingent negative variation, integrated gradients attribution, frontoparietal multiple demand system, temporal graph convolutional network, cue-to-target interval, diffusion tensor imaging, cortical functional connectivity, stimulus-response mapping, executive function, global field power, Desikan-Killiany atlas