Functional connectivity MRI, the technique at the heart of the study, exploits the fact that spatially distant brain regions fluctuate together in their activity even when a person is simply resting with eyes open. These temporally coherent fluctuations reveal the brain’s intrinsic architecture: distinct functional systems that support sensorimotor, cognitive, and affective processes. In adults, this organization has been characterized extensively, both in group averages and in densely sampled individuals. But questions about how these networks mature through childhood and adolescence have persisted, partly because pediatric neuroimaging is notoriously difficult.

Children move in the scanner. Head motion induces systematic artifacts in functional connectivity estimates, and variable data quality has long complicated efforts to establish what is real signal and what is noise. Subcortical structures present their own challenges, including reduced temporal signal-to-noise ratio, small anatomical size, partial volume effects, and sensitivity to preprocessing choices. As a result, many developmental studies have focused on the cortex alone, leaving the organization of cortico-subcortical loops—which connect the cortex with the basal ganglia, thalamus, and cerebellum through recurrent circuits supporting integrated processing—less systematically characterized in children.

The new study tackled these obstacles with scale and statistical discipline. The team divided the ABCD baseline sample into two demographically matched halves: a discovery cohort of 3,624 children and a replication cohort of 3,692 children, all aged 9 to 10 years, matched across ten demographic variables including site location, age, sex, ethnicity, grade, parental education, handedness, family income, and exposure to anesthesia. Family members were kept together in the same set to avoid confounding. Strict motion exclusion criteria were applied: only children with at least 600 frames—eight minutes—of low-motion data (framewise displacement below 0.2 millimeters) were included, with a minimum of five contiguous retained frames. The imaging data themselves were collected across 21 U.S. sites on Siemens Prisma, Philips, and GE 3 Tesla scanners, with 20 minutes of eyes-open resting-state fMRI per child.

After preprocessing with the ABCD-HCP pipeline—which included anatomical normalization, cortical surface parcellation, distortion correction, nuisance regression, band-pass filtering, and respiratory motion correction—the researchers computed pairwise correlations among all 59,412 cortical vertices for each participant and averaged them within each cohort. They then applied the Infomap community-detection algorithm, a graph-theoretical method originally developed for mapping communities in social and biological networks, across a range of edge densities from 0.1 to 5 percent. Communities were assigned functional identities by template matching against canonical network templates using the Jaccard index, with a consensus identity selected at the sparsest threshold where a reliable match was found. For subcortical voxels, each voxel in the basal ganglia, thalamus, and cerebellum was assigned to the cortical network with which it showed the strongest correlation, in a winner-take-all fashion.

The reproducibility results were striking. Vertex-wise functional connectivity between the discovery and replication cohorts correlated at r = 0.99, and parcel-level matrices showed the same near-perfect agreement. Network topography was similarly consistent: normalized mutual information (NMI), a metric ranging from 0 for no agreement to 1 for perfect agreement, reached 0.90 in the cortex, 0.87 in the basal ganglia and thalamus, and 0.89 in the cerebellum, all statistically significant under spatial permutation testing that rotated network labels on a spherical surface 1,000 times. The maps recovered all the canonical systems previously described in adults, including the default-mode, frontoparietal, cingulo-opercular, salience, dorsal attention, ventral attention, somatomotor, visual, and auditory networks.

The large sample also confirmed classic anatomical motifs with unprecedented statistical power. On the medial wall of the cortex, the team observed a foot-hand-foot somatomotor arrangement consistent with the known somatotopy of the supplementary motor area, a pattern first described through stimulation studies in the 1950s. The parietal operculum showed a hand-face topography matching the established organization of secondary somatosensory cortex. In the subcortex, the posterior putamen and central thalamus displayed topographically organized somatomotor assignments, and the lateral geniculate nucleus of the thalamus appeared exactly where visual relay anatomy predicts it. In the cerebellum, the characteristic dorsal and ventral representations of motor cortex were evident, reflecting the well-documented doubled somatomotor topography of the motor cerebellum. Every one of these features replicated independently across the two halves of the sample.

The study also extended its reach to a recently discovered system: the Somato-Cognitive Action Network, or SCAN, first described in 2023 as a network interdigitated with effector-specific motor regions and implicated in integrating motor control with cognitive and goal-directed processes. Using a manual seed-mapping approach with two independent raters and conservative connectivity thresholds around the 96th to 98th percentile, the team identified SCAN’s characteristic distributed topography in the child data. A split-half analysis yielded an NMI of 0.88, providing novel evidence that SCAN is a reproducible feature of large-scale functional organization in children, though the authors caution that automated approaches remain variable in their sensitivity to the network’s ventral and integrative components.

When the child maps were compared with adult data, the picture became more nuanced. To ensure comparable sample sizes, the researchers subsampled 1,000 children from each ABCD cohort, sex-matched to the 1,000 HCP adults aged 22 to 37. Mantel tests confirmed robust adult-child correspondence in both the rank ordering and the absolute strength of functional connections: Spearman correlations of roughly 0.80 and Pearson correlations of roughly 0.85 between HCP and each ABCD subsample. But these cross-age values were consistently lower than the near-identical agreement between the two child cohorts (Spearman r = 0.998). Community structure showed moderate adult-child similarity (NMI around 0.62 to 0.63) compared with 0.88 between child partitions, and Dice coefficients quantifying spatial overlap of individual networks exceeded spatially constrained null expectations for every network, with sensorimotor networks showing the highest overlap and association networks the lowest.

A graph-theoretical measure called the participation coefficient, which indexes how broadly a region’s connections span multiple networks, revealed further developmental nuance. Adults showed higher participation in dorsal attention, somatomotor-hand, somatomotor-foot, and visual networks, suggesting that in maturity these systems engage in more distributed cross-network communication, potentially reflecting greater integration with higher-order association and control systems. Children, by contrast, showed relatively higher participation in the auditory network. The authors interpret the lower participation of sensory and motor systems in children as possibly reflecting a more modular organization, with these systems remaining relatively segregated from higher-order networks during late childhood. The researchers emphasize that these differences should be read as descriptive features of network architecture rather than definitive markers of developmental progression, since the metric is sensitive to parcellation boundaries and cross-dataset acquisition differences.

The benchmark carries practical weight beyond basic neuroscience. By establishing a stable, reproducible reference for functional network organization at ages 9 to 10, the framework gives researchers a foundation for longitudinal analyses as additional waves of ABCD imaging data become available, helping distinguish true within-subject developmental change from measurement variability. It may also aid the detection of atypical network organization underlying emerging cognitive, emotional, or behavioral difficulties in conditions such as attention-deficit/hyperactivity disorder, autism spectrum disorder, mood disorders, and psychotic-spectrum disorders. The authors note that group averages necessarily smooth over individual variability, which is substantial in higher-order association regions, and that population-level and individualized mapping approaches should be treated as complementary. As the ABCD cohort matures, this benchmark will serve as the yardstick against which the adolescent remodeling of the brain’s functional architecture can finally be measured.

Subject of Research: Population-scale mapping of functional brain network organization in childhood and its similarity to adult network architecture

Article Title: Benchmarking functional brain network organization in childhood and its similarity to adults

Article References: Ali, S. A., Demeter, D. V., Baim, A. R., Koithan, E. M., Feigelis, M., Zreik, S., Ahern, J., Chang, S. E., Park, S., Gordon, E. M., Marek, S., & Greene, D. J. (2026). Benchmarking functional brain network organization in childhood and its similarity to adults. Frontiers in Neuroimaging, 5, Article 1818662. https://doi.org/10.3389/fnimg.2026.1818662

Image Credits: AI Generated

DOI: 10.3389/fnimg.2026.1818662

Keywords: brain networks, functional connectivity, childhood development, resting-state fMRI, ABCD Study, Human Connectome Project, neuroimaging, somatomotor networks, association networks, Somato-Cognitive Action Network, graph theory, brain maturation

The research team, led by investigators affiliated with hospitals and medical colleges in Yancheng and Nantong, China, took a fundamentally different approach to the puzzle. Rather than asking where in the brain gray matter changes occur, they asked what functional network those changed regions belong to. The underlying idea, drawn from a rapidly growing movement in neuroscience, is that brain disorders are best understood not as problems of isolated spots but as disturbances of interconnected networks. When different anatomical locations produce similar clinical pictures, it is often because they are wired into the same functional circuit. The team used a technique called functional connectivity network mapping, or FCNM, to test whether the spatially scattered gray matter findings in IGE might converge on a shared network architecture.

To build their dataset, the researchers systematically searched PubMed, Embase, and Web of Science following PRISMA guidelines, ultimately identifying 18 whole-brain VBM studies of IGE published up to May 2026. These studies collectively reported 111 peak coordinates of gray matter alteration from 1,492 participants, including 694 patients and 798 healthy controls. The epilepsy subtypes represented ranged from juvenile myoclonic epilepsy—the most frequently studied—to childhood absence epilepsy, juvenile absence epilepsy, and epilepsy with generalized tonic-clonic seizures alone. Patient ages spanned from early childhood to middle adulthood, and disease durations ranged from a few months to two decades. Crucially, the team pooled coordinates reflecting both increased and decreased gray matter in their primary analysis, treating the direction of change as secondary to the question of network membership.

The methodological heart of the study lies in how those coordinates were transformed into a network map. Each of the 18 patient-control contrasts yielded its own set of gray matter coordinates, which were converted into 4-millimeter-radius spherical seeds. Using resting-state functional MRI data from 1,093 healthy adults in the Human Connectome Project S1200 release, the researchers calculated whole-brain functional connectivity maps for each seed by correlating the seed’s mean time series with that of every other brain voxel. Pearson correlations were converted to Fisher’s z values, individual maps were combined through one-sample t-tests, and the resulting group-level maps were thresholded using voxel-wise false discovery rate correction at p < 0.05. Only positive connectivity was considered, reflecting ongoing debate about the biological interpretation of negative connectivity measures. The binarized maps from all 18 contrasts were then overlaid, and only voxels showing positive connectivity with at least 60 percent of the contrast-specific seeds were retained in the final network—a criterion previously validated in FCNM studies.

The result was striking. Despite the anatomical scatter of the input coordinates, the analysis revealed a coherent, distributed normative connectivity pattern anchored in four principal regions: the bilateral middle temporal gyri, the bilateral thalami, the left superior frontal gyrus, and the right putamen. When the team measured how much of this network overlapped with eight canonical brain networks—seven cortical networks defined by the Yeo parcellation plus a subcortical network delineated from the Human Brainnetome Atlas—the subcortical network dominated, accounting for 42.77 percent of the overlap. The default mode network, a cortical system involved in self-referential thinking and internally directed cognition, came in a distant second at 11.06 percent. Overlap with the visual, somatomotor, dorsal attention, ventral attention, limbic, and frontoparietal networks each fell below 10 percent.

Replication is where many neuroimaging findings falter, so the team put their network to the test. They repeated the entire FCNM procedure using an independent, cross-scanner validation cohort: the Southwest University Adult Lifespan Dataset, comprising 329 healthy individuals scanned on a different 3 Tesla Siemens system. The validation analysis produced a broadly similar spatial pattern, with the subcortical network again showing the greatest overlap at 43.98 percent. The default mode network correspondence, however, was noticeably weaker in the validation dataset than in the discovery analysis—a discrepancy the authors attribute partly to the wider age range of the SALD sample and to differences in preprocessing pipelines, including the use of ICA-FIX denoising in the Human Connectome Project data but not in SALD.

Sensitivity analyses reinforced the robustness of the subcortical signal while exposing the fragility of the cortical component. When the researchers varied the seed radius from 4 millimeters to 1 and 7 millimeters, the resulting maps preserved their dominant subcortical and default mode overlap. When they separated gray matter increases from decreases, a potentially intriguing dissociation emerged: decreased gray matter foci showed their greatest overlap with the subcortical network at 65.84 percent, while increased gray matter foci aligned more with the default mode network at 37.23 percent. The authors caution that these direction-specific analyses rest on fewer contrasts and should be treated as hypothesis-generating rather than definitive. A leave-one-contrast-out analysis, in which each of the 18 contrasts was excluded in turn and the FCNM procedure repeated, found that 14 of the 18 resulting maps—77.7 percent—reproduced the major network pattern.

The convergence on the thalamus and putamen carries substantial biological weight. The thalamus has long been implicated in generalized epilepsies: structural MRI studies have documented thalamic volume and morphology abnormalities in childhood absence epilepsy, juvenile myoclonic epilepsy, and generalized tonic-clonic seizures, and functional imaging has captured thalamic activation during typical absence seizures coinciding with generalized spike-and-wave discharges. Neurochemical work has revealed altered GABA concentrations in the frontal lobe and thalamus of patients with juvenile myoclonic epilepsy. The putamen, meanwhile, has emerged as a subcortical hub in generalized epilepsy pathogenesis, with multimodal imaging revealing microstructural integrity deficits and diminished dopamine D2/D3 receptor binding that may weaken the basal ganglia’s inhibitory grip on pathological cortical discharges. Prior intracranial EEG research has further shown that interictal thalamocortical network configuration tracks susceptibility to consciousness-impairing seizures, and thalamic deep brain stimulation and vagus nerve stimulation are thought to act partly by restoring normative thalamocortical rhythms.

The default mode network component of the findings is more tentative. The middle temporal gyrus and superior frontal gyrus nodes identified in the study are known to participate in cognitive and emotional processing, and previous work has documented both structural and functional abnormalities in these regions in juvenile myoclonic epilepsy. Broader epilepsy literature has consistently reported altered functional connectivity within the default mode network across multiple syndromes, particularly at the posterior cingulate cortex and precuneus, and dynamic analyses suggest these subtle network anomalies may be more sensitively captured by temporal variability measures than by static connectivity. Yet because the DMN overlap was not comparably reproduced in the validation dataset, the authors interpret it as a secondary, exploratory observation rather than a definitive network-level finding.

The researchers are careful to frame their conclusions within important limits. Because the functional connectivity was derived entirely from healthy adults, the identified network represents the normative connectivity profile of reported gray matter alteration sites, not direct evidence of abnormal connectivity in patients. No coordinate-based permutation or spatially constrained null-model analysis was performed, so the study cannot establish that the observed overlap exceeds chance expectation or that it is disease-specific—particularly given the thalamus’s inherently broad connectivity. The developmental mismatch between the adult normative connectome and pediatric patient samples, the heterogeneity of scanners and analytic methods across the included studies, and the absence of negative findings from the original literature all add uncertainty. Still, the authors argue that their findings provide preliminary, descriptive evidence that the scattered gray matter alterations of IGE are embedded within a reproducible subcortical normative connectivity pattern, reframing IGE as a distributed cortico-subcortical network disorder and offering a framework that may guide future network-based biomarkers and candidate targets for neuromodulation.

Subject of Research: Normative functional network localization of gray matter alterations in idiopathic generalized epilepsy

Article Title: Normative functional network localization of gray matter alterations in idiopathic generalized epilepsy

Article References: Wang, X., Zhang, Z.-P., Zheng, B.-C., Liu, L., Wang, G.-D., Yang, H.-C., & Ma, L.-L. (2026). Normative functional network localization of gray matter alterations in idiopathic generalized epilepsy. Frontiers in Neurology, 17, Article 1957963. https://doi.org/10.3389/fneur.2026.1957963

Image Credits: AI Generated

DOI: 10.3389/fneur.2026.1957963

Keywords: idiopathic generalized epilepsy, gray matter, voxel-based morphometry, functional connectivity network mapping, thalamus, putamen, default mode network, subcortical network, resting-state fMRI, Human Connectome Project, juvenile myoclonic epilepsy, brain networks

The research, led by Johanna L. Popp and Kirsten Hilger of the University of Würzburg together with Joshua Faskowitz of Indiana University, took advantage of the Human Connectome Project, one of the largest open-access neuroimaging datasets in existence. The team analyzed ultra-high-field 7-Tesla fMRI data from 174 young adults who had watched four short films—roughly fifteen minutes each—while lying in the scanner. Crucially, the participants had also completed the NEO-Five-Factor Inventory, a standard questionnaire measuring the Big Five personality traits, allowing the researchers to link each person’s neuroticism score to the way their brain responded to the films.

Rather than asking whether particular brain regions light up more strongly in neurotic individuals, the team employed a technique called inter-subject representational similarity analysis, or IS-RSA. The logic is elegant. First, the researchers measured how similarly each pair of participants’ brains responded to the movies, region by region, using inter-subject correlations of activity time courses across 200 functionally defined brain areas. They then compared this matrix of neural similarity to a matrix of similarity in neuroticism scores. If people with similar personalities show similar brain responses, the two matrices should align—and the precise form of that alignment can reveal something deeper about how the trait is organized in the brain.

To test competing hypotheses about that form, the researchers constructed three mathematical models of trait similarity. The Nearest Neighbors model assumes that people with similar neuroticism scores, whether high or low, show similar brain activity. The Anna Karenina model—named after Tolstoy’s famous opening line about happy and unhappy families—assumes that high scorers resemble one another while low scorers diverge. The Gaussian model predicts that people in the middle of the distribution are the most alike. When the team compared these models against the actual neural data, a clear winner emerged, but with an unexpected twist: it was the inverse of the Anna Karenina model that best described reality. In other words, people with low neuroticism responded to the films in highly similar ways, while people with high neuroticism diverged from one another—and from everyone else—in their patterns of brain activity.

This finding of increasing neural heterogeneity with increasing neuroticism is more than a statistical curiosity. The researchers speculate that it may reflect a well-documented feature of the trait: a propensity for worry and rumination. According to Attentional Control Theory and the Dual Mechanisms of Control framework, highly neurotic individuals rely on reactive, bottom-up attentional systems that are vulnerable to disruption by internally generated thoughts. While watching a film, a person high in neuroticism may drift inward, activating autobiographical memories and personal concerns that are unique to their life history. Such idiosyncratic internal processing would naturally produce brain activity that differs from person to person, whereas low scorers, attending more uniformly to the screen, would generate closely matched responses. The authors also note that distinct subtypes of neuroticism, such as the so-called healthy neuroticism profile combining high neuroticism with high conscientiousness, could contribute to this diversity of neural responses.

The study’s second major contribution came from a clever two-stage design. Before analyzing any brain data, the researchers ran an online pilot study in which 86 independent participants rated 88 thirty-second scenes from the same four movies on how strongly each scene evoked emotions related to the six facets of neuroticism: anxiety, angry hostility, self-consciousness, depression, impulsiveness, and vulnerability. Scenes falling in the top quartile of these ratings were classified as trait-relevant, while those in the bottom quartile were classified as trait-irrelevant. This allowed the team to ask whether the link between neuroticism and brain activity depends on what is happening on screen—a direct test of interactionist personality theories, such as Cybernetic Big Five Theory and Trait Activation Theory, which hold that traits reveal themselves most clearly in situations that engage them.

The answer was emphatic. When the researchers restricted their analysis to trait-relevant scenes, the association between neuroticism and neural response heterogeneity was roughly twice as strong as during trait-irrelevant scenes. At the whole-brain level, the effect size grew from a Cohen’s d of about 0.64 to about 1.43, and the difference between the two scene types was itself highly significant. The amplification was visible across multiple large-scale functional networks, including the default mode network, the ventral attention and salience network, the dorsal attention network, and the somatomotor network, with the default mode network—a system closely tied to self-referential thought and autobiographical memory—showing one of the most robust enhancements. The pattern replicated almost entirely in an independent sample of 85 participants drawn from the same dataset but analyzed separately.

The authors interpret this amplification as a kind of neural stress test. Just as a cardiac stress test exposes hidden weaknesses in the heart that are invisible at rest, emotionally demanding scenes appear to engage the neural machinery of neuroticism more fully, making each individual’s trait-related response patterns more pronounced and therefore more detectable. This idea converges with earlier work showing that brain-trait associations in general become stronger when neural activity is constrained by trait-relevant tasks, and it carries a practical message for the field: stimulus selection is not a trivial detail but a powerful lever for improving the sensitivity of neuroimaging studies of individual differences, an area of research that has struggled with small effect sizes and replication failures.

The findings also resonate with a growing literature on clinical populations. Reduced similarity of brain responses during naturalistic viewing has been reported in patients with melancholic depression, first-episode psychosis, and in children and adolescents with elevated social anxiety or depressive symptoms. The new results extend this picture to neuroticism in healthy adults, lending support to dimensional models of psychopathology, such as the Hierarchical Taxonomy of Psychopathology, which view mental disorders as extreme ends of continuous trait distributions rather than categorically distinct illnesses. If heightened neural heterogeneity during emotional stimulation marks a vulnerability that precedes clinical onset, naturalistic neuroimaging could eventually help identify at-risk individuals before illness strikes, opening a window for personalized prevention.

The study is not without limitations, which the authors acknowledge candidly. The sample, though larger than most brain-wide association studies, is modest by the standards now considered necessary for robust individual-differences research, and no fully independent external dataset with comparable movie fMRI and personality data currently exists. Trait-relevance was rated after viewing rather than measured during scanning, no direct measure of attention such as eye-tracking was collected, and demographic covariates cannot yet be formally controlled within the IS-RSA framework. Even so, the work stands as a compelling demonstration that who we are shapes how our brains respond to the world—and that the right moments in a film can bring those hidden differences into sharp focus.

Subject of Research: Neural correlates of neuroticism during naturalistic movie-watching fMRI

Article Title: Trait-relevant movies amplify neuroticism-associated brain responses

Article References: Popp, J. L., Weiß, M., Faskowitz, J., & Hilger, K. (2026). Trait-relevant movies amplify neuroticism-associated brain responses. Communications Psychology, 4(1), Article 134. https://doi.org/10.1038/s44271-026-00541-5

Image Credits: AI Generated

DOI: 10.1038/s44271-026-00541-5

Keywords: neuroticism, personality neuroscience, fMRI, naturalistic stimuli, inter-subject correlation, IS-RSA, Human Connectome Project, mental health risk, default mode network, attentional control, psychopathology, movie watching

The research, led by Yuhang Wu, Shi Gu, and colleagues at Zhejiang University, the University of Electronic Science and Technology of China, New York University, and the University of Pennsylvania, including network neuroscientist Dani S. Bassett, took a deliberately controlled approach. Rather than embedding their networks in physical space and imposing wiring-cost constraints, the team trained recurrent neural networks (RNNs) on batteries of cognitive tasks of the kind long used in systems neuroscience: working memory, perceptual decision-making, context-dependent categorization, and other tasks that probe the computational repertoire of prefrontal and parietal cortex. The question was simple but profound: if you strip away all spatial and metabolic pressure, does modularity still appear when a network must learn many demanding tasks at once?

The answer was a resounding yes. Networks trained under multitask learning paradigms developed significantly higher modularity than networks trained on a single task, and the effect grew stronger as the task load pushed against the network’s capacity. When the number of simultaneous tasks strained what the fixed pool of units could compute, the networks responded by reorganizing their internal connectivity into functionally segregated communities. This is a striking result because nothing in the training objective rewarded modularity directly. The networks were optimized only for task performance, yet the pressure of limited capacity and diverse demands was sufficient to carve the connectivity matrix into modules—much as the pressure of diverse cognitive demands may have shaped the brain’s own architecture.

The study went further by comparing different training regimes. Networks trained with incremental multitask learning—in which tasks were introduced sequentially and the network had to integrate each new demand into an already functioning system—developed the highest degree of modularity of all, while also maintaining superior performance across the full task set. This detail matters because it mirrors the developmental trajectory of biological brains, which do not acquire all cognitive abilities simultaneously but build them progressively over years of experience. The finding suggests that the order and pacing of task acquisition, not merely the total computational load, shapes the topology that emerges. Modularity, in this view, is not a static design feature but an adaptive response to the sequential introduction of complex problems.

Technically, the team quantified modularity using established network-science measures, including community-detection methods of the kind pioneered by Leicht and Newman for directed networks, applied to the learned weight matrices of the RNNs. They tracked how modular structure unfolded over the course of training, revealing that community boundaries sharpened as learning progressed and as additional tasks accumulated. They also examined the incremental addition of connections during learning, drawing an intriguing parallel to the lottery ticket hypothesis from deep learning research—the idea that sparse, trainable subnetworks exist within larger networks and are the components that effectively carry the computational load. In the task-trained RNNs, sparse modular substructures appeared to play an analogous role, suggesting a possible computational rationale for why both artificial and biological learning systems might favor segregated, sparsely interconnected architectures.

Perhaps the most consequential finding came when the researchers compared their task-induced networks against biological data. Using structural connectivity data from the Human Connectome Project, covering 84 cortical areas, they evaluated how closely the artificial networks matched the brain’s own organization. The task-trained networks exhibited structural properties that more closely resembled biological brain networks than models based solely on spatial constraints such as wiring-cost minimization. In other words, functional demand—the need to compute—appears to be a stronger organizing principle for brain-like topology than physical economy alone. This does not mean spatial constraints are irrelevant; the brain is certainly shaped by the geometry of the skull and the metabolic cost of axons. But the new results demonstrate that spatial models alone cannot fully explain the functional organization of brain networks, and that computational pressure fills a substantial part of that explanatory gap.

The work builds on a rich lineage of research at the intersection of machine learning and neuroscience. Previous studies had shown that RNNs trained on many cognitive tasks develop mixed selectivity and shared dynamical motifs that support flexible behavior, and that spatially embedded RNNs recapitulate numerous structural and functional findings from neuroscience. Other work demonstrated that brain-like functional specialization can emerge spontaneously in deep networks trained on naturalistic tasks. The new study adds a crucial piece: a controlled computational demonstration that modularity itself—arguably the signature feature of brain network organization—can be induced purely by the functional demands of multitask learning under capacity constraints. It thereby offers a causal, mechanistic account where earlier work offered correlations or spatial explanations.

The implications run in both directions. For neuroscience, the study provides a testable framework: if modularity is an adaptive response to cumulative cognitive demands, then developmental changes in brain network segregation should track the acquisition of complex abilities, a hypothesis consistent with prior findings that modular segregation of structural brain networks supports the development of executive function in youth. For artificial intelligence, the results hint at design principles for more adaptable machines. Modular deep learning has become a vibrant field precisely because modular systems can learn new skills without catastrophically forgetting old ones, and this study suggests that simply structuring the training curriculum—introducing tasks incrementally under realistic capacity limits—can coax modularity into existence without hand-engineered architectural constraints. That could inform how researchers build continual-learning systems, from robotics to large multimodal models, where flexibility and stability must coexist.

The study also speaks to a long-standing debate about the economy of brain network organization. The brain has often been described as a compromise between wiring cost and topological value, with small-world architecture emerging from that trade-off. The new findings suggest the ledger has more entries than previously appreciated: computational value, capacity limits, and the temporal sequence of learning demands all leave structural fingerprints. Modularity may be less a consequence of saving wire and more a consequence of solving problems—a functional adaptation that spatial economy then refines rather than creates. As the authors put it in their abstract, modularization emerges as an adaptive response to the sequential introduction of complex tasks, a framing that reframes the brain’s architecture as the product of a computational curriculum written by evolution and experience.

The team has released its code and processed connectivity data publicly via GitHub and Zenodo, allowing other researchers to reproduce the simulations and extend the approach to new task sets, architectures, and species comparisons. Cross-species network comparisons, representational similarity analyses, and generative models of the connectome are natural next steps, and the framework could eventually be applied to clinical questions, since many neuropsychiatric conditions involve disruptions of modular brain organization. For now, the study stands as a vivid example of a growing trend in modern science: artificial neural networks are no longer just engineering tools but instruments for asking why questions about the brain—and, increasingly, they are answering with structures that look strikingly familiar. When a machine learns the way a mind must, it begins, it seems, to build itself the way a brain is built.

Subject of Research: Emergence of task-driven modular network organization in recurrent neural networks and its alignment with human brain architecture

Article Title: Task-structured modularity emerges in artificial networks and aligns with brain architecture

Article References: Wu, Y., Deng, S., Du, K., Mattar, M. G., Wu, Y., Bassett, D. S., Tang, H., Pan, G., & Gu, S. (2026). Task-structured modularity emerges in artificial networks and aligns with brain architecture. Nature Machine Intelligence. https://doi.org/10.1038/s42256-026-01306-9

Image Credits: AI Generated

DOI: 10.1038/s42256-026-01306-9

Keywords: recurrent neural networks, multitask learning, modularity, brain networks, connectome, cognitive tasks, network neuroscience, Human Connectome Project, continual learning, wiring cost, lottery ticket hypothesis, Nature Machine Intelligence

Daniele Orzechowski of the Federal University of Santa Catarina and Ronaldo Martins da Costa of the Federal University of Goiás have built and released a containerised pipeline for Linear Mixed-Effects (LME) analysis of task fMRI data, and they have stress-tested it against one of the largest neuroimaging datasets in existence: the Human Connectome Project. Their work, published in September 2026, is notable not only for what the new model finds but for how transparently the tool reports its own behaviour. The pipeline requires the user to declare which fitting engine is being used, refuses to silently substitute one for another, versions every output file so repeated runs cannot overwrite each other, and writes a machine-readable manifest recording the software versions, hardware, and parameters behind every step. In a field still reckoning with reproducibility failures, those design choices are as consequential as the statistics themselves.

The validation is unusually thorough. The authors cross-validated their Python implementation, built on the statsmodels library, against the R reference standard lme4 by fitting identical data from 21 brain regions under both engines in separately archived executions. The two implementations agreed on the intraclass correlation coefficient, a key variance measure, to a mean absolute difference of 2.2 multiplied by ten to the power of minus six, with significance concordance in all 21 regions. Monte Carlo simulations across five variance regimes confirmed that the pipeline recovers known parameters with only tiny biases. The only practical divergence between the two backends was inferential rather than numerical: R’s lmerTest applies Satterthwaite degrees of freedom while statsmodels uses Wald z-tests, which placed a single region, the right hippocampus, on opposite sides of the 0.05 threshold even though the underlying estimates agreed to within one part in a million.

For the case study, the team turned to the emotion task from the Human Connectome Project S1200 release, in which participants view fearful faces and match neutral shapes in a blocked design. After quality control excluded three runs with excessive head motion, 142 healthy adults aged 22 to 35 contributed 283 functional runs acquired on a 3 Tesla scanner with 2-millimetre resolution. Each run produced a beta map for the fearful-face-versus-shape contrast, and all 283 maps entered the mixed-effects model, which modeled a fixed intercept, a fixed effect for run type, a random intercept per subject, and residual noise. The intraclass correlation coefficient then quantified how much of the variance was between subjects rather than within them.

The headline result is a surprise that reframes a long-running methodological debate. Compared head to head over an identical mask of 215,265 voxels with the same false discovery rate correction, the full LME model flagged 74,004 significant voxels while the conventional GLM flagged 108,627, a ratio of 1.47, with a Dice overlap of 0.77. That looks like the familiar story of mixed-effects models being more conservative. But the authors then fitted an intermediate LME model that kept the hierarchical structure yet dropped the run covariate, and it produced 109,273 significant voxels, almost indistinguishable from the GLM, with a correlation of 0.994 and a Dice of 0.993. The entire sensitivity difference, in other words, comes from the run covariate absorbing systematic run-related variance, not from hierarchical variance decomposition itself. Previous reports of reduced LME sensitivity may have conflated two distinct effects.

The run effects themselves proved scientifically interesting. They were significant in 13 of the 21 regions of interest, with the largest in the right amygdala, and they survived adjustment for framewise displacement, a head-motion measure, at the group level, on top of the twelve motion regressors already included in every first-level model. Because the Human Connectome Project acquires its two runs in a fixed order without counterbalancing, the authors caution that this effect may partly reflect practice, adaptation, or fatigue rather than the phase-encoding direction of acquisition, and they suggest reading it as a run-and-order effect. Still, the finding demonstrates the kind of signal that summary-statistics approaches simply average away.

Perhaps the most reusable product of the study is a set of whole-brain, voxelwise ICC maps covering all 215,265 voxels, released unthresholded. The maps reveal a pronounced subcortical-to-cortical gradient in within-subject reliability: the mean voxelwise ICC was 0.199, with values reaching 0.412 in the left amygdala and around 0.37 in the amygdala on average, while lateral frontal cortex sat near 0.10 to 0.14 and the frontal pole was effectively zero. High-ICC regions such as the amygdala are the ones most suitable for brain-behavior correlation studies and biomarker development, whereas low-ICC regions show high within-subject variability that conflates genuine neural fluctuation, measurement noise, and model misspecification. The authors also quantify how imprecise two-run ICC estimates are, with a simulation root-mean-square error of 0.081, improving substantially with three or more runs per subject.

Computationally, the pipeline makes whole-brain LME genuinely feasible. Benchmarked on synthetic data, the mixed model cost 19 times more per voxel than the GLM, 37.2 milliseconds against 1.99, but on real data the full analysis of 215,265 voxels completed in 5 hours and 41 minutes on twelve processor cores, with 100 percent convergence as read from the optimizer’s own flag. The authors note that convergence rates computed merely from the absence of exceptions, a common shortcut, overstate model fit. Stability checks reinforced confidence in both methods: split-half reliability exceeded 0.98 for the GLM and the LME alike, and leave-one-out cross-validation showed the group estimates were robust to dropping any individual subject.

So when should researchers abandon the GLM? The study’s answer is nuanced. For single-run designs or rapid exploratory work, the GLM remains a practical and nearly equivalent choice, since hierarchical modeling alone produces negligible divergence from summary statistics. LME earns its cost when multiple runs are available, when run-level covariates need explicit modeling, when conservative inference is a priority, or when within-subject reliability is itself the question. The decomposition strategy the authors propose, fitting full, intermediate, and standard models over an identical voxel set, generalises to any run-level covariate and can serve as a routine diagnostic for attributing sensitivity differences between statistical frameworks.

The pipeline is released under an MIT license with a Dockerfile and pinned environment manifests, and it is designed to travel: the task labels and atlases are configuration values rather than code, the statistical core is indifferent to whether data arrive as volumes or surface vertices, and the authors flag the two dataset-specific conventions, motion-regressor file layouts and rotation units, most likely to break outside the Human Connectome Project. Limitations remain, including a single contrast and dataset, a random-intercept-only specification forced by the two-run design, and volumetric rather than surface-based analysis. But the message to the field is clear: mixed-effects modeling of task fMRI is no longer a statistical luxury or a computational gamble. It is a validated, benchmarked, reproducible tool that reveals what the standard approach has been quietly discarding, and it now runs on hardware most labs can access overnight.

Subject of Research: Linear mixed-effects modeling of task fMRI data from the Human Connectome Project

Article Title: Beyond the General Linear Model: Linear Mixed-Effects Modeling of fMRI Data from the Human Connectome Project

Article References: Orzechowski, D., & Martins da Costa, R. (2026). Beyond the General Linear Model: Linear Mixed-Effects Modeling of fMRI Data from the Human Connectome Project. Neuroinformatics, 24(4), Article 64. https://doi.org/10.1007/s12021-026-09820-2

Image Credits: AI Generated

DOI: 10.1007/s12021-026-09820-2

Keywords: fMRI, linear mixed-effects models, General Linear Model, Human Connectome Project, intraclass correlation coefficient, neuroimaging, reproducibility, statsmodels, lme4, emotion task, brain reliability, Neuroinformatics