We introduce LibriBrain100, a large-scale MEG dataset for speech decoding designed from the ground up for reproducible, standardised evaluation. LibriBrain100 more than doubles the size of the original LibriBrain release, resulting in over 100 hours of high-quality MEG acquired while subjects listened to naturalistic continuous speech. With ∼80 hours from a single subject, LibriBrain100 sets a new record for deep, within-subject neural data (8× more than the next comparable dataset and roughly 80× more than other datasets). To demonstrate the payoff of this depth-first design, we evaluate on a word-classification benchmark—an increasingly well-established stepping stone towards the open challenge of noninvasive brain-to-text decoding. Using an existing decoding model, we achieve state-of-the-art performance—validating both the quality of the recordings and the value of within-subject data at scale. Because collecting 80 hours of data per user is impractical for real-world applications, we also collected ∼40 minutes of additional data from each of 32 subjects. Using the same word-classification benchmark, we demonstrate the value of broad multi-subject data: supervised finetuning of a pre-trained model can substantially compensate for limited per-subject data. We provide standard train, validation, and test splits, all reproducible through an open-sourced Python library that supports easy downloading, optional preprocessing, and data loading for common deep learning frameworks. In addition, the dataset and evaluation infrastructure are being released alongside an open machine-learning competition with a public leaderboard for standardised benchmarking. Ultimately, our hope is that LibriBrain100 will accelerate progress towards practical non-invasive brain-computer interfaces, capable of restoring communication to people living with severe paralysis.

The ambition of the 2025 PNPL competition (Landau et al., 2025) was to launch a multi-year curriculum for non-invasive speech decoding. Designed to progress from foundational tasks toward the linguistic complexity required for a practical brain-computer interface (BCI), it set the stage with speech detection and phoneme classification tasks. Winning submissions reached F1-macro scores of 95.6 The 2026 PNPL competition responds to this challenge with LibriBrain100 (Mantegna et al., 2026), an extended LibriBrain dataset with 32 additional subjects (∼40 minutes each) plus even more within-subject data (∼80 hours). Advancing the curriculum of tasks to focus on word classification, two complementary tracks are presented in this competition: the Deep track targets within-subject word classification at scale, aiming at the best possible performance; the Broad track targets cross-subject generalisation, progressively reducing the amount of subject-specific fine-tuning data from ∼40 to ∼20 to ∼10 minutes, a duration that falls within a clinically feasible range and brings us a step closer to a non-invasive BCI capable of restoring communication to people living with profound paralysis.

Modelling the complex spatio-temporal patterns of large-scale brain dynamics is crucial for neuroscience, but traditional methods fail to capture the rich structure in modalities such as magnetoencephalography (MEG). Recent advances in deep learning have enabled significant progress in other domains, such as language and vision, by using foundation models at scale. Here, we introduce MEG-GPT, a transformer-based foundation model that uses time-attention and next time-point prediction. To facilitate this, we also introduce a novel data-driven tokeniser for continuous MEG data, which preserves the high temporal resolution of continuous MEG signals without lossy transformations. We trained MEG-GPT on tokenised brain region time courses extracted from a large-scale MEG dataset (N = 612, eyes-closed rest, Cam-CAN data), and show that the learnt model can generate data with realistic spatio-spectral properties, including transient events and population variability. Critically, it performs well in downstream decoding tasks, improving downstream supervised prediction task, showing improved zero-shot generalisation across sessions (improving accuracy from 0.56 to 0.59) and subjects (improving accuracy from 0.45 to 0.49) compared with a PCA baseline method. Furthermore, we show the model can be efficiently fine-tuned on a smaller labelled dataset to boost performance in cross-subject decoding scenarios. This work establishes a powerful foundation model for electrophysiological data, paving the way for applications in computational neuroscience and neural decoding.

Abstract Functional connectivity (FC) profiles derived from fMRI and MEG offer complementary perspectives on large-scale brain organization, while showing reasonable correspondence at the population-average level. However, how their individual variability relates between these modalities remains unclear. Using the Cam-CAN dataset, we derived neural fingerprints from subject-level resting-state fMRI FC and MEG FC obtained from the same participants (N=543). Fingerprints derived from each modality separately showed robust within-subject, cross-session consistency and successfully predicted age and cognition, confirming that these features capture stable and behaviourally relevant individual traits. We then quantified shared individual variability between modalities using variance partitioning analyses and representational similarity measures. Two main findings emerged. First, despite strong similarity at the population-average level, correspondence between MEG and fMRI neural fingerprints at the subject level was low, as reflected in both cross-modal shared variance and the preservation of pairwise inter-subject similarity patterns, quantified by linear Centred Kernel Alignment (CKA). Second, structural fingerprints accounted for the majority of age-related variance in functional neural fingerprints, almost entirely explaining the age-related variance in, and shared between, fMRI and MEG. MEG functional fingerprints did have unique information not accounted for by structure when explaining variability in cognitive traits, but this was not shared with fMRI. Together, these findings demonstrate that there is a surprisingly lack of similarity in the way that subjects vary between fMRI and electrophysiology, especially when structural variability is accounted for.

Abstract Transcranial magnetic stimulation (TMS) targeting the left dorsolateral prefrontal cortex is known to progressively reduce symptoms of depression. However, the neural mechanisms supporting this effect are poorly understood. To address this gap, we analysed longitudinal EEG recordings from 70 people undergoing TMS therapy and fitted an established dynamic network model of resting-state activity. Greater baseline symptom severity was associated with reduced occupancy of and fewer transitions into an anterior default mode brain state, alongside increased activity in a posterior default mode state. During treatment, decreases in anterior default mode state engagement following TMS predicted symptom improvement in the latter half of the intervention. Brain state activity exhibited structured, cyclical dynamics, with slower cycles linked to greater baseline severity. These findings suggest that symptoms of depression are characterised by gradual alterations in brain state dynamics, highlighting a central and dissociable role of default mode brain states in the persistence and remission of symptoms.

Functional neuroimaging techniques allow us to estimate functional networks that underlie cognition. However, these functional networks are often estimated at the group level and do not allow for the discovery of, nor benefit from, subpopulation structure in the data, that is, the fact that some recording sessions may be more similar than others. Here, we propose the use of embedding vectors (c.f. word embedding in Natural Language Processing) to explicitly model individual sessions while inferring networks across a group. This vector is effectively a "fingerprint" for each session, which can cluster sessions with similar functional networks together in a learnt embedding space. We apply this approach to estimate dynamic functional networks using a hierarchical Hidden Markov Model (HMM). We call this approach HIVE (HMM with Integrated Variability Estimation). Using simulated data, we show that HIVE can uncover true subpopulation structure and show improved performance over existing approaches. Using real magnetoencephalography data, we show the learnt embedding vectors (session fingerprints) reflect meaningful sources of variation across a population. Overall, HIVE provides a powrful new approach for modelling individual sessions while leveraging information available across an entire group.

To accelerate new treatments for Alzheimer’s disease, there is the need for human pathophysiological biomarkers that are sensitive to treatment and disease mechanisms. In this proof-of-concept study, we assess new biophysical models of non-invasive human MEG imaging to test the pharmacological and disease modulation of NMDA-receptor inhibition. Magnetoencephalography was recorded during an auditory mismatch negativity paradigm from (1) neurologically-healthy people on memantine or placebo (n = 19, placebo-controlled crossover design); (2) people with Alzheimer’s disease at baseline and 16-months (n = 42, amyloid-biomarker positive, longitudinal observational design). Optimised dynamic causal models inferred voltage-dependent NMDA-receptor blockade using Parametric Empirical Bayes to test group effects. The mismatch negativity amplitude was attenuated when Alzheimer’s disease was more severe (lower baseline mini-mental state examination) and after follow-up (versus baseline). Memantine increased NMDA-receptor inhibition, compared to placebo. Alzheimer’s disease reduced NMDA-receptor inhibition in proportion to severity and over time. In line with preclinical studies, we confirm in humans that memantine and Alzheimer’s disease have opposing effects on NMDA-receptor inhibition. The ability to infer such receptor dynamics and pharmacology from non-invasive physiological recordings has wide applications, including the assessment of other neurological disorders and novel drugs intended for symptomatic or disease-modifying treatments.

OBJECTIVE:A non-invasive measure of cerebral motor system dysfunction would be valuable as a biomarker in amyotrophic lateral sclerosis (ALS). Task-based magnetoencephalography (tMEG) measures the magnetic fields generated by cortical neuronal oscillatory activity during task performance. Gamma activations are periods of high-power and high-frequency cortical oscillations integral to motor control. METHODS:tMEG was undertaken during 60 bilateral isometric hand grip exercises in ALS (n = 42) and compared with healthy controls (HC, n = 33). Gamma activation spread (GAS) was estimated by calculating the number of activated regions during each 100 ms time-bin and compared statistically between groups. Gamma activation patterns were visualised by plotting each participant's brain activity separately as a 2-dimensional video. RESULTS:There was no difference in grip strength between groups. GAS was greatly increased in the ALS group compared to HC (p < 0.001) and correlated positively with rate of ALSFRS-R progression (t = 1.35, p = 0.023) and a fine motor sub-score (t = -1.18, p = 0.047). CONCLUSIONS:ALS was associated with a marked increase in regional spread of gamma frequency activation, greater in those with higher disease progression rates. SIGNIFICANCE:The regional spread of gamma activity may reflect disease activity in ALS, with potential application as an experimental medicine readout.

The advance of speech decoding from non-invasive brain data holds the potential for profound societal impact. Among its most promising applications is the restoration of communication to paralysed individuals affected by speech deficits such as dysarthria, without the need for high-risk surgical interventions. The ultimate aim of the 2025 PNPL competition is to produce the conditions for an "ImageNet moment" or breakthrough in non-invasive neural decoding, by harnessing the collective power of the machine learning community. To facilitate this vision we present the largest within-subject MEG dataset recorded to date (LibriBrain) together with a user-friendly Python library (pnpl) for easy data access and integration with deep learning frameworks. For the competition we define two foundational tasks (i.e. Speech Detection and Phoneme Classification from brain data), complete with standardised data splits and evaluation metrics, illustrative benchmark models, online tutorial code, a community discussion board, and public leaderboard for submissions. To promote accessibility and participation the competition features a Standard track that emphasises algorithmic innovation, as well as an Extended track that is expected to reward larger-scale computing, accelerating progress toward a non-invasive brain-computer interface for speech.

Information processing in the brain spans from localised sensorimotor processes to higher-level cognition that integrates across multiple regions. Interactions between and within these subsystems enable multiscale information processing. Despite this multiscale characteristic, functional brain connectivity is often either estimated based on 10-30 distributed modes or parcellations with 100-1000 localised parcels, both missing across-scale functional interactions. We present Multiscale Probabilistic Functional Modes (mPFMs), a new mapping which comprises modes over various scales of granularity, thus enabling direct estimation of functional connectivity within- and across-scales. Crucially, mPFMs emerged from data-driven multilevel Bayesian modelling of large functional MRI (fMRI) populations. We demonstrate that mPFMs capture both distributed brain modes and their co-existing subcomponents. In addition to validating mPFMs using simulations and real data, we show that mPFMs can predict ~900 personalised traits from UK Biobank more accurately than current standard techniques. Therefore, mPFMs can offer a paradigm shift in functional connectivity modelling and yield enhanced fMRI biomarkers for traits and diseases.

BACKGROUND:Stress leads to neurobiological changes, and failure to regulate these can contribute to chronic psychiatric issues. Despite considerable research, the relationship between neural alterations in acute stress and coping with chronic stress is unclear. This longitudinal study examined whole-brain network dynamics following induced acute stress and their role in predicting chronic stress vulnerability. METHODS:Sixty military pre-deployment soldiers underwent a lab-induced stress task where subjective stress and resting-state functional magnetic resonance imaging were acquired repeatedly (before stress, after stress, and at recovery, 90 min later). Baseline depression and post-traumatic stress symptoms were assessed, and again a year later during military deployment. We used the Leading Eigenvector Dynamic Analysis framework to characterize changes in whole-brain dynamics over time. Time spent in each state was compared across acute stress conditions and correlated with psychological outcomes. RESULTS:Findings reveal significant changes at the network level from acute stress to recovery, where the frontoparietal and subcortical states decreased in dominance in favor of the default mode network, sensorimotor, and visual states. A significant normalization of the frontoparietal state activity was related to successful psychological recovery. Immediately after induced stress, a significant increase in the lifetimes of the frontoparietal state was associated with higher depression symptoms (r = 0.49, p < .02) and this association was also observed a year later following combat exposure (r = 0.49, p < .009). CONCLUSIONS:This study revealed how acute stress-related neural alterations predict chronic stress vulnerability. Successful recovery from acute stress involves reducing cognitive-emotional states and enhancing self-awareness and sensory-perceptual states. Elevated frontoparietal activity is suggested as a neural marker of vulnerability to chronic stress.

We describe OHBA Software Library for the analysis of electrophysiology data (osl-ephys). This toolbox builds on top of the widely used MNE-Python package and provides unique analysis tools for magneto-/electro-encephalography (M/EEG) sensor and source space analysis, which can be used modularly. In particular, it facilitates processing large amounts of data using batch parallel processing, with high standards for reproducibility through a config API and log keeping, and efficient quality assurance by producing HTML processing reports. It also provides new functionality for doing coregistration, source reconstruction and parcellation in volumetric space, allowing for an alternative pipeline that avoids the need for surface-based processing, e.g., through the use of Fieldtrip. Here, we introduce osl-ephys by presenting examples applied to a publicly available M/EEG data (the multimodal faces dataset). osl-ephys is open-source software distributed on the Apache License and available as a Python package through PyPi and GitHub.

The past few years have seen remarkable progress in the decoding of speech from brain activity, primarily driven by large single-subject datasets. However, due to individual variation, such as anatomy, and differences in task design and scanning hardware, leveraging data across subjects and datasets remains challenging. In turn, the field has not benefited from the growing number of open neural data repositories to exploit large-scale deep learning. To address this, we develop neuroscience-informed self-supervised objectives, together with an architecture, for learning from heterogeneous brain recordings. Scaling to nearly 400 hours of MEG data and 900 subjects, our approach shows generalisation across participants, datasets, tasks, and even to novel subjects. It achieves improvements of 15-27% over state-of-the-art models and matches surgical decoding performance with non-invasive data. These advances unlock the potential for scaling speech decoding models beyond the current frontier.

INTRODUCTION:Alzheimer's disease (AD) affects neurophysiology by loss of neurons, synapses, and neurotransmitters. A mechanistic understanding of the human disease will facilitate new treatments. METHODS:Magnetoencephalography was recorded during an auditory mismatch negativity paradigm from healthy adults (n = 14) and people with symptomatic AD (n = 45, amyloid biomarker positive) at baseline and after 16 months. Fourteen people with AD had repeat magnetoencephalography at 2 weeks to assess test-retest reliability. Dynamic causal models were fitted to the evoked responses and analyzed using parametric empirical Bayes. RESULTS:Sensor data confirmed that AD and its progression reduce the mismatch negativity amplitude, which had excellent test-retest reliability. Parametric empirical Bayes analyses confirmed that AD progressively reduces extrinsic connectivity between pyramidal cells and superficial pyramidal cell gain modulation. DISCUSSION:Dynamic causal modeling revealed cellular-level causes of the neurophysiological deficits observed in AD. This approach may help facilitate experimental medicine studies of candidate treatments. HIGHLIGHTS:Magnetoencephalography scanning provides reliable biomarkers that are sensitive to Alzheimer's disease (AD) and its progression, and informative about disease mechanisms underlying cognitive decline. In vivo assays of pyramidal cell function during cognitive processes in humans improve our understanding of AD mechanisms. The amplitude of the mismatch negativity response is progressively reduced in AD. Reduced pyramidal cell gain and connectivity underlie this neurophysiological deficit. These measures are potential biomarkers for interventional studies.

There is growing interest in studying the temporal structure in brain network activity, in particular, dynamic functional connectivity (FC), which has been linked in several studies with cognition, demographics and disease states. The sliding window approach is one of the most common approaches to compute dynamic FC. However, it cannot detect cognitively relevant and transient temporal changes at time scales of fast cognition, that is, on the order of 100 ms, which can be identified with model-based methods such as the HMM (Hidden Markov Model) and DyNeMo (Dynamic Network Modes) using electrophysiology. These new methods provide time-varying estimates of the 'power' (i.e., variance) and of the functional connectivity of the brain activity, under the assumption that they share the same dynamics. But there is no principled basis for this assumption. Using a new method that allows for the possibility that power and FC networks have different dynamics (Multi-dynamic DyNeMo) on resting-state magnetoencephalography (MEG) data, we show that the dynamics of the power and the FC networks are not coupled. Using a (visual) task MEG dataset, we show that the power and FC network dynamics are modulated by the task, such that the coupling in their dynamics changes significantly during the task. This work reveals novel insights into evoked network responses and ongoing activity that previous methods fail to capture, challenging the assumption that power and FC share the same dynamics.

Non-invasive recordings of magnetoencephalography have been used for developing biomarkers for neural changes associated with Parkinson's disease that can be measured across the entire course of the disease. These studies, however, have yielded inconsistent findings. Here, we investigated whether analysing motor cortical activity within the context of large-scale brain network activity provides a more sensitive marker of changes in Parkinson's disease using magnetoencephalography. We extracted motor cortical beta power and beta bursts from resting-state magnetoencephalography scans of patients with Parkinson's disease (N = 28) and well-matched healthy controls (N = 36). To situate beta bursts in their brain network contexts, we used a time-delay-embedded hidden Markov model to extract brain network activity and investigated co-occurrence patterns between brain networks and beta bursts. Parkinson's disease was associated with decreased beta power in motor cortical power spectra, but no significant differences in motor cortical beta-burst dynamics occurred when using a conventional beta-burst analysis. Dynamics of a large-scale sensorimotor network extracted with the time-delay-embedded hidden Markov model approach revealed significant decreases in the occurrence of this network with Parkinson's disease. By comparing conventional burst and time-delay-embedded hidden Markov model state occurrences, we observed that motor beta bursts occurred during both sensorimotor and non-sensorimotor network activations. When using the large-scale network information provided by the time-delay-embedded hidden Markov model to focus on bursts that were active during sensorimotor network activations, significant decreases in burst dynamics could be observed in patients with Parkinson's disease. In conclusion, our findings suggest that decreased motor cortical beta power in Parkinson's disease is prominently associated with changes in sensorimotor network dynamics using magnetoencephalography. Thus, investigating large-scale networks or considering the large-scale network context of motor cortical activations may be crucial for identifying alterations in the sensorimotor network that are prevalent in Parkinson's disease and might help resolve contradicting findings in the literature.

Replay has been implicated in organising experiences into cognitive maps offline, yet how this process evolves through development remains unclear. We studied 106 participants (ages 8–25) who learned a hidden two-dimensional (2D) structure and then underwent magnetoencephalography (MEG) during a map-based inference task and subsequent rest, allowing us to detect spontaneous replay. Younger participants relied more on replay alone to represent the 2D associations, whereas older participants showed increasingly precise alignment between replay and the default mode network (DMN), particularly when replay events were timed to the DMN theta (2–6 Hz) trough. This alignment further predicted grid-cell-like codes in the entorhinal cortex, previously identified in the same cohort using fMRI. Resting-state fMRI indicated that DMN connectivity also strengthened with age, which explained reduced reliance on replay and faster inference across development. These findings illuminate a developmental progression where replay shifts from an isolated hippocampal process to a coordinated hippocampal–DMN mechanism. This shift may underpin maturing grid-like schema representations, offering insight into how children gradually build internal knowledge structures for flexible inference. ### Competing Interest Statement The authors have declared no competing interest.

BACKGROUND:Systems neuroscience studies have shown that baseline brain activity can be categorized into large-scale networks (resting-state-networks, RNSs), with influence on cognitive abilities and clinical symptoms. These insights have guided millimeter-precise selection of brain stimulation targets based on RSNs. Concurrently, Transcranial Magnetic Stimulation (TMS) studies revealed that baseline brain states, measured by EEG signal power or phase, affect stimulation outcomes. However, EEG dynamics in these studies are mostly limited to single regions or channels, lacking the spatial resolution needed for accurate network-level characterization. OBJECTIVE:We aim at mapping brain networks with high spatial and temporal precision and to assess whether the occurrence of specific network-level-states impact TMS outcome. To this end, we will identify large-scale brain networks and explore how their dynamics relates to corticospinal excitability. METHODS:This study leverages Hidden Markov Models to identify large-scale brain states from pre-stimulus source space high-density-EEG data collected during TMS targeting the left primary motor cortex in twenty healthy subjects. The association between states and fMRI-defined RSNs was explored using the Yeo atlas, and the trial-by-trial relation between states and corticospinal excitability was examined. RESULTS:We extracted fast-dynamic large-scale brain states with unique spatiotemporal and spectral features resembling major RSNs. The engagement of different networks significantly influences corticospinal excitability, with larger motor evoked potentials when baseline activity was dominated by the sensorimotor network. CONCLUSIONS:These findings represent a step forward towards characterizing brain network in EEG-TMS with both high spatial and temporal resolution and underscore the importance of incorporating large-scale network dynamics into TMS experiments.

We propose the Gaussian-Linear Hidden Markov model (GLHMM), a generalisation of different types of HMMs commonly used in neuroscience. In short, the GLHMM is a general framework where linear regression is used to flexibly parameterise the Gaussian state distribution, thereby accommodating a wide range of uses-including unsupervised, encoding, and decoding models. GLHMM is available as a Python toolbox with an emphasis on statistical testing and out-of-sample prediction-that is, aimed at finding and characterising brain-behaviour associations. The toolbox uses a stochastic variational inference approach, enabling it to handle large data sets at reasonable computational time. The GLHMM can work with various types of data, including animal recordings or non-brain data, and is suitable for a broad range of experimental paradigms. For demonstration, we show examples with fMRI, local field potential, electrocorticography, magnetoencephalography, and pupillometry.