Nature Neuroscience

Conformal prediction is increasingly proposed for clinical decision support because it provides distribution-free coverage guarantees that hold for any model and any data distribution. The guarantee routinely reported, however, is marginal, and we show that in multi-class diagnostic classification it can be satisfied exactly while individual diagnoses are covered very unevenly. On a four-way mood and psychosis classification task over 1,520 subjects from three studies and 14 acquisition sites, marginal split-conformal calibration attained 0.9000 empirical coverage against a nominal 0.90, while healthy controls were covered at 0.941 and schizoaffective disorder at 0.818. The surplus and the deficit cancel, so the aggregate figure is uninformative about either. Class-conditional (Mondrian) calibration reduced this 12.3-point disparity to 0.4 points at a cost of 0.07 labels in mean set size, under 3%. Holding coverage fixed across diagnoses has a further consequence: the residual variation in set size can no longer be attributed to class prior or per-class accuracy, and becomes interpretable as a property of the subject. We show that the resulting prediction sets separate subjects into confident, boundary, ambiguous and unresolved strata, and that the proportion of subjects independently flagged as label-ambiguous by a structural-MRI model trained separately on the same cohort rises monotonically across these strata (34.1%, 57.3%, 68.1%, 81.8%; p = 8.8e-18). No schizoaffective subject and 0.9% of bipolar subjects reach the confident stratum, against 18.0% of controls and 13.4% of schizophrenia subjects. Set size at matched coverage also provides a comparison between representations that accuracy cannot. Applied to structural MRI, functional MRI and their fusion, it reveals a sign reversal that accuracy conceals: fusion reduces set size for bipolar (-0.25) and schizoaffective (-0.23) subjects and increases it for controls (+0.16), and the aggregate fusion benefit itself changes sign below alpha = 0.10, while top-1 accuracy rises monotonically from 0.495 to 0.578 to 0.611 across the three models. Reporting marginal coverage alone is insufficient whenever diagnostic classes are unbalanced.

Understanding how the brain supports flexible, goal-directed behavior is a central aim of neuroscience. Yet such behavior is highly variable and unfolds over long timescales, making it challenging to understand its organization. Advances in behavioral tracking quantify movements but do not reveal the objectives that guide them. Motivated by this, we develop inverse reinforcement learning (IRL) into a general framework for inferring these objectives from naturalistic behavior. Our approach represents an animal's objectives as internal rewards, and infers these rewards as well as the policies they induce from observed behavior. It applies across diverse behavioral settings, from freely moving behaviors to structured tasks. It can infer objectives that switch over time or gradually change with experience. In mice capturing crickets, it identified a small set of objectives that the animals switched between over the course of a trial. In mice learning a cued labyrinth, it inferred how the animals' internal rewards changed as learning progressed. Across both datasets, the models predicted held-out actions and generated extended behavior that reproduced key features of animal performance. By transforming long, variable behavioral sequences into time-resolved estimates of the animal's objectives, our approach provides a quantitative basis for studying how the brain organizes natural behavior.

Can we tell what animals are dreaming about from their brain activity? This fascinating question has been explored extensively using the rodent hippocampus, where place cells are sequentially reactivated during sleep, consistent with replay of paths explored while awake. However, dreams involve vivid visual experiences, and whether analogous replays occur in the visual cortex has not been explored, that too in primates. To address this gap, we performed wireless neural recordings from the high-level visual (IT) and motor regions (PMv, PFC) in two monkeys while they viewed food images on a screen while awake, and afterwards while they slept in their natural setting. Decoders trained on sleep/wake neural activity in each region were able to accurately predict the ongoing sleep/wake state, as well as upcoming transitions between states. Interestingly, decoders trained to categorize neural responses to food (but not faces/objects) showed increased probability during REM-like sleep compared to NREM or awake states. Thus, behaviorally relevant visual stimuli seen while awake are reactivated during REM-like sleep. In other words, these monkeys saw food, faces and objects while awake, and dreamt selectively about food while sleeping.

Extracellular recordings carry information about neuronal location, and pooling per-unit predictions across neighboring electrodes improves its readout, but whether this local pooling should be learned is untested. We present MYCEL (Message-passing Yields Cellular Embeddings of Location), a graph neural network whose nodes are spike-sorted units carrying waveform and spike-timing features selected on cell-type ground truth, and whose edges encode recording geometry. On the Allen Brain Observatory Visual Coding Neuropixels dataset, learned message passing outperforms fixed aggregation, with the largest gains where single units are least informative. On the International Brain Laboratory Brain Wide Map, MYCEL matches the strongest population model at one-tenth the pairwise interactions. On held-out animals, the embeddings recover laminar depth within visual cortex. Retrained with only the edge geometry changed, the same framework decodes anatomy in macaque visual cortex and human medial temporal lobe. Thus, how neighboring neurons are combined matters as much as the features that describe them.

To act adaptively, animals must infer hidden states of the world from incomplete sensory information and update beliefs as new observations accrue. While dopamine signals are well explained by reinforcement learning models that incorporate belief states, how the brain implements belief-state inference remains unknown. Prior modeling work showed that recurrent neural networks trained to predict value (Value-RNN) develop task-specific, attractor-like dynamics that mirror evolution of beliefs. Here we performed high-density electrophysiological recordings from orbitofrontal cortex (OFC) and other brain areas of mice performing two variants of a Pavlovian task that differ in reward probability, which produce different within-trial belief dynamics. We find that OFC population activity exhibits task-specific attractor-like dynamics that mirror the within-trial dynamics in the Value-RNN. These dynamics are absent in motor and olfactory regions, are not explained by behavioral differences, and emerge progressively with learning. Our findings indicate that the brain approximates belief-state inference through learned, task-specific attractor-like dynamics.

The adult mammalian brain has limited regenerative capacity yet retains substantial potential for functional reorganization after experience, injury or disease1-3. While local plasticity at injury sites has been described4,5, the brain-wide consequences of a focal injury and the biological processes that drive them remain unknown. To address this, we performed an unbiased, whole brain screen of neuronal activity at single cell resolution following a focal injury to the primary visual cortex (V1) in mice. Olfactory processing areas emerged as loci of remote activation. We further show that V1 injury promotes the recruitment of adult born neurons into the olfactory bulb (OB) and potentiates cortical feedback onto bulbar circuits. Combining high-density multielectrode array recordings with two-photon calcium imaging revealed OB circuit refinement characterized by increased synchrony and sharpened tuning of principal output neurons. These circuit-level functional enhancements coincide with improved behavioral performance in odor-guided tasks. Motor cortex lesions do not elicit similar cellular and behavioral responses, suggesting that such distal adaptation only develops when injury is coupled with increased olfactory demand. Together, these data demonstrate that the adult brain can recruit alternative sensory circuits distant from a focal lesion to undergo adaptive, functionally relevant reorganization, and implicate adult neurogenesis as a contributing mechanism. Our findings expose a previously underappreciated degree of remote plasticity and reveal a novel role of adult neurogenesis in sensory compensation.

Learning from outcomes requires weighing each surprise by how much it is likely to mean. Two properties of the environment set that weight in opposite directions: volatility, the rate at which contingencies change, and stochasticity, the noise in outcomes around a stable contingency. Whether animals separate the two is unresolved, in part because varying reward probability in a reversal task makes outcomes noisier and the two options harder to tell apart at once. We trained male and female mice on a probabilistic reversal task for intracranial self-stimulation, scaling reward magnitude with reward probability so that the expected reward of each option and the difference between them stayed constant while outcome variance varied, and crossing this with different reversal rates. Simulations show that adapting the learning rate alone and adapting the decision policy alone are near-equivalent solutions across the environments we tested. Mice adopted the first adjustment: their fitted learning rate rose with volatility and fell with stochasticity, in both directions predicted by reward maximization, while their inverse temperature stayed fixed. This joint dependence was reproduced only by a meta-learning rule that reads stochasticity from the size of prediction errors and volatility from how often recent outcomes favored switching

Decisions often involve trade-offs between benefits and costs, such as a small payment now vs. a larger payment after a delay. Although decision variables are distributed across frontal cortex, how these regions transform competing attributes into a choice remains unclear. Combining large-scale electrophysiology with causal perturbations as mice chose between different reward amounts and delays, we found that population geometry of decision variables differed across regions, reflecting distinct computations. Dorsal frontal areas, including motor cortex, collapsed competing attributes onto a decision axis shared across reward-delay contexts, with projection amplitude predicting choice probability; silencing these areas impaired action selection regardless of delay. In contrast, ventral prefrontal cortex (vPFC) represented choice in higher-dimensional subspaces organized by reward delay, whose geometry predicted how strongly delay influenced choice. Notably, inhibiting vPFC preserved the dorsal decision representation but selectively removed its modulation by reward delay, causing animals to behave as if delay no longer influenced choice. vPFC inhibition had no effect when mice chose between reward amounts. Frontal cortex therefore solves multi-attribute decisions through a generator-modulator architecture: dorsal frontal circuits maintain a shared choice axis guiding decisions across contexts, whereas vPFC maintains competing attributes in distinct subspaces that modulate decisions when options conflict.

Cerebral cortical maturation supports emotional, social, and cognitive development during adolescence1. Neurodevelopmental theories of psychiatric illnesses, which peak in onset around age 182, implicate subtle alterations in adolescent cortical development3. However, mechanistic links between cortical maturation and emergent psychopathology have remained elusive. This gap likely reflects a need for more nuanced understanding of cortical development4, larger, longitudinal samples powered to detect small effects,5 and more uniform image quality control.6 Using 29,482 structural MRI scans from 11,720 adolescents in the Adolescent Brain Cognitive Development (ABCD) study, we demonstrate reproducible differences in the contour and tempo of maturation across the cortical surface. Regional trajectories are robust to image quality variance and align with increasing functional specialization over time. Thicker cortex in early-maturing sensory regions predicts psychiatric symptoms as early as four years prior to their emergence, a pattern that subsequently spreads to association cortex. Overlap with spatially and temporally organized gene co-expression networks implicates prenatal programming of this sequence of maturation and its links to psychopathology. These findings present a novel, versatile neurodevelopmental framework for cortical maturation during adolescence that sequentially implicates sensory and association cortex in risk for emergent psychiatric symptoms.

Animals continuously transform sensory inputs into representations that guide behavior. Although the amygdala is a brain structure critical for learning and choice, how natural sensory experience is represented within this circuit remains unknown. Here, we identify a representation of the visual environment in the human basolateral amygdala (BLA) that contains information relevant for learning and choice. Using image-computable encoding models trained on responses to naturalistic videos, we recover a continuous, low-dimensional representation of visual experience in the BLA. The representation generalizes across individuals and independent datasets and prospectively predicts BLA responses to novel visual stimuli generated through closed-loop image synthesis. Signals derived from the representation also informed reinforcement-learning models of human choice. Together, these findings show that the human BLA organizes complex visual experience into a structured representation that can support learning and flexible behavioral responses.

Dopamine in the nucleus accumbens Core powerfully influences motivation and learning, yet the information encoded by Core dopamine remains unresolved. Competing theories typically assume a spatially uniform signal, and standard recording methods average dopamine over substantial tissue volumes. Using two-photon imaging through microprisms in behaving mice, we reveal a pronounced spatial organization of Core dopamine dynamics. Posterior Core showed phasic dopamine increases to both rewards and aversive stimuli, with transfer to a predictive cue after learning. Anterior Core showed phasic decreases to aversive stimuli, weaker increases to reward, and higher tonic dopamine that was further amplified by cocaine. Rather than uniform or highly fragmented signals, the Core contains a smoothly graded mixture between two distinct sets of dopamine dynamics, neither matching canonical reward prediction errors.

Mouse ultrasonic vocalizations (USVs) provide a rapidly evolving readout of social interaction but are typically analysed only after acquisition. Here we introduce DeepFisFis, a waveform-based neural network that detects USVs while they are being produced. DeepFisFis classifies consecutive 5 ms audio segments directly from the waveform with high accuracy, processing each segment in approximately 2.5 ms and therefore faster than the incoming audio stream. This enables ongoing vocalizations to guide experimental interventions in real time. In a deployed closed-loop system, detections triggered an external stimulus, demonstrating online control of ongoing vocal behaviour. DeepFisFis also enables event-triggered data acquisition: gating storage around detected calls preserved more than 99% of vocalization time while retaining only approximately 22% of the continuous recording. DeepFisFis thus transforms USVs from a retrospective behavioural readout into a real-time experimental signal for selective acquisition and closed-loop causal interrogation of vocal communication and its underlying neural circuits.

Human brain assembloids provide a tractable platform for studying mutation-specific network dysfunction, but most studies do not distinguish whether abnormal activity arises from increased spontaneous initiation or stronger history-dependent amplification. We introduce a low-dimensional discrete-time self-exciting population model for calcium-imaging event sequences that separates baseline initiation (%[mu]), integrated history-dependent gain (%[eta]), and memory decay (%[tau]). Across 175 recordings from 13 MAPT p.R406W or CRISPR-corrected isogenic assembloids, biological-unit inference identified a 4.31-fold increase in %[eta] in mutants (95% CI 3.20- 5.84; exact p=7.8x10-4), without increased baseline activity. Parametric predictive checks exposed residual Poisson overdispersion; fixed-decay and beta-binomial sensitivity analyses nevertheless preserved the mutant history effect. Median %[eta] separated all 13 held-out assembloids, while a multifeature recording-level RBF-SVM achieved AUC 0.816 under nested whole-assembloid holdout. These results identify robust history-dependent temporal amplification as a functional MAPT-mutant network phenotype.

Brain states are conventionally defined as discrete categories, such as wakefulness, NREM sleep and REM sleep. Yet, whether these labels fully capture the organization of cortical population dynamics remains unclear. Using large-field-of-view two-photon calcium imaging of layer II/III pyramidal neurons in mouse primary visual cortex (V1), we show that population activity organizes into recurring micro-states that span conventional sleep-wake boundaries. A Hidden Markov Model fitted independently of behavioral state labels identified temporally stable micro-states with structured transition pathways. Micro-state organization was more strongly aligned with a continuous behavioral arousal axis, indexed by pupil diameter and orofacial motion, than with discrete NREM-wake categories, a relationship confirmed by mediation analysis. At the circuit level, two spatially intermingled but functionally opposing neuronal ensembles tracked this arousal axis in opposite directions. These findings establish behavioral arousal, rather than discrete wake-sleep state, as the primary organizing dimension of V1 population dynamics.

Prader-Willi syndrome (PWS) is a neurodevelopmental disorder characterized by hypothalamic dysfunction, developmental delay, hyperphagia, and cognitive impairment, yet the cellular mechanisms underlying circuit abnormalities remain unclear. Here, we show that loss of the imprinted gene Magel2 remodels microglial state and function in a mouse model of PWS. Three-dimensional morphometric analysis of hypothalamic microglia revealed reduced structural complexity and contracted territorial domains, emerging early in development and largely independent of obesity. Single-nucleus RNA sequencing of hypothalamic brain tissue identified extensive transcriptional reprogramming involving immune signaling, cellular communication, ribosome function, synaptic organization, and protein translation, with stronger effects in females. Magel2 deficiency also increased microglia-neuron contacts and synaptic engulfment, accompanied by reduced synaptic marker density, suggesting aberrant refinement of hypothalamic circuits. Analyses of human post-mortem hypothalamic tissue from individuals with PWS confirmed microglial structural alterations and increased complement engagement. Ligand-receptor inference further revealed reorganization of microglia-centered communication networks, including altered adhesion, trophic, and immune signaling consistent with enhanced phagocytic states. Together, these findings identify microglial remodeling and increased synaptic engagement as central features of Magel2 deficiency, implicate microglia in hypothalamic circuit dysfunction in PWS, and highlight excessive synaptic pruning as a novel dysfunctional mechanism of PWS.

Large-scale Neuropixels recordings sample many brain areas, but no single experimental session records every area of interest. NeuroPaint, a transformer-based autoencoder, integrates such recordings by aligning area-specific latent dynamics across sessions and "inpainting" activity in unrecorded areas. Could a linear model, which would be computationally efficient and potentially more interpretable, suffice to achieve this? To address this question, we introduce Linear NeuroPaint, a linear variant of NeuroPaint, and evaluate it on two multi-area Neuropixels datasets. Linear NeuroPaint achieves strong predictive performance on recorded areas, however, it is less accurate than nonlinear models at cross-area inpainting, despite extensive hyperparameter optimization. Moreover, its parameters are not readily interpretable: incomplete cross-session alignment leads parameter-defined neuron clusters to exhibit heterogeneous response profiles. Together, these results show that linear transformations are insufficient for cross-session alignment and cross-area prediction. Nevertheless, Linear NeuroPaint provides a computationally efficient baseline for benchmarking multi-session neural models.

The habenula (Hb) is a small epithalamic midline structure critical for reward processing and mood regulation that is implicated in many complex brain disorders, including depression and addiction. The molecular anatomy of the Hb has been well-characterized in model organisms, but little is known about its organization in the human brain, including the transcriptomic and epigenetic signatures of spatially-organized lateral (LHb) and medial (MHb) cell types. Here, we generated a high resolution spatiomolecular map of the human Hb using both multiomic single nucleus sequencing (snMultiome) and spatially-resolved transcriptomics. We measured gene expression and chromatin accessibility across 4 MHb and 5 LHb neuronal cell types as defined by transcriptomic conservation with rodents and zebrafish. We identified cell type-specific molecular specializations related to GABAergic and glutamatergic signaling as well as regulatory relationships between open chromatin regions, transcription factors, and target genes associated with depression and addiction. High resolution spatial transcriptomics identified the anatomical location of heterogeneous LHb subpopulations, including two putative inhibitory populations in the LHb. We further characterized cell-type relationships between Hb neurons and glial subtypes and predicted ligand-receptor interactions between LHb and MHb neurons. Finally, we delineated Hb cell types and spatial domains enriched in genes associated with genetic risk for psychiatric disorders, identifying the strongest enrichment in the MHb. Collectively, the generated data and analyses resolve the spatiomolecular organization of the human Hb and identify molecular-genetic associations with neuropsychiatric disease.

Neuroimaging studies have shown close overlap between neural representations of nociceptive stimuli and non-nociceptive, salient stimuli (such as low-probability sounds), leading to the proposal that nociception and salience share a common neural substrate in the human brain. Here we use high-resolution intracranial recordings from 13 neurosurgical patients (2,018 total electrodes) to show that neural representations of nociception and salience are anatomically and functionally distinct in the human brain. 77% of 107 electrodes that encoded nociceptive stimuli did not encode low-probability sounds ("nociceptive representations"), whereas 23% encoded both nociceptive and low-probability sounds ("salience representations"). Nociceptive representations were distinct from salience representations in terms of anatomical prevalence (posterior insula, inferior parietal, and perirolandic cortex vs. anterior insula, thalamus, and inferior frontal gyrus), timing (nociception preceded salience by 120 ms on a trial-by-trial basis), and function (intracranial electrical stimulation was more likely to cause sensory experience when applied near nociceptive representations, as compared to salience representations). These results provide strong evidence that nociception and salience have distinct neural substrates in the human brain, and that selective measurement of these distinct processes may help understand patient-to-patient variability in chronic pain and develop personalized brain stimulation treatments.

Modern recording technologies enable simultaneous measurement of high-dimensional neural activity and rich behavioral variables, creating both an opportunity and a modeling challenge: identifying the shared latent structure that mediates the brain-behavior relationship and the underlying dynamics through which it evolves across behavioral states. Existing approaches focus on different parts of this challenge: shared representation learning methods identify neural-behavioral subspaces but typically do not model dynamical structure, while dynamical models capture temporal variation and discrete state transitions in neural activity, but do not explicitly disentangle shared neural-behavioral structure from modality-specific variability. We introduce Switching Shared Latent Dynamics (SSLD), a unified framework that learns a latent representation shared between neural activity and behavior and endows this shared space with nonlinear switching recurrent dynamics to reflect the hypothesis that behaviorally relevant neural representations exhibit structured dynamical changes across behavioral states. Complementary private latent variables capture modality-specific variability, providing a window into neural activity that does not directly interact with behavior. We evaluate SSLD on a simulated dataset and four diverse experimental datasets spanning species, brain regions, and recording modalities: monkey motor and premotor cortex during reaching, somatosensory cortex during a bump task, widefield calcium imaging across mouse dorsal cortex during (a) self-initiated decision-making and (b) spontaneous movements. Across all four datasets, SSLD accurately reconstructs neural and behavioral signals, recovers discrete states in the dynamics that align with experimentally defined behavioral epochs, and isolates behaviorally-relevant neural information in the shared latent while preserving private neural variability. Ablations confirm that shared representation learning, behavioral supervision, and switching dynamics each contribute to performance. SSLD offers an interpretable approach to modeling shared neural-behavioral dynamics that complements ongoing efforts toward foundation models for neuroscience.

The goal of building a whole-brain interface - capable of generating a real-time digital representation of the state of the entire central nervous system - has long been framed as a hardware challenge of recording simultaneously from ever-larger numbers of microelectrodes. Advances in high-resolution neural interfaces now enable electrophysiologic interaction with the human nervous system at spatial and temporal resolutions previously unattainable without disrupting tissue architecture or physiology. These capabilities create new opportunities for mechanistic neuroscience, neuroprosthetics, and precision-guided neurosurgery. With the advent of modern microelectrode array technology, however, the whole-brain interface has also become a data challenge. Rather than pursuing large-scale human brain recording purely through a single device placed in a single patient, we demonstrate here that systematic multi-subject aggregation, across brain regions and behavioral paradigms, in a common anatomical framework, can achieve something functionally similar to a whole-brain interface. Here we present an approach to a functional atlas of the human central nervous system derived from high-resolution surface electrophysiology recordings obtained in the first 70 patients implanted with an anatomically conformal, thin-film neural interface positioned in more than 100 locations. The atlas comprises >100,000 electrode deployments and catalogs placement across a broad range of neuroanatomic targets spanning the motor, sensory, language, auditory, and associative domains, as well as brainstem and spinal cord structures that have historically been difficult to access at high electrophysiologic resolution. Across participants, aggregated recordings reveal somatotopic organization, including digit representations that follow a lateral-to-medial ordering consistent with classical homuncular maps, with transitions between adjacent representations occurring over spatial scales of approximately 0.8 - 2.0 mm; canonical frequency-band signatures, such as movement-related beta-band suppression and focal high-gamma (70 - 200 Hz) increases during motor and speech behaviors; and task-modulated population dynamics in both evoked and spontaneous activity across cortical, brainstem, and spinal cord regions. These findings yield insight into patterns of mesoscale electrophysiological dynamics conserved across the entire human central nervous system, suggesting that fine-scale functional organization is preserved across individuals when aligned to common anatomical landmarks. We further demonstrate several applications of these high-resolution recordings, including real-time neural decoding from upper extremity motor cortex to enable real-time neuroprosthetic control, and from ventrolateral cortex to decode cortical correlates of spoken language, facial movements, and complex facial expressions of emotion, supporting initial feasibility of combined verbal and nonverbal communication interfaces. Together, these results establish a paradigm for approaching large-scale interfaces with the human brain, as well as a starting point for a scalable, extensible atlas of human electrophysiologic function designed to accelerate discovery across basic neuroscience, translational research, and clinical neurotechnology.