Nature Neuroscience

Chronic pain mechanisms are complex, spanning multiple brain regions and networks. We ask whether resting brain activity carries a readout of that state. From a few minutes of resting-state electroencephalography (EEG), we generate a spectrogram to represent how each region of the cortex oscillates across frequency and time and pass it through CREST (Cortical Resting-state EEG Spatial Transformer): a frozen image-recognition network that reads each region as an image--here, a spectrogram--paired with a graph model that weighs the 56 cortical regions together to classify chronic-pain status. Across 125 people (74 with chronic pain, 51 healthy controls), evaluated through a leave-one-subject-out cross-validation, CREST separates the two groups with an area under the receiver operating characteristic curve (AUROC) = 0.782 (permutation p < 0.005). Control experiments implicate each persons individual alpha rhythm. Clinical relevanceA resting-state EEG readout of chronic MSK pain could clarify pathophysiology and inform treatment.

How sensory cortical activity is read out by downstream circuits to guide behavior is a fundamental unsolved problem. Substantial work has examined correlations between sensory neural responses and animals perceptual judgments, but their interpretations remained controversial due to many intervening variables, such as responses of unrecorded neurons. Furthermore, stimulus and choice encoding in sensory populations are often not well aligned, and it remains contested whether this misalignment indicates a limitation in sensory readout. Here, we introduce a closed-loop, neurofeedback paradigm that directly interrogates the capacity of sensory readout: the key idea is to train subjects to report specific patterns of population activity in a sensory area recorded online, rather than the actual stimuli presented. To test this, we trained macaque monkeys on a visual change-detection task using shape stimuli, implanted an electrode array in visual area V4, and tested whether they could be further trained to rely on their own V4 activity along specific axes in neural state space. Strikingly, monkeys successfully increased the neuron-choice correlation along trained axes in neural population state space. No detectable changes in stimulus selectivity or noise correlations were found within the recorded population, and further model simulations confirmed that adjustment of sensory readouts best accounted for the results. A control experiment that merely disrupted the stimulus-reward contingency without a closed loop failed to enhance neuron-choice correlation. Together, these results demonstrate that closed-loop neural feedback achieves neuron-choice alignment beyond the ceiling of natural perceptual training, suggesting that the misalignment in perceptual tasks reflects constraints on learning within naturally available training regimes.

Adaptive behavior requires external input to be interpreted together with internal state and ongoing behavior. During pain, noxious somatosensory input evokes movement and arousal, and injury reshapes this relationship, yet how cortical activity organizes stimulus content with behavioral state remains unclear. In awake mice, we delivered hindpaw stimuli while tracking movement, arousal and facial expression, and studied primary somatosensory cortex (S1) using widefield and two-photon calcium imaging, single-action-potential activation of nociceptors, and S1 silencing. Here, we show that S1 neurons were broadly recruited by stimulus and state, whereas latent population dimensions carried mechanical stimulus content. Inflammatory injury caused a reorganization of S1 geometry, binding state and protective responses tighter together. Noxious heat drove S1 as strongly, but engaged mostly the state axis. S1 silencing reduced mechanical hypersensitivity, arousal, and facial expressions. S1 thus embeds mechanical input within a stimulus-state geometry, which is reorganized during inflammatory injury to support adaptation of a coordinated protective response.

Neural recordings from different individuals vary substantially even when behavior is broadly shared. The sampled neurons differ, and the same behavior occurs at different times. Standard cross-subject analyses rely on matched time points or anatomical correspondence, which excludes many datasets. We recently introduced Shared Representation Discovery (ShaReD), which identifies neural-behavioral relationships conserved across subjects by learning a shared behavioral projection together with subject-specific neural projections. Here we develop and benchmark this method using synthetic, primate, and rat data. On synthetic data, ShaReD recovers common structure across noise levels, sample sizes, and subject counts, and separates components confined to different groups of subjects. In non-human primate motor cortex, ShaReD identifies kinematic representations that generalize across individuals and across reaching tasks with different movement statistics. In rats navigating a spatial alternation task, ShaReD isolates behavior-aligned directions within the CA1-to-PFC communication subspace. ShaReD thus extends multi-subject analysis to datasets in which comparable behaviors occur without cross-subject temporal correspondence.

Leading deep neural network encoding models predict visual cortical responses with nearly indistinguishable accuracy, raising the strong inference that these models have converged on the same underlying brain-aligned parameterization of natural image space. Here we demonstrate that this is not the case. We introduce axis-aligned feature accentuation, which converts each model's fitted encoding axis into graded stimulus perturbations that are predicted to parametrically control neural firing within and beyond the natural-image range. We generated over 27,500 controller stimuli from ten leading vision models and presented them to five macaques in closed-loop experiments targeting early, mid-, and high-level visual areas. Despite matched natural image predictivity, models diverged strongly in their ability to control neural firing using accentuated stimuli, revealing that most model encoding axes failed to capture the precise tuning of their corresponding neurons. The two adversarially trained models showed a consistent advantage, though adversarial robustness was only weakly predictive of neural control across other models. Instead, control was better predicted by the spatial frequency structure of the input gradient: the distribution of pixels influencing each encoding axis. Overall, these results establish neural control via axis-aligned feature accentuation as a causal method to assess the alignment between how neurons and models parameterize the visual world.

Animals must balance persistence with flexibility when outcomes change. Behavioral extinction--the reduction in responding to cues that no longer predict reward--is widely thought to be driven by pauses in dopamine-neuron firing upon omission of expected rewards. However, whether these pauses are necessary remains unknown. We used a connectomic intervention to weaken inhibitory synapses onto dopamine neurons in the ventral tegmental area, attenuating electrophysiologically defined pauses while sparing tonic and burst firing. Contrary to canonical accounts, this intervention accelerated rather than delayed behavioral extinction, demonstrating that these synapses normally postpone behavioral extinction. Learning about a newly rewarded cue was spared, arguing against a general disruption of learning. Dopamine photometry showed that the intervention attenuated reward-omission dopamine dips, whose dissipation preceded and predicted behavioral extinction across mice. These findings show that pause-generating inhibitory inputs to dopamine neurons sustain behavioral persistence. This mechanism may protect established associations from premature abandonment as outcomes fluctuate.

Schizophrenia presents with several partially independent symptom dimensions, including positive symptoms, negative symptoms, and cognitive impairment; yet no neuroimaging framework has provided individual-level markers of symptom severity that remain anatomically interpretable. Here, we present an interpretable AI-based framework that addresses this gap by mapping high-dimensional resting-state rs-fMRI dynamics onto a low-dimensional latent manifold using self-supervised contrastive learning with a new attribution method to localize the highest decodable regions. Applied to two independent schizophrenia-spectrum cohorts, the label-free latent space supports individual-level prediction across clinical features of the disorder, including symptom severity and cognitive function. The attribution maps identify a disease-specific pathological footprint concentrated in prefrontal, posterior cerebellar and temporal areas that diverge from the manifold organization observed in healthy controls, which was dominated by auditory, limbic, and ventral-striatal circuits. These results establish an interpretable latent space framework for characterizing the distributed neural substrates of schizophrenia symptoms at the level of the individual patient, and provide an anatomically grounded route toward precision decoding of symptom severity.

Humans retain memories of individual experiences for a lifetime, an ability attributed to a complementary learning system in which a fast process encodes episodes and a slow process integrates them into semantic knowledge. In classical Hebbian models such as Hopfield networks, memory traces are superposed in shared weights. This makes learning naturally continual but causes strong interference among correlated memories, a failure that reappears as catastrophic forgetting in deep networks. Here we use a large language model as a model system for continual episodic memory, with its pretrained weights supplying the semantic context in which new episodes are embedded. Fast learning is implemented by a hippocampus-like module that assigns each episode to a dedicated, extremely sparse low-rank adapter; competitive gating then selects among these separated traces during recall. Across streams of up to 1,000 factual and autobiographical episodes, each adapter requires only 2-3 parameters per token while preserving excellent recall. An internal retrieval-augmented generation mechanism reconstructs the selected episode in context and supports high-accuracy question answering over stored memories. Finally, slow cortical consolidation is modeled by fine-tuning the base weights through batch replay, enabling reconstruction and direct question answering without episodic adapters. Together, fast storage and slow consolidation implement both components of a complementary learning system within a single language model, yielding a neural-network model that stores, recalls, and consolidates naturalistic episodic memories, thereby capturing key functional features of human memory.

Inflammatory activation and lipid remodeling are linked features of microglial states, but how inflammatory transcription factors shape microglial lipid handling is unclear. Here we show that STAT1 sets neutral-lipid content in microglia through a route not predicted by lipid-handling transcription. Acute STAT1 depletion in primary microglia lowered neutral-lipid content while lipid-uptake and lipid-storage programs were induced, and interferon-{gamma} activation moved inflammatory transcription in the opposite direction yet lowered lipid content alike. Single-cell transcriptomic and chromatin profiling of Stat1- and Irf1-deficient mice showed that STAT1 and IRF1 organize overlapping inflammatory and lipid-handling programs, with genome-wide accessibility changes that did not predict transcriptional output at individual lipid-handling loci. Microglia co-expressing STAT1 and APOE recurred across Alzheimer's disease and multiple sclerosis datasets. Transcriptional program engagement is therefore separable from cellular lipid state, and lipid-handling gene expression cannot be read as a proxy for microglial lipid content.

Sensory neuroscientists seek to model the neural computations that encode complex stimuli. Distinct encoding models often make similar predictions for natural stimuli such as speech, posing a challenge for model comparison. We developed a method to synthesize stimuli that decorrelate model predictions across a neural population, termed neural prediction decorrelation (NPD). Using fMRI responses to NPD sounds, we compared standard and adversarially robust deep neural network models of human auditory cortex. Prediction accuracy for NPD sounds was substantially better for the adversarially robust model in every region tested, an effect completely masked with natural sounds. Population responses to natural and synthesized NPD sounds shared an interpretable low-dimensional organization that was reproduced by the robust encoding model. NPD provides a general approach for comparing encoding models and reveals that adversarial robustness expands the predictive power of DNNs beyond natural stimuli, which is likely critical for targeting population activity through stimulus synthesis.

Gathering information before committing to a choice is critical in real-world decision making, and biases thereof are hallmark features of psychiatric disorders. Here, we study the behavioural and neural mechanisms that guide information gathering and characterise several key cognitive constituents, including escalating urgency and systematically biased temporal weighting of information. Using fMRI, we identify an integrated information gathering signal in ventromedial and dorsomedial prefrontal cortices (dmPFC), signalling an overall likelihood for continued sampling of information. Teasing this signal apart, we find distinct neural circuits encoding separable information-gathering constituents: whilst an urgency signal primarily engaged locus coeruleus and dmPFC, accumulated evidence was represented in anterio-medial PFC, and evidence-strength prediction errors were computed in ventral striatum and dmPFC. These findings indicate that information gathering arises from functionally distinguishable prefrontal-subcortical computations that converge within medial prefrontal cortex, providing a mechanistic framework for understanding aberrant sampling in psychiatric conditions, including schizophrenia and obsessive-compulsive disorder.

The amygdala is a central hub for emotional learning that is dysregulated across numerous psychiatric disorders, yet its molecular architecture in humans remains poorly defined. We here present a spatial transcriptomic atlas of the human amygdala from nine neurotypical donors. Integrating Visium, Xenium, and VisiumHD technologies, we profiled over 1.1 million cells across 13 spatial domains spanning the basolateral complex, central, medial, and cortical nuclei, as well as the intercalated islands, each defined by distinct marker genes and gene co-expression networks. By integrating snRNA-seq reference atlases, we found that each subdivision of the basolateral subnucleus contains distinct excitatory neuron classes. We additionally found three transcriptionally distinct populations of intercalated neuron types, two of which form spatially distinct islands neighboring the basolateral complex, and were able to refine cell type diversity across the central nucleus and related amygdalostriatal transition areas. Finally, we localized psychiatric genetic risk across the amygdala, revealing both broad neuronal enrichment and subnuclear specificity. Together, this atlas provides a foundational resource and framework for accelerating cross-species comparisons and human disease-focused investigations.

Parabrachial Calca neurons are necessary for chronic pain and sufficient to drive nociplastic pain in mice, but how they sustain a pain state that outlasts its trigger is unknown. We performed temporal single-cell mRNA sequencing of the parabrachial nucleus (PBN) across the onset, chronic, and recovery phases of Calca neuron-driven tactile allodynia, using fixed-tissue profiling and reference-atlas registration to track molecularly defined populations over time. Activation broadly induced immediate-early genes, after which Calca neurons displayed changes in expression of genes that affect signaling and synaptic plasticity, with bidirectional changes that mirrored the onset and resolution of allodynia. The gene encoding brain-derived neurotrophic factor (Bdnf) remained persistently elevated in the chronic phase. BDNF infusion in the PBN prolonged allodynia, whereas blockade of its receptor (TrkB) attenuated it, and inactivating the Bdnf gene in Calca neurons abolished their hyperexcitability, attenuated allodynia, and relieved pain in a migraine model. These observations pinpoint BDNF as a driver of persistent nociplastic pain.

Injury, stroke, and neurological diseases cause persistent accumulation of cellular debris that deteriorates lesion microenvironment and impedes central nervous system (CNS) repair. Debris clearance in the injured CNS has long been attributed primarily to microglia and infiltrating macrophages. Here, we identify perivascular cells as previously unrecognized phagocytes that expand after injury and exhibit robust phagocytic activity. Perivascular cell phagocytosis is conserved across multiple mouse models of CNS injury and human stroke lesions. These cells exhibit key hallmarks of phagocytosis, including LC3-associated phagocytosis for efficient lysosomal degradation. Mechanistically, phosphatidylserine serves as the eat-me signal and Axl mediates myelin debris uptake. Myelin phagocytosis drives perivascular cell proliferation, fibrosis and lesion progression. Genetic deletion of Axl in perivascular cells or pharmacological inhibition with the FDA-approved Axl inhibitor Gilteritinib reduces pathology and improves functional recovery after spinal cord injury. Together, these findings establish Axl-dependent perivascular cell phagocytosis as a therapeutic target for CNS repair.

Sleep architecture is organized by neural ensembles operating across multiple timescales, yet the organizing principles remain unclear. By monitoring neural populations across >40 brain regions in mice, we reveal a distributed sleep code spanning multiple temporal scales. On a slow (minutes-to-seconds) timescale, vigilance state reorganized brain-wide firing and was decodable from every region, with REM sleep emerging as a globally activated state. State transitions followed low-dimensional trajectories, with cortical and subcortical ensembles evolving in antiphase at wake-NREM boundaries but in register during transitions into REM sleep. On a fast (seconds-to-milliseconds) timescale, NREM slow/delta oscillations served as a global rhythm while hierarchically nesting spindles, sharp-wave ripples, and pontine (P) waves, whereas REM theta-P-wave coupling coordinated firing across regions. Regional sleep-related activities covaried with neuromodulatory innervation, while infraslow, history-dependent firing dynamics tracked NREM-REM cycles. Together, these findings provide a cell-resolved atlas of distributed neural population dynamics across sleep stages.

Population-level measurements portray object representations in primate inferotemporal cortex (IT) as smooth, low-dimensional, and predictable by deep neural networks (DNNs). However, it remains unclear whether this structured population-level picture is representative of the full diversity of its constituent neurons. Here, we used Neuropixels 2.0 probes to record large populations of well-isolated units from an fMRI-localized face patch in macaque anterior IT while monkeys viewed more than 3,000 natural images. Single-unit responses were substantially sparser and more heterogeneous than multi-unit activity (MUA). Sparse neurons collectively formed a higher-dimensional code, supported efficient image identification, and responded later than broadly tuned neurons. Partly independently of sparseness, some neurons exhibited reliable feature randomness: their stimulus preferences were reproducible across repeated presentations but discontinuous across DNN feature spaces, rather than reflecting trial-to-trial response variability or noise. These neurons were virtually uncorrelated with the surrounding population despite being located within the same face patch. By contrast, MUA responses were denser, more correlated, lower-dimensional, and better predicted by DNNs. Together, these findings suggest a dual coding architecture in anterior IT: shared low-dimensional structure supports category generalization, whereas sparse and reliably feature-random single-neuron responses expand the representational space and support efficient identification of individual visual inputs. Response sparseness and reliable feature randomness may therefore constitute fundamental computational resources for object recognition.

Decision making is associated with frontal brain circuits and spatial navigation with the hippocampus. In addition, recent work in spatial decision making tasks found single neurons in both areas encoding space conjunctively with other task-relevant variables. However, circuit function is not determined by tuning alone, but also by representational geometry, i.e. the representation of task-relevant variables in neural state space. Here, using Neuropixel recordings in a complex spatial decision making task combined with nonlinear dimensionality reduction, we show an intrinsically low-dimensional neural manifold in medial prefrontal cortex (mPFC) on which key task variables were represented as smooth gradients. This geometry resembled the hippocampal (HPC) map. The mPFC and HPC manifolds from one mouse can predict the behavior across other mice and brain areas. A non-linear representational map between the mPFC and HPC manifolds demonstrates alignment in time. Our work suggests that the representational geometry in HPC and mPFC is distributed and time-aligned using low-dimensional neural codes.

Understanding which genes are reproducibly dysregulated in which cell types is foundational knowledge for efforts to slow or reverse pathologies. For neuropathologies, such efforts rely primarily on differential expression analysis of single-nucleus RNA-seq (snRNA-seq) data. However, this strategy suffers from experimental and statistical challenges that limit marker gene reproducibility. We describe a novel strategy called Covariation Projection Analysis (CoPA) that combines the power of bulk sampling with the precision of single-cell methods. By projecting bulk gene coexpression modules onto pseudobulked snRNA-seq cell types, CoPA reveals the cellular origins of highly reproducible genomic programs and their relative importance among cell types. By comparing CoPA projection patterns between normal and pathological human brain samples using differential CoPA (dCoPA), we identify gene coexpression modules that are uniformly and reproducibly dysregulated in specific neocortical cell types in Alzheimers disease or schizophrenia. We share our findings through a novel web application called CoPA Cabana (https://oldhamlab.shinyapps.io/copacabana/).

B cell-derived IgG in the dorsal root ganglia (DRG) drives neuropathic pain after peripheral nerve injury (PNI), but the site of B cell organization is unclear. Here, PNI induced leukocyte clusters in the DRG meninges, enveloped by lymphatic endothelium and apposed to high endothelial venules. These clusters resemble tertiary lymphoid structures (TLSs) with germinal center-like features, including germinal center B cells and plasma cells, and follicular dendritic and follicular helper T cells. Single-cell RNA sequencing revealed enrichment of germinal center B cells in the DRG meninges after PNI. Germinal center B cells regulate TLS organization: TLSs were absent after deletion of Ezh2 from germinal center-experienced B cells. Intrathecal CD20 monoclonal antibody to locally deplete B cells also disrupted TLS organization. Conversely, intrathecal B cell transfer to B cell-deficient (muMT) mice was sufficient for TLS organization after PNI. Allodynia did not develop when TLS organization was disordered. Similar TLSs formed in pig DRG after tail docking and in human donors with chronic pain, where B cell receptor clonotype analysis confirmed functional maturity. Together, these data establish that germinal center B cells are required for TLS organization, and that disrupting this process abolishes the development of neuropathic pain after PNI.

Abstract Objective. Deep Brain Stimulation (DBS) of the ventral capsule/ventral striatum (VCVS) offers therapeutic potential for refractory psychiatric conditions, but clinical success is hindered by time-consuming, trial-and-error parameter programming reliant on subjective self-reports. Tracking objective behavioral markers allows contact settings to be evaluated rapidly. Here, we evaluate a direct closed-loop Multi-Armed Bandit (MAB) optimization framework designed to rapidly identify optimal stimulation contacts using raw reaction time (RT) during a cognitive control task. Approach. We leveraged empirical data demonstrating that VCVS DBS enhances cognitive control during the Multi-Source Interference Task (MSIT) in a site-specific manner. Using a synthetic patient simulation environment across 1,000 replicates, we benchmarked adaptive MAB algorithms under noisy, non-stationary conditions. Crucially, we eliminated intermediate state-space sensor models to evaluate raw RT directly, transitioned from discrete daily resets to an uninterrupted continuous optimization architecture, and implemented a rolling arena mechanism to scale contact selection under real-world hardware constraints. Main results. Eliminating the intermediate sensor model prevented high-frequency noise amplification (where state variance was inflated by 51.7% in baseline and 148.0% in conflict states) and reduced contact ranking failure rates from 31.9% down to 11.9%. Operating within a continuous trial architecture preserved historical sample density, driving mean trial-level regret down steadily over 4,200 trials and enabling dynamic re-convergence across unannounced mid-session change-points. Additionally, a 4-contact sub-arena successfully scaled search efficiency across 8-contact arrays without sacrificing selection accuracy. Significance. Direct MAB optimization within a continuous rolling arena provides a noise-resilient, hardware-compatible architecture for automated DBS programming. By bypassing latent state estimation and utilizing standard task-based behavioral metrics without specialized recording hardware or complex state-space modeling, this framework reduces search timelines to clinically feasible durations, establishing a scalable foundation for real-time, patient-tailored neuromodulation.