Neuron

Adaptive choice requires transforming the evaluation of available options into commitment to a specific action but understanding how this transformation is implemented across neural circuits remains a central challenge. Here we recorded well-isolated neurons in the substantia nigra pars reticulata (SNr) and frontal eye field (FEF), which send parallel projections to the superior colliculus for driving the eye movement choice, while monkeys performed a sequential-offer choice task designed to partially dissociate scene-defined ordinal rank from the categorical commitment. Before target onset, SNr displayed stronger scene-related modulation than FEF. During target evaluation, SNr activity showed ordered modulation across behavioral outcomes dominated by ordinal rank, whereas FEF activity categorically separated acceptance from rejection and strongly encoded target direction. Behavioral model decomposition revealed that ordinal rank alone best explained SNr activity, outperforming both reward magnitude and even a composite acceptability measure that incorporated rank together with reward, scene context, and waiting cost, whereas FEF activity was best explained by categorical commitment. This dissociation was consistent across multivariable modeling, single-neuron response patterns, and all three monkeys. Together, these findings support a division of labor in which context-dependent evaluation and categorical commitment are distributed across parallel basal ganglia and frontal cortical output pathways to efficiently guide voluntary choices.

Internal states such as motivation and task engagement influence cognitive functions. Working memory, which maintains information over time, is an essential component of cognition and is modulated by motivation. Here, we show motivational states modulated attractor dynamics that supported working memory. Combining population recordings from mouse medial prefrontal cortex (mPFC) with data-constrained recurrent neural network (RNN) modeling, we found task engagement selectively modulated attractor dynamics within a memory-maintenance subspace, while stimulus-evoked responses remained intact. Reverse-engineering the RNNs revealed that task engagement reorganized the dynamical landscape by stabilizing memory-specific attractors. Specifically, task engagement modulated interactions between neurons to change the attractor dynamics. Finally, gradual changes in behavioral engagement were predicted by continuous modulation of attractor geometry in RNNs and mPFC. Together, these results suggest that internal state modulate working memory function by controlling the dynamical regime of mPFC circuits, providing a mechanistic link between internal state, neural dynamics, and cognitive function.

Planning is critical for adaptive behaviour in a changing world, because it lets us anticipate the future and adjust our actions accordingly. While prefrontal cortex is crucial for this process, it remains unknown how planning is implemented in neural circuits. Prefrontal representations were recently discovered in simpler sequence memory tasks, where different populations of neurons represent different future time points. We demonstrate that combining such representations with the ubiquitous principle of neural attractor dynamics allows circuits to solve much richer problems including planning. This is achieved by embedding the environment structure directly in synaptic connections to implement an attractor network that infers desirable futures. The resulting spacetime attractor excels at planning in challenging tasks known to depend on prefrontal cortex. Recurrent neural networks trained by gradient descent on such tasks learn a solution that precisely recapitulates the spacetime attractor - in representation, in dynamics, and in connectivity. Analyses of networks trained across different environment structures reveal a generalisation mechanism that rapidly reconfigures the world model used for planning, without the need for synaptic plasticity. The spacetime attractor is a testable mechanistic theory of planning. If true, it would provide a path towards detailed mechanistic understanding of how prefrontal cortex structures adaptive behaviour.

Cues that predict reward can acquire strong power to bias orientation and approach, directing seeking behaviors into close proximity with reward. The parafascicular thalamus (PF), part of the intralaminar complex, is well-placed as a neural hub for integrating subcortical sensory and attentional signals to generate actions that support cue-guided behavior. Neurons in the PF respond to sensory cues and encode features of head position, but little is known about how the PF is engaged in vivo during learning. Here, we recorded calcium activity in PF neurons using fiber photometry throughout a pavlovian conditioning task. We found that PF neurons developed sustained cue-evoked responses that scaled with associative learning and diminished with extinction. PF neurons preferentially signaled cue-directed orientations and body movements made during the cue, and their activity paused during reward consumption. Using optogenetics, we found that stimulation of PF neurons disrupted normal cue-directed orientation and impaired associative learning by promoting exaggerated ipsilateral turning behavior. Broadly, these results suggest an encoding profile by which the PF can support cue and reward approach behavior via dynamic regulation of head and body position. Together, our data demonstrate a thalamic region that is important for steering attention to facilitate cue-reward learning.

Humans and animals plan actions to achieve goals in worlds that are complex and continually changing. While planning is critically dependent on the prefrontal cortex in humans, little is known about its cellular underpinnings. Mechanistic understanding has been limited by a scarcity of controlled animal experiments in which subjects must flexibly plan novel behaviours on every trial. Here we characterise the neural representations and dynamics of mouse medial frontal cortex (mFC) during flexible navigation in structured environments. We trained mice to navigate complex mazes, to goals that changed location on every trial. Optogenetic silencing established that mFC was necessary for efficient navigation. mFC activity was dominated by two factorised components: (i) a structured representation of subjects position within the maze that formed an efficient code for behavioural trajectories, and (ii) a flexible representation of the shortest path-distance to the current goal. Both representations oscillated within local field potential (LFP) theta cycles, processing from further to closer to the goal at a systematic offset. These data suggest a computation in which mFC evaluates possible futures by their distance-to-goal to update a structured behavioural policy.

Escape from threat is one of the most ancient and conserved sensorimotor transformations and must satisfy two competing demands. It must be fast and reliable, because failing to escape genuine threats risks death, yet also selective, because escaping indiscriminately costs energy and missed resources. Speed and reliability favour a system with a low threshold that responds to the slightest indication of danger, whereas selectivity demands a high threshold that filters out innocuous stimuli - yet both must be achieved simultaneously. While previous work has identified mechanisms for implementing selectivity, it is not known how the mammalian brain achieves speed and reliability once genuine threats have been identified. Here we use whole-cell recordings from dorsal periaqueductal gray (dPAG) neurons in behaving mice to show that the escape circuit resolves this tension via high intrinsic excitability at the single cell level. We find that while the membrane potential of dPAG neurons does not reach action potential threshold during exploratory behaviour, they require only a small amount of current to fire. Threat stimuli cause a sparse increase in synaptic input rate that, because of the high excitability, produces a sustained depolarizing voltage step that drives spiking and escape behaviour. Individual dPAG neurons respond similarly on escape and non-escape trials - what determines escape is the fraction of the dPAG population that is recruited, a finding we confirm with single-unit recordings in freely moving mice. The membrane voltage step response is also invariant to threat type, in contrast to superior colliculus neurons where membrane potential dynamics reflect stimulus identity. These findings reveal a single neuron mechanism for fast and reliable escape, in which the high intrinsic excitability of dPAG neurons transforms sparse, threat-evoked synaptic input into a rapid and stereotyped voltage signal, converting diverse threat representations into a uniform escape command.

Flexible behavior requires using past experiences to configure cortical computations to suit current task demands. A central question in neuroscience is how memory representations control such reconfigurations. Although hippocampal (HPC) engrams can drive learned behaviors, prior studies have been largely limited to a fixed stimulus-response computation. Thus, whether engrams can drive retrieval of a set of stimulus-response mappings, rather than a specific response itself, remains unresolved. Moreover, how engrams affect cortical task representations and dynamics so as to produce engram-consistent behavior remains unstudied. Here, we address both questions by combining tagging and reactivation of HPC engrams with simultaneous large-scale recordings in medial prefrontal cortex (mPFC) during a context-dependent task-switching paradigm in mice. We report that HPC engram reactivation caused mice to apply engram-consistent decision-rules rather than a specific motor output. Simultaneous mPFC recordings revealed that reactivating the HPC engram for a given context reinstated the representation of that context in mPFC within hundreds of milliseconds, indicating that it was mediated by rapid network effects, leading to choice behavior consistent with the reactivated HPC engram. Other aspects of endogenous dynamics in mPFC were left remarkably intact. Together, our findings provide direct causal evidence that HPC engrams can configure task-relevant population states of downstream cortical circuits in real time, establishing a neural mechanism by which memory traces control flexible behavior.

Humans solve problems adaptively by selecting strategies suited to the situation. For example, when missing the bus to an appointment, we may wait for the next bus, call a taxi, cancel, or reschedule depending on the circumstances. Yet the neural and computational principles that support such flexible problem solving remain poorly understood. To address this question, we designed a moderately complex decision task for monkeys that allows multiple plausible solution strategies. Animals learned the task rapidly, generalized to novel maze geometries, and their choices were inconsistent with any single fixed strategy. We then recorded large-scale neural activity from the frontal cortex and found that population dynamics varied systematically with maze geometry. Neural responses clustered into two distinct dynamical regimes with separable initial states, consistent with hierarchical and sequential strategies. A decoder trained on population activity revealed time-resolved decision dynamics that aligned with these regimes, and an unsupervised latent-space analysis provided convergent evidence that strategy use varied across trials. A behavioral model grounded in neurally inferred strategies accounted for choices better than fixed-strategy alternatives and captured trial-by-trial variability. Together, these results provide a neural and computational account of how the brain selects and implements distinct strategies during adaptive problem solving.

Learning operates across different brain circuits to associate population activity patterns with desired outcomes, and to enable volitional reactivation of those patterns to control behavior. These circuits differ profoundly in their architecture and dynamical regimes, yet which features of learning are shared across them and which arise from circuit-specific implementations remains unknown. Here, we use a brain-computer interface (BCI) to train mice to modulate the activity of selected neuronal ensembles toward configurations that trigger reward delivery. By making reward delivery contingent directly on population activity, we impose an identical associative learning problem on two circuits with distinct dynamical regimes: the primary motor cortex (M1) and the hippocampal area CA3. Mice acquired robust volitional control in both regions, and learning produced a set of shared signatures across circuits, including modulation of reward-controlling neurons, network-level sparsification, and greater exploration of reward-related activity patterns. These signatures were underpinned by distinct population dynamics: M1 activity flowed continuously through reward-associated states, whereas CA3 activity traced approach-and-return dynamics around them. Recurrent network models endowed with distinct minimal connectivity constraints chosen to reflect the dominant dynamical regime associated with each region captured key features of these shared signatures and region-specific dynamics, indicating that local architectural constraints are sufficient to account for the distinct implementations of learning. These findings indicate that equivalent learning outcomes arise from divergent dynamical implementations across architecturally distinct circuits. This principled degeneracy reveals that learning is not a single canonical solution, but is implemented through multiple circuit-specific mechanisms shaped by local network architecture.

Adaptive behavior requires animals to use the outcomes of recent actions (action-outcomes) to guide future decisions. While the striatum is critical for decision-making, it is currently unclear how it is involved in the storage and retrieval of short-term associative memories. Here, we implemented a head-fixed memory-guided decision-making task in which mice use the outcome of a previous choice to determine whether to repeat or switch their next action. We show that dopamine fluctuations in the ventrolateral striatum are modulated by reward receipt or omission and recent outcome history, while the activity of direct and indirect pathway striatal projection neurons encodes recent action-outcome associations and predicts future switch/repeat choices. Closed-loop optogenetic activation and inhibition of direct and indirect pathway neurons during either the action-outcome association period or the delay preceding the next choice bidirectionally biased future actions away from those favored by reward history. Together, these findings suggest that striatal activity maintains a short-term action-outcome associative memory that links completed actions to future motor plans during adaptive decision-making.

Human genetics implicates innate immunity as a key modifier of tau toxicity, yet human-specific neuroimmune mechanisms remain difficult to test in vivo. Here, we developed HuMiNAX, the first humanized iPSC-based neuroimmune xenograft model of tau-associated neurodegeneration, enabling human microglia to interact with human neurons and astrocytes in the adult mouse brain. In HuMiNAX, tau seeding induced aggregation only in mutation-carrying human neural grafts, causing neuron loss and inflammatory activation of human microglia. Progranulin-overexpressing human microglia dampened tau-associated inflammation, preserved neurons, and restored neuronal gene-expression and RNA-splicing programs, supporting microglial control of neuronal resilience. CRISPRi knockdown of the human-specific lncRNA HNRNPK-AS1 also protected neurons in HuMiNAX. These findings establish HuMiNAX as a human neuroimmune model of tauopathy and identify microglial and RNA-mediated strategies of neuronal resilience.

Large language models have the potential to transform scientific research and analysis, but without domain-specific structure they produce silent methodological errors, unreported decisions, and irreproducible results. Here we present SpikeLab, a text-to-analysis framework for neural spike data that combines composable data structures with a skill-based agentic system enforcing bounded autonomy: mandatory use of expert-vetted methods, correctness over efficiency, and clarification-seeking on ambiguous requests. In a controlled benchmark on electrophysiology data, Sonnet 4.6 with SpikeLab produced correct and reproducible results across all tasks, outperforming both the unassisted Sonnet and the more capable Opus 4.6, which exhibited deterministic failures including ad hoc method invention, silent data reduction, and inconsistent experimental designs. We demonstrate versatility across in vivo mouse, human, and in vitro brain organoid recordings, and apply the framework to a pharmacological dose-response study spanning single-unit dynamics, pairwise network structure, burst-level temporal sequences, and latent population states, all through natural language prompts without writing analysis code.

Strong prior beliefs can render individuals resistant to change, even when outcomes contradict expectations, yet the neural mechanisms supporting such stability remain poorly understood. Here we show that the ventral hippocampus (vCA1) actively stabilizes inferred contextual state (belief) during aversive memory extinction in a behavior-dependent manner. Using in vivo calcium imaging, we identify a subset of vCA1 neurons that exhibit selective reductions in activity when expected aversive outcomes do not occur during extinction. Axonal calcium imaging further reveals that reduced signals are broadcast to basal amygdala, nucleus accumbens, and medial prefrontal cortex. Optogenetic manipulations demonstrate that these signals actively suppress extinction learning and promote persistence of defensive response. Computational modeling indicates that vCA1 omission-related signals reflect discrepancies between beliefs and observed omission of the expected outcome, thereby maintaining the aversive belief. Crucially, convergent evidence from calcium imaging, modelling, and closed-loop optogenetic manipulations reveals that vCA1 omission-related signals are specific to the period during which animals execute defensive behavior driven by their inferred aversive state. These findings identify a novel mechanism in vCA1 that stabilizes inferred aversive state despite changes in stimulus contingencies and suggest a key role of vCA1 in belief-based control of learning.

Frontal cortex plays critical roles in action control and motor skill learning. Within the layer 1 apical tuft dendrites of layer 5 (L5) neurons in frontal cortex, precise input patterns and back-propagating action potentials can trigger powerful regenerative events that may be essential for flexible computation and learning. However, it remains unclear whether tuft activity in frontal cortical L5 circuits encodes sensorimotor information that differs from the information conveyed by their outputs to downstream targets. Using longitudinal two-photon calcium imaging, we investigated sensorimotor encoding in the apical tuft dendrites and somata of L5 extratelencephalic (ET) neurons in the frontal cortex during learning of a discrete change to a cued dexterous action. During learning, movement errors either triggered corrective action or did not, allowing us to dissociate error signals from signals selective for corrective action. Somatic activity tracked both sensory cues and action, whereas tuft activity predominantly tracked sensory cues. Movement errors during learning revealed additional distinct tuft activity that was selectively associated with corrective actions. Furthermore, learning induced divergent changes in the response gain and net selectivity of tuft dendrites compared to somata. Our measurements uncover systematic differences between the tuft dendrites and somata in sensorimotor selectivity, sensitivity to corrective action, and functional plasticity, providing a foundation for investigating the contributions of dendritic computation to motor skill learning.

Animals must execute learned behaviors and update them when outcomes change, yet the neural substrates controlling this phenomenon are not fully understood. Here, we show that N-Methyl-D-Aspartate Receptors (NMDARs) in the dorsomedial striatum are necessary for learning from previously rewarded actions. Moreover, blocking of striatal NMDARs almost fully abolished striatal calcium dynamics, but not action potential activity, suggesting a unique function of NMDAR-driven striatal calcium activity in updating action policy.

Dopamine is essential for basal ganglia function. Striatal dopamine release can be triggered by dopamine cell firing, but also by coordinated cholinergic interneuron activity, which stimulates dopamine release via presynaptic nicotinic acetylcholine receptors on dopamine axons. While acetylcholine-dependent dopamine release is well-documented ex vivo and under artificial optogenetic stimulation in vivo, its role during natural behavior has remained unclear. One possible natural driver of acetylcholine-dependent dopamine release is thalamic input, which provides strong excitatory drive to cholinergic interneurons. To examine whether thalamic input provokes acetylcholine-dependent dopamine release during behavior, we performed simultaneous fiber photometry recordings of striatal dopamine (GRAB-rDA3m) and thalamic axon activity (gCaMP8m) in the dorsomedial (DMS) and dorsolateral striatum (DLS) of mice learning the accelerating rotarod, a striatal-dependent task that demands precise and effortful motor control. Recordings were obtained on- and off-task and across days of training to capture the full arc of learning. Dopamine transients in DMS, but not DLS, were frequently coupled to peaks in thalamic axon activity via an acetylcholine-dependent mechanism. The occurrence of these thalamic-evoked dopamine transients depended on learning, task engagement, and the recent history of striatal dopamine activity, but did not appear to signal motor errors. Together, these findings establish thalamic input as a physiological driver of acetylcholine-dependent dopamine release. Moreover, they reveal that striatal sensitivity to this local release mechanism is dynamically gated by dopaminergic history, providing a compelling framework for understanding how local and soma-triggered dopamine signals are coordinated to support learning.

Flexible decisions require the brain to transform sensory evidence into abstract, task-relevant variables and then into actions. Understanding this process requires identifying how distributed neural populations represent sensory, cognitive, and motor variables, and how interareal interactions mediate transformations between them. We simultaneously recorded population activity in frontal eye field (FEF), lateral intraparietal area (LIP) and superior colliculus (SC) while monkeys performed a flexible yet urgent visual motion-categorization task. Within this network, FEF first encoded abstract categories, followed by SC and then LIP. LIP showed the earliest encoding of visual stimulus features, but a later encoding of upcoming saccades. Single-trial analyses revealed directed information flow from FEF to LIP populations for category- and choice-related signals. Reversible FEF inactivation impaired categorization and saccadic choice, causally implicating FEF in category-guided action. These findings reveal a differentiated FEF-LIP-SC circuit for transforming sensory evidence into abstract categorical decisions and the actions used to report them.

The hippocampus uses external stimuli and self-motion to construct cognitive maps critical for navigation. However, these maps can also be activated independently of external inputs and movements, reflecting an internal ability to control the map. The neural basis of such internal control remains unknown. To address this, we used a brain-machine interface (BMI) to drive navigation in mice directly from real-time hippocampal activity. In this condition, CA1 place codes rapidly reconfigured to disregard locomotion-related input. By comparing BMI-controlled navigation, locomotion-controlled navigation, and passive playback of predetermined routes, we found evidence of CA1 place cell responses that emerge specifically in conditions in which animals can causally influence their travel. Our findings thus indicate that agency is represented by a distinct place cell code.

Fear extinction suppresses learned fear without erasing the original memory, yet how competing fear and extinction ensembles are selectively updated remains unknown. Here, we show that microglia preserve fear memory by selectively editing extinction-ensemble synapses during retrieval. In medial prefrontal cortex, microglial processes, but not somata, expand, ramify, and tighten their engagement with extinction-ensemble dendrites and spines, where they preferentially engulf excitatory postsynaptic material and bias spine remodeling toward elimination. This selectivity is instructed by local find-me, eat-me, and dont-eat-me cues: purinergic signaling recruits microglial processes, phosphatidylserine exposure licenses engulfment of extinction-ensemble synapses, and CD47-SIRP protects fear ensembles from removal. Weakening microglial recruitment or engulfment, or removing this protection, accelerates extinction without impairing fear acquisition. These findings identify microglial processes as active gatekeepers of ensemble competition and reveal a neuroimmune mechanism that preserves fear memory by limiting extinction. In BriefDuring memory updating, microglial processes selectively engage extinction-ensemble synapses rather than globally pruning active circuits. Local recruitment, engulfment, and protection cues determine this choice, allowing microglia to preserve fear memory by limiting extinction. HighlightsO_LIExtinction retrieval selectively recruits microglial processes to extinction ensembles. C_LIO_LIMicroglia preferentially engulf excitatory postsynaptic material from extinction ensembles. C_LIO_LILocal "find-me," "eat-me," and "dont-eat-me" cues determine synapse selection. C_LIO_LIDisrupting microglial pruning facilitates extinction without impairing fear learning. C_LI

Accurate navigation requires egocentric representations of environmental geometry to be continuously updated by self-motion while remaining stable over time. The retrosplenial cortex (RSC) is central to this process, yet how local inhibitory circuits support this balance remains unclear. We show that parvalbumin (PV) and somatostatin (SST) interneurons regulate distinct components of egocentric spatial coding in RSC. PV interneurons are strongly modulated by self-motion and exhibit bearing-aligned synchrony that precedes SST activation, linking movement to egocentric coding precision. In contrast, SST interneurons display weak self-motion modulation but robust boundary-anchored activity with globally coherent dynamics that stabilize representations over time. Optogenetic silencing revealed dissociable effects: PV perturbation degraded egocentric coding precision while SST perturbation disrupted global population organization. Behaviorally, PV silencing impaired initial egocentric orientation while SST silencing preserved initial orientation but impaired its sustained update. These findings identify separable inhibitory mechanisms balancing rapid updating with representational stability during navigation.