Rui M. Costa

Reinforcement learning requires the selective modification of synapses that contribute to outcome-relevant neural activity, yet the principles underlying these selective modifications remain unclear. Here we combined longitudinal two-photon imaging of identified dendritic spines with a closed-loop optical brain–machine interface (BMI), in which activity from selected apical dendrites of layer 5 motor-cortical neurons controlled an auditory cursor, with target acquisition triggering optogenetic reinforcement. Mice learned to reliably generate the specific dendritic activity that leads to reinforcement. Learning was accompanied by selective strengthening and weakening of individual spines on reinforced dendrites. At individual synapses, the direction of plasticity depended on both the timing of activity relative to reinforced dendritic events and the pre-existing temporal organization of that activity. Furthermore, plasticity was enriched in small task-relevant synaptic groups: spines in synergistic groups were preferentially strengthened, whereas those in antagonistic groups were preferentially weakened. These changes were accompanied by the emergence of increasingly reproducible, selectively recruited, and temporally focused synaptic ensembles, and the incorporation of newly formed spines into these ensembles. Finally, blocking striatal dopamine receptor signaling impaired BMI learning and cortical synaptic reorganization of target dendrites, demonstrating a role for striatal reinforcement in cortical plasticity. Together, these findings reveal how reinforcement of specific dendritic activity sculpts synaptic ensembles that support learned operant dendritic control.

Action selection using predictive models of the environment plays a fundamental role in human and animal behaviour, yet is poorly understood at circuit and algorithmic levels. Spatial navigation is an attractive domain for characterising how world models guide action selection. However spatial behaviours are shaped by multiple control systems including habits, vector-navigation using a Euclidean model of spatial relationships, and route planning using models of environment structure. Understanding how world models support navigation requires assays that dissociate control systems and decorrelate behavioural variables, while generating large datasets that allow precise quantification of brain-behaviour relationships. Here we developed and computationally optimised a behavioural assay to quantify flexible navigation using knowledge of environment structure. Mice navigated to visually cued goals in complex mazes, with randomised start and goal locations on each trial, generating thousands of non-repetitive goal-directed navigation trajectories. They navigated efficiently - strongly favouring options on the shortest path-to-goal, and learnt rapidly - demonstrating knowledge of maze structure from their first sessions in new environments. We anticipate the assay will be useful for characterising how world models support flexible behaviour.

The external globus pallidus (GPe) is traditionally viewed as a homogeneous relay in the indirect basal ganglia pathway that broadly suppresses movement. Using whole-brain anterograde axon mapping, rabies tracing, single-neuron reconstruction, and single-nucleus RNA sequencing, we reveal that the GPe is a major basal ganglia output nucleus, composed of anatomically and molecularly distinct populations. Besides the neurons with canonical projections, and projections to striatum and cortex, the GPe contains specific neuronal populations that target the thalamus and brainstem directly, including the parafascicular thalamus (GPe ^Pf ) and the pedunculopontine nucleus (GPe ^PPN ). These projection-defined subpopulations exhibit distinct behavioral functions. GPe-PPN neurons are selectively suppressed at locomotor onset, and bidirectional manipulation of this pathway is sufficient to promote or suppress locomotion. Notably, D2-MSN stimulation upstream of these neurons evokes locomotion, and co-activation of GPe ^PPN pathway blocks this D2-MSN-evoked locomotion, demonstrating a disinhibitory circuit for action stemming from D2-MSNs. In contrast, GPe ^Pf neurons are not engaged during locomotion but are selectively recruited during skilled forelimb actions, and their activation disrupts forelimb movements without affecting locomotion. Together, these findings establish the GPe as a basal ganglia output hub composed of modules that mediate distinct behaviors. This organization revises canonical models of the indirect pathway by demonstrating that D2-MSNs can also facilitate, rather than only suppress, movement depending on the downstream GPe output channels they engage. In brief: The external globus pallidus (GPe), classically considered a relatively homogeneous relay within the indirect basal ganglia pathway, is revealed here as a major basal ganglia output hub. Distinct GPe subpopulations, defined by transcriptomic identity and projection target, form parallel channels to cortical, thalamic and brainstem structures with action-specific functions. These findings uncover a disinhibitory mechanism through which D2-MSNs can facilitate behavior, revising canonical models of basal ganglia function. Highlights: Brain-wide mapping reveals the GPe as a major basal ganglia output node GPe contains novel cell types defined by transcriptomic identity and projection specificity Projection-specific GPe circuits differentially control locomotion and skilled forelimb actions D2-MSNs can disinhibit downstream targets via specific GPe projections and promote action.

The striatum is classically thought to modulate action vigor and reinforcement, but whether it specifies and controls ongoing actions at a fine scale remains unclear. Here, we test whether striatal activity can resolve and control actions as granular as distinct patterns of muscle co-contraction within the same limb. We developed a closed-loop system combining two-photon imaging and holographic optogenetics to identify and manipulate action-specific striatal ensembles in real time. Mice performed isometric push or pull actions engaging the same forelimb muscles with different activation patterns. Ensembles of both D1- and D2-spiny projection neurons encoded action identity with equal fidelity. Causal stimulation of small ensembles biased ongoing force, but only when the ensembles encoded the ongoing action. These results show that the striatum specifies and controls ongoing actions in real time at a previously unappreciated granularity, redefining its role in motor control.

Animals can initiate movements either in response to external cues or from internal drive, yet how the brain flexibly supports both remains unclear. Disorders such as Parkinson’s disease disrupt these modes differently, suggesting distinct underlying mechanisms. Such differences could arise from specialized circuits dedicated to each action mode or from shared neuronal populations that shift their dynamics across contexts. To distinguish between these possibilities, we performed two-photon calcium imaging in the dorsolateral striatum as mice executed the same lever press either spontaneously or in response to a cue, enabling direct comparison of internally and externally triggered actions. Unsupervised clustering revealed subpopulations of neurons modulated during cue, movement, or post-action periods. Critically, the same neurons were tuned to movement across initiation contexts. However, the dynamics of this shared population evolved within distinct “context” and “action” subspaces, even before movement initiation. Both dopamine D1- and D2-receptor-expressing spiny projection neurons contributed to both subspaces, with D1-SPNs more active at the time of the sensory stimulus. These results show that context shapes neural dynamics within a shared movement-encoding population, revealing a context-generalizable striatal code that supports flexible movement initiation across internal and external drives.

The primary motor cortex (M1) is a central hub for motor learning and execution. M1 is composed of heterogeneous cell types with varying relationships to movement. Here, we tagged active neurons at different stages of motor task performance in mice and characterized cell type composition. We identified corticothalamic neurons (M1^CT) as consistently enriched with training progression. Using two-photon calcium imaging, we found that M1^CT activity is largely suppressed during movement, and this negative correlation augments with training. Increasing M1^CT activity through closed-loop optogenetic manipulations during forelimb movement significantly hinders execution, an effect that became stronger with training. Similar manipulations, however, had little effect on locomotion. In contrast, M1 corticospinal neurons positively correlate with movement, with an increase during training. We uncovered that M1^CT neurons suppress corticospinal activity via feedforward inhibition, also scaling with training. These results identify a permissive role of corticothalamic neurons in movement execution through disinhibition of corticospinal neurons.

Reaching movements are critical for survival, and are learned and controlled by distributed motor networks. Even though the thalamus is a highly interconnected node in these networks, its role in learning and controlling reaches remains underexplored. We report dissociable roles of two thalamic forelimb circuits coursing through parafascicular (Pf) and ventroanterior/ventrolateral (VAL) nuclei in refining reaches to a spatial target. Using 2-photon calcium imaging as mice learn directional reaches, we observe high reach-related activity from both circuits early in learning, which decreases with learning. Pf activity encodes reach direction early in learning, more so than VAL. Consistently, bilateral lesions of Pf before training impairs refinement of reach direction. Pre-training lesions of VAL does not affect reach direction, but increases reach speed and target overshoot. Lesions of either nucleus after training does not affect the execution of learned reaches. These findings reveal different thalamic circuits governing distinct aspects of learned reaches.

In Parkinson’s disease (PD), dopaminergic neurons (DANs) in the midbrain gradually degenerate, with ventral substantia nigra pars compacta (SNc) DANs exhibiting greater vulnerability. However, it remains unclear whether specific molecular subtypes of ventral SNc DANs are more susceptible to degeneration in PD, and if they contribute to the early motor symptoms associated with the disease. We identified a subtype of Sox6+ DANs, Anxa1+ , which are selectively lost earlier than other DANs, and with a time course that aligns with the development of motor symptoms in MitoPark mice. We generated a knock-in Cre mouse line for Anxa1+ DANs and showed differential anatomical inputs and outputs of this population. Further, we found that the inhibition of transmitter release in Anxa1+ neurons led to bradykinesia and tremor. Therefore, Anxa1+ is not only a biomarker of a selectively vulnerable subtype of DANs, but is also sufficient to drive early motor symptoms in Parkinson’s disease.