Cortico-basal ganglia plasticity in motor learning

One key function of the brain is to control our body’s movements, allowing us to interact with the world around us. Yet, many motor behaviors are not innate but require learning through repeated practice. Among the brain’s motor regions, the cortico-basal ganglia circuit is particularly crucial for acquiring and executing motor skills, and neuronal activity in these regions is directly linked to movement parameters. Cell-type-spe-ciﬁc adaptations of activity patterns and synaptic connectivity support the learning of new motor skills. Func-tionally, neuronal activity sequences become structured and associated with learned movements. On the synaptic level, speciﬁc connections become potentiated during learning through mechanisms such as long-term synaptic plasticity and dendritic spine dynamics, which are thought to mediate functional circuit plasticity. These synaptic and circuit adaptations within the cortico-basal ganglia circuitry are thus critical for motor skill acquisition, and disruptions in this plasticity can contribute to movement disorders.


INTRODUCTION
Throughout our lifespan, learning new motor skills and adapting our motor behavior are crucial for our survival.Whether we are learning to ride a bike in childhood or learning to play a new instrument in adulthood, our brain adapts and undergoes plasticity to acquire and execute these skills after practice.In parallel, the brain needs to adapt its motor output to changes in our body, such as muscle growth and decline throughout life.Importantly, plasticity of the brain motor system is also required to adapt to body injuries or recover from brain diseases.Thus, understanding the mechanisms by which the brain undergoes plasticity during motor learning is critical for gaining insight into one of the most fundamental roles of the brain.These insights also guide the development of new treatment strategies for neurological diseases and injury rehabilitation.
Research on brain motor function has been a longstanding focus of neuroscientists, but technical advances in the recent decade have propelled our understanding of the brain regions, cell types, and circuits involved in motor control and how they adapt during motor learning.Additionally, we are beginning to uncover the synaptic mechanisms that help shape the adaptations in circuit activity.Such adaptations evolve over the course of learning, and learned neuronal activity patterns could be maintained for long periods, reflecting the long-lasting nature of motor memories.The functional plasticity observed in motor learning is accompanied by structural adaptations in neuronal circuits and changes in the efficacy of their synaptic connections, i.e., synaptic plasticity.Functional and structural synaptic changes likely influence each other-for example, during learning, circuits that are repeatedly engaged cause synaptic plasticity, which, in turn, can lead to these circuits being prefer-entially activated during the execution of a learned behavior.Despite this implied interplay between functional, structural, and synaptic adaptations, how these forms of plasticity are linked during motor learning has been challenging to parse experimentally.
The rise in neural recording devices with potential for human use has fueled a recent push to understand how neuronal activity patterns in the brain encode movement, given the prospect of using such devices to help patients control prosthetics.Though there has been tremendous progress in decoding brain motor activity and generating computational models linking brain activity to movement parameters, experimental studies underlying these models often rely on well-trained-if not over-trainedmotor tasks.Focusing on brain mechanisms involved in the acquisition of motor skills is, on one hand, important for gaining a deeper knowledge of the basic science of how the brain optimizes its neural circuits to acquire and refine motor skills.This can potentially inform more efficient and effective training regimens for learning new skills.On the other hand, this knowledge is critical for developing interventions for neurological disorders that affect motor function, such as Parkinson's disease (PD) and Huntington's disease, as well as for engaging neural plasticity in the recovery from neurological injuries, such as strokes or traumatic brain injuries.
In this review article, we will summarize the latest research on the role of synaptic and circuit mechanisms in motor learning, with a focus on cortico-basal ganglia motor circuits.We will discuss the intricate interplay between the cortex and basal ganglia with an emphasis on the rodent models, given their advantages for circuit and synaptic studies.We will describe the role of cortico-basal ganglia circuits in motor control, followed by a discussion of circuit adaptations during motor learning.
We will further explore the intricate synaptic plasticity mechanisms underlying these circuit changes as well as examine how these plasticity mechanisms can be disrupted in movement disorders.

CIRCUITS FOR PRECISE MOTOR CONTROL
The orchestration of precise movements is governed by a complex set of neuronal circuits within the central nervous system.Inside the brain, several major structures are involved in motor control (Figure 1).These include the motor cortex, the basal ganglia, the cerebellum, the motor thalamus, and brain stem motor areas.Classically, loss-of-function and electrical brain stimulation studies, whether observed in diseases or induced in experiments, were used to delineate the distinct roles of individual brain regions and circuits.Such approaches, for example, defined the motor map in the cortex (homunculus), where different regions of the motor cortex control muscles in corresponding parts of the body.More recent optogenetic activation or inactivation experiments are able to delineate the role of individual brain regions and circuits with higher temporal precision during different aspects of motor behavior. 1,2Over the last few decades, the broad functional roles of these brain motor regions have been established, with the motor cortex involved in movement planning and initiation 3,4 ; the basal ganglia in voluntary movement, action selection, and habit formation 5,6 ; and the cerebellum in sensory integration and motor predictions. 7,8The motor thalamus is thought to act as a relay of motor information from subcortical regions, 9 and the brain stem motor areas control movement elements or behavioral syllables and locomotion. 10,11On a macro-circuitry level, the motor cortex, basal ganglia, cerebellum, and thalamus form two loops that are largely regarded as separate and thought to primarily interact with each other at the cortical level.Yet, there is evidence that subcortical connections between the basal ganglia and cerebellum exist 12,13 and that the thalamus might play a role in integrating projections from the basal ganglia and cerebellum. 14Both systems can send direct projections to lower motor neurons in the spinal cord or target brain stem motor areas, which, in turn, act as an output to control the spinal cord.
In this review, we will focus primarily on the cortico-basal ganglia system, given its importance for the learning and acquisition of skilled motor tasks.For further reading on the role of the cerebellum or the brainstem in motor learning, please refer to Manto et al., 7 Carey, 8 Arber and Costa, 10 Leiras et al., 11 and De Zeeuw and Ten Brinke. 15rtico-striatal circuit Within the motor cortex, the primary motor cortex (M1) is involved in the execution of movements, whereas higher-order motor areas, such as the secondary motor cortex (M2) and the supplementary motor area (SMA), are involved in motor planning.These cortical areas are heavily interconnected with each other and send long-range excitatory output projections to the striatum, brainstem, and spinal cord.The striatum-the input nucleus of the basal ganglia-is predominantly comprised of dopamine D1 receptor (D1R)-expressing direct pathway spiny projection neurons (dSPNs) and D2 receptor (D2R)-expressing SPNs of indirect pathway (iSPNs).Within the striatum, the dorsolateral part (DLS) is particularly involved in the control of movement.In addition to these projection neurons, both the motor cortex and the striatum contain a variety of local interneurons (Figure 2).][21][22] Striatal innervations from the motor cortex mainly originate from two classes of excitatory glutamatergic neurons: pyramidal tract (PT) and intratelencephalic (IT) neurons (Figure 2). 25 PT neurons are exclusively located in layer 5B (L5B) of the motor cortex and send axonal projections toward the brainstem and spinal cord with collateral axons targeting the striatum, thalamus, and other cortical areas.IT neurons are located at layers 2/3 (L2/3), 5A (L5A), and 5B (L5B), sending their axons to the striatum and other cortical areas. 23,247][28][29] For example, on a forelimb reaching task, IT neurons represent reaching amplitude, whereas PT neurons represent movement direction. 26It is important to note that while striatal input from collateral axons Overview of prominent brain motor circuits, including primary and secondary motor cortex (M1 and M2, blue); the basal ganglia system (red) with its input structure, the striatum, external globus pallidus (GPe) and subthalamic nucleus (STN), as well as its outputs, the internal globus pallidus (GPi) and substantia nigra pars reticulata (SNr), and the dopaminergic substantia nigra pars compacta (SNc); the thalamus (orange) with the motor thalamus (MoThal) consisting of the ventromedial (VM) and ventral anterolateral (VAL) nuclei as well as the parafascicular nucleus (Pf); the cerebellum (gray) with its outputs, the deep cerebellar nuclei (DCN); and the brainstem motor nuclei (green).Major projection pathways are shown by arrows.
3][34] Cortico-striatal inputs in primates might predominantly originate from IT-type cortical neurons. 32,35In addition to inputs from the motor cortex, the striatum also receives excitatory input from the thalamus, mainly from the parafascicular (Pf) and central medial nuclei (CM). 36,37These inputs play a crucial role in movement initiation; for example, when rodents are trained on a sequential lever push task. 38mportantly, the striatum is also innervated by projections from midbrain dopaminergic neurons, which are thought to carry movement, reward, and aversion-related signals. 39Dopaminergic projections to the striatum most notably encode a reward prediction error (RPE) and act as a teaching signal for reinforcement learning. 402][43] Striatal dopamine projections arise from a diverse population of midbrain neurons 39,44,45 ; for example, reward-related and movement-related signals arise from separate populations of midbrain dopaminergic neurons. 42Axons from the substantia nigra pars compacta (SNc) are mostly activated by locomotion, while dopaminergic axons from the ventral tegmental area (VTA) are mostly activated by reward presentation. 42Interestingly, recent studies have found that dopamine axonal activity is possible without an action potential generated at the midbrain soma, likely mediated by local cholinergic signaling. 46,47Such studies suggest a diverse and complex role of neuromodulatory activity that can influence striatal signaling during movement.
Overall, recording the neuronal activity during movement in these brain circuits revealed a variety of movement parameters encoded within these different neuronal populations, ranging from movement planning and initiation to vigor and velocity, as well as reward and trial outcome.These observations together have been used to develop theoretical models that can start to explain how the brain controls movement.Many models are taking the overall population activity of neurons into account because the heterogeneity and variability in the activity of individual neurons across repeated movements and trials suggests that it is unlikely that such parameters are encoded by individual neurons. 48Therefore the dynamical systems perspective of movement control, spearheaded by Krishna Shenoy, has emerged as a prominent theoretical framework to start describing the relationship between neural population activity and movement parameters. 4

CIRCUIT ADAPTATIONS DURING MOTOR LEARNING
One key aspect of the motor system is our ability to learn and acquire new motor skills throughout adulthood, which is supported by the brain's ability to adapt and undergo plasticity.Adaptations of neural activity have been recorded in a variety of animal species, but given their advantages for circuit and synaptic studies, rodent models have been widely used to study motor learning.A range of motor tasks have been used to train mice or rats, to study how novel motor skills are acquired and what adaptations in the brain accompany this learning (Table 1).Some tasks involve whole-body-wide movement adaptations, while others focus on the training of novel forelimb movements.For example, a popular task is the forelimb reaching task in which mice or rats are food restricted and trained to reach for, grasp, and retrieve a small food pellet through a small opening in the training chamber (Figure 3A).Though this task is typically performed in freely moving rodents, it has been recently adapted for head-fixed mice to be compatible with in vivo recording and imaging. 1,49 similar head-fixed forelimb movement task is the lever push or pull task in which head-restrained mice are trained to move a lever across a certain distance threshold to receive a water reward (Figure 3E).A task that, by design, involves whole body movements is the treadmill running task in which rodents are trained to run on a treadmill while being head-fixed (Figure 3I).Although these tasks appear simple to us as humans, rodents typically require 20-30-min-long training session each day for 1-2 weeks to acquire these new motor skills and gain proficiency, with increasing rates of success to retrieve food pellets, move the lever across threshold, or increase the distance run on the treadmill.Over the course of Two classes of excitatory neurons in motor cortex-pyramidal tract (PT) and intratelencephalic (IT) neurons-innervate the striatum.Although PT neurons are exclusively located in layer 5B (L5B) and send axonal collaterals toward the brainstem, spinal cord, thalamus, and striatum, IT neurons are in layers 2/3, 5A, and 5B and only target the striatum as well as other cortical areas. 23,24In addition, the cortex is comprised of three major classes of inhibitory interneurons-somatostatin (SOM)-, parvalbumin (PV)-, and vasoactive intestinal peptide (VIP)-expressing neurons.The striatum acts as the input nucleus of the basal ganglia system and is predominantly comprised of D1R-expressing direct pathway spiny projection neurons (dSPNs) and D2R-expressing SPNs of the indirect pathway (iSPNs).Cholinergic (ChIn) and GABAergic (including PV-, SOM-, calretinin-, and tyrosine-hydroxylase-expressing) neurons make up the majority of local interneurons in the striatum.Additional input projections to the cortex include thalamo-cortical and cortico-cortical projections, and striatal neurons receive additional input from thalamo-striatal projections.Dopaminergic projections predominantly target the striatum but also motor cortex.
motor learning, rodents also acquire more stereotypic movements to execute the tasks.For example, movement trajectories become more consistent with the learning progress in a reaching task (Figures 3C, 3D, 3G, and 3H) and, on a forced movement version of the treadmill task, steps become more structured (Figures 3K and 3L).

Plasticity of neuronal activity in motor cortex
95,96 Such studies revealed that the representation of individual movements changes with learning, and the activity correlation of neurons with similar response properties increases. 95,96An important study from Peters et al. further showed that learning a lever push task leads to the formation of precise spatiotemporal activity patterns within M1 (Figures 4A-4D).Specifically, excitatory neurons in M1 L2/3 exhibit activity that becomes more aligned with movement onset following motor learning and the temporal activity sequence of L2/3 neurons becomes more stable and correlated to the learned movement trajectory. 81Interestingly, this is not observed in M1 L5 corticospinal PT neurons where, instead of an increased association of similar movements and similar activity patterns, dissimilar movements become associated with more dissimilar activity patterns. 82

Plasticity of neuronal activity in striatum
In the striatum, neuronal activity patterns also adapt over the course of motor learning in adult mice.For example, following motor learning, neuronal activity becomes correlated with the initiation and termination of action sequences. 78Similar to the motor cortex, striatal SPNs develop reproducible firing sequences that are specifically associated with the learned stereotyped movements (Figures 4E-4H). 84This study found that the fraction of neurons active during movement on a lever-pull task increases for both D1 and D2 SPNs with motor learning.Notably, in trained mice, a larger fraction of D1 SPNs are movement selective compared with D2 SPNs.
Overall, these results suggest that neurons in both the M1 and the DLS adapt their activity during motor learning by changing their response selectivity to specific aspects of movement.In general, they become more active at movement onset 78,81,84 and form precise spatiotemporal activity patterns that reflect the stereotypic movements formed after learning. 81,84However, the exact adaptations differ between individual neuron types.
An interesting question these studies raise is whether cortical and subcortical areas like the striatum are correlated and equally contribute to motor learning and movement execution.Although cortical and striatal activity are overall strongly linked, 97 their precise relationship can change over the course of learning.For example, some studies found that acute optogenetic inhibition of the motor cortex can impair the execution of a learned forelimb reaching task, 1 yet other studies found that the motor cortex is only required for learning new motor skills but not for executing trained skills. 79This seems to be particularly the case when motor skills are trained for extended periods of time (60 days in mice), at which point acute inhibition of the motor cortex no  50 Yang et al., 51 Hayashi-Takagi et al., 52 and Yang et al. 53 Complex wheel freely moving McKenzie et al. 54 and Miyamoto et al. 55 Split treadmill freely moving Darmohray et al. 56 Head-fixed locomotion head-fixed Yang et al., 53 Adler et al., 57 Cichon and Gan, 58 and Ma et al. 59 Forelimb specific Forelimb reaching task freely moving Arroyo et al., 60 Rioult-Pedotti et al., 61 Albarran et al., 62 Suresh and Dunaevsky, 63 Fu et al., 64 Hwang et al., 65 Guo et al., 66 Sohn et al., 67 Bacmeister et al., 68 Roth et al., 69 Biane et al., 70 Padmashri et al., 71 Withers and Greenough, 72 Greenough et al., 73 Gloor et al., 74 Harms et al., 75 Wang et al., 76 and Xu et al. 77 Sequential lever push freely moving Jin and Costa, 78 Kawai et al., 79 and Wolff et al. 80 Forelimb reaching task head-fixed Guo et al. 1 and Yang et al. 49 Lever push/pull head-fixed Peters et al. 81,82 Masamizu et al., 83 Sheng et al., 84 Hwang et al., 85 Ren et al., 86 and Hedrick et al. 87 Robotic manipulandum head-fixed Wagner et al. 88,89 and Sibener et al. 90 longer has an effect on task execution. 85This disengagement of the motor cortex is accompanied by further adaptation of neuronal activity, such that M1 L2/3 activity, which becomes more correlated and associated with movement trajectories within the first 2 weeks of learning, 81 becomes decorrelated and unassociated from movement trajectories with further training up to 60 days. 85Similarly, lesioning the motor cortex in rats trained to perform a sequential lever press affects the behavior when the lesion occurs pre-training and not after rats are well trained. 79A follow-up study showed that, specifically, the activity of cortico-striatal projections is required for learning and dispensable for learned behaviors, but thalamostriatal projections are required for the execution of learned skills. 80

Role of inhibitory interneurons
Currently, the adaptation of neuronal activity during motor learning is best described for the major projection neuron classes in the motor cortex (excitatory L2/3 and L5 pyramidal neurons) and striatum (dSPNs and iSPNs).These activity patterns are likely influenced by excitatory input projections, but also local inhibitory interneurons.All three major subtypes of cortical inhibitory neurons (SOM, PV, and VIP) in the motor cortex are activated by movement, 57,49,86,98 though, overall, VIP and SOM neurons appear to be the most dynamic over the course of motor learning.Specifically, the movement-related activity of VIP neurons is reduced and SOM activity is increased with learning.Interestingly, such adaptations are not limited to the motor cortex but are distributed across several cortical sensory areas. 86 neurons undergo a modest increase in activity, which is mostly limited to motor cortical areas.
Although the activity of cortical SOM interneurons is increased overall during motor learning, 86 a subset of SOM interneurons that are task-related and express the immediate early gene (IEG) NPAS4 during early stages of learning show relatively lower activity during movement.This reduction in activity is required for learning a forelimb reaching task, as chemogenetic activation of NPAS4-expressing SOM neurons impairs learning. 49Similarly, increasing the activity in all motor cortex SOM neurons also impaired learning a treadmill running task in young adult mice. 57The task-specific activity of SOM neurons is further corroborated by the finding that SOM neuron activity is differentially regulated by different motor tasks, such as showing increased activity during forward running on a treadmill and reduced activity during backward running. 57hus, proper SOM neuron activity is critical for activity adaptations in L2/3 neurons, as described above, 81 and chemogenetically increasing SOM activity disrupts such L2/3 activity adaptation during learning. 57SOM activity during learning is likely regulated by VIP interneurons because, in the cortex, VIP interneurons have a strong inhibitory effect on SOM neurons, 16 and VIP neurons are activated during movement. 57However, how VIP neurons can specifically regulate task-relevant SOM neurons and whether there are task-relevant VIP neurons that increase in activity over the course of learning remains an open question.Following motor learning, the temporal activity pattern of SOM neurons is shifted toward later onsets during the movement execution, 57 while VIP activity shifted toward earlier onsets and even shows preparatory activity before movement onset in M2. 60 These temporal shifts in VIP and SOM activity might play a role in adaptations of L2/3 neuron activity toward earlier onsets following learning. 81lthough these studies have begun to illuminate how inhibitory activity in the motor cortex adapts during motor learning and contributes to the plasticity of excitatory neuron activity, little is known about the role of local inhibitory interneurons in the striatum for motor learning.Like the cortex, the striatum contains a variety of interneuron classes, 19 including PV-positive fast spiking interneurons.PV neuron activity, for example, inhibits SPN output activity 99 overall and plays a role in cue-guided discrimination tasks 100 as well as associative learning. 99Inhibition of PV interneurons in the DLS can also lead to abnormal movements or dyskinesia. 101verall, it is clear that the plasticity of neuronal activity in both the motor cortex and striatum plays a crucial role in motor learning.Although more evidence is emerging regarding how distinct cell types and circuits within the cortico-striatal system contribute to these adaptations, the stability and long-lasting nature of activity changes following motor learning likely require physical rewiring of the synaptic con-nections between these cell types and circuits to support long-lasting motor memories.

SYNAPTIC PLASTICITY MECHANISMS UNDERLYING CIRCUIT ADAPTATIONS
Synaptic plasticity is the ability of synapses to undergo long-lasting modifications in strength and efficacy and is fundamental to the adaptive nature of the nervous system.There are many forms of synaptic plasticity, including postsynaptic long-term potentiation (LTP) as well as long-term depression (LTD), which increases or decreases synaptic strength, respectively.On a molecular and cellular level, synaptic plasticity can be mediated by increasing or decreasing the number of synaptic connections (e.g., dendritic spine dynamics) 102 or by changing the efficacy of individual synapses through dynamics in the number of neurotransmitter receptors (e.g., AMPA-type glutamate receptors) 103 or altering receptor signaling. 104Many forms of synaptic plasticity are engaged during motor learning and may play a role in shaping the functional adaptations in circuit activity described above.It is important to note that these plasticity mechanisms are differentially engaged at different ages of animals.6][107][108] Unless specifically stated, all motor learning studies mentioned here were performed in adult mice or rats (over 2 months of age).

Synaptic plasticity in the motor cortex
Acquisition of novel motor skills is accompanied by changes in the number and strength of synaptic connections in the M1.
Early electrophysiological studies demonstrated that brain slices from rats trained on a forelimb reaching task showed an overall strengthening of intracortical connections in L2/3 of the motor cortex, which is specific to the hemisphere contralateral to the trained paw and likely occurs through an LTP-like mechanism. 61,109This increase in synaptic strength is medi-ated through an increase in postsynaptic AMPA-type glutamate receptors, which are the main mediators of fast excitatory synaptic transmission.Following training on a rotarod or forelimb reaching task, brain slices from young adult (1-2 months old) rats showed an increased AMPA receptor (AMPAR)/NMDA receptor (NMDAR) ratio as well as an increase in excitatory postsynaptic current (EPSC) amplitude and frequency. 50,70These electrophysiological findings are further corroborated by biochemistry experiments showing increased synaptic AMPAR levels following learning of a rotarod or forelimb reaching task in young adult and adult rodents. 50,71,69Further studies have shown a significant structural reorganization that can underly these changes in synaptic strength.3][74] Additional electron microscopy (EM) and slice immunohistochemistry experiments have shown that the number of dendritic spines on L2/3 and L5 neurons is increased during learning. 75,76,110lthough these experiments relying on post hoc measurements of trained mice establish that motor learning is accompanied by an overall increase in synaptic strength, they do not provide information on the time course of synaptic plasticity, how individual synapses change, or the spatial distribution of synaptic changes within the motor cortex.By using in vivo two-photon imaging, more recent studies were able to longitudinally monitor neuronal and synaptic structures over the course of motor learning in mice.Specifically, these studies measured the dynamics of dendritic spines, which serve as the location of most excitatory synapses on pyramidal cortical neurons, using mice expressing a fluorescent protein in a sparse subset of neurons, such as the Thy1-YFP mouse line, which sparsely expresses yellow fluorescent protein (YFP) in the cortex (Figure 5). 102,111During learning of either a rotarod or forelimb reaching task, a transient increase in the total number of dendritic spines on L5 neurons in the M1 could be observed in young and adult mice.This was caused by an increase in spine formation rates within the first days of training and a delayed increase in spine elimination. 51,77Supporting a link between spine formation and motor skill learning, the rate of spine formation tightly correlated with the motor task performance.Newly formed spines are further selectively stabilized during motor learning and are maintained for at least 4 months after learning, at which point mice are still performing the trained task with high proficiency. 62,77ecent studies further examined how the strength of stable synapses changes during motor learning by imaging fluorescently tagged AMPARs (e.g., the AMPAR subunit GluA1 tagged with Super-Ecliptic pHluorin [SEP-GluA1]) in mice, which serves as a proxy for synaptic strength given the correlation between the amount of AMPARs and excitatory currents at synapses. 112These studies found that motor learning in adult mice induces the potentiation of a subset of stable dendritic spines and the potentiation of these spines is maintained after daily training ended. 63,69Importantly, this increase in spinal AMPAR levels correlates with the degree of motor learning.(F) Schematic summary of spine dynamics and plasticity during motor learning.During early phases of learning new spines are formed and existing synapses are strengthened.In late phases, initially potentiated synapses remain potentiated and spines formed in spatial clusters are selectively stabilized. 64,77,69uch in vivo imaging studies are also able to reveal the spatial distribution of dynamic synapses and found that spine formation during motor learning is preferentially occurring in local clusters along individual dendritic segments 64 around movement-related spines. 87Similarly, the potentiation of stable spines occurs in local clusters during learning, 69 potentially reflecting that spines with similar activity during the execution of motor tasks are spatially clustered. 87Functionally, this clustering of synaptic connections can further strengthen the effect of synaptic signaling from clustered inputs through dendritic mechanisms such as non-linear integration. 113he majority of current work has been focused on different forms of synaptic strengthening, such as new spine formation and LTP.Yet, there is clear evidence that synaptic weakening or LTD plays an important role in learning and memory as well, 114,115 and the removal of dendritic spines in the motor cortex has been shown to accompany motor learning. 62,77,116hether such spine elimination is an active process with a specific role in supporting learned movements or is a homeostatic response to synaptic strengthening and LTP is less understood.Further studies examining the functional role of LTD in motor learning will be needed to gain a more comprehensive understanding of the interplay between these different plasticity mechanisms.

Synaptic plasticity in the striatum
Despite the important role synaptic plasticity in the striatum plays for motor learning, 117 as well as the breadth of studies on plasticity mechanisms within the striatum, [118][119][120] little is known about the spatiotemporal distribution of plasticity at striatal synapses.This is partly due to the technical limitation of optically accessing such deep brain structures for high-resolution imaging and the widespread and converging cortico-striatal projections.To circumvent this challenge, a recent study combining genetic strategies to target task-relevant neurons with slice electrophysiology found that cortico-striatal synaptic connections from task-relevant motor cortex neurons are strengthened following motor learning. 65Similar to the motor cortex, these projections converge onto spatially clustered spines in SPNs.
An important regulator of synaptic plasticity in the striatum is dopaminergic signaling, [120][121][122][123][124] and disrupted dopamine signaling in movement disorders disrupt proper striatal plasticity (see section on role of synaptic plasticity in motor disorders below).Particularly, the precise timing of dopamine input in relation to glutamatergic input plays a crucial role in promoting postsynaptic plasticity at striatal dSPNs. 125At the same time, dopaminergic signaling also increases the intrinsic excitability of dSPNs on longer timescales of up to minutes, 126 providing a further possible mechanism to facilitate striatal synaptic plasticity.
Cell-type and input-specific spine plasticity One significant, if not entirely surprising, finding from studies on synaptic plasticity in vivo is that not all synapses undergo plasticity during motor learning.Only a small subset gets potentiated, added, or eliminated.Depending on the study, 10%-15% of spines are potentiated and 5%-10% of spines are added and eliminated specifically during motor learning.This raises important questions about whether such synaptic plasticity occurs on specific populations of neurons and what the presynaptic inputs to these dendritic spines are.
Within the M1, spine potentiation, either by measuring synaptic AMPAR levels or spine size, is observed in excitatory pyramidal neurons in L2/3 63,81,87 as well as in L5. 58,69 Interestingly, synaptic potentiation over the course of learning a reaching task has also been observed in sensory brain regions such as the primary visual cortex, 69 suggesting that motor learning also involves long-lasting changes in sensory processing.Similarly, new spine formation has been clearly demonstrated in L5 pyramidal neurons, 51,64,66,77 whereas results from L2/3 neurons are ambiguous. 81,127Although most of these forms of synaptic plasticity are specific to the cortex contralateral of the trained paw in tasks involving the training of one paw specifically (for example, see Rioult-Pedotti et al., 61 Suresh and Dunaevsky, 63 and Xu et al. 77 ), increases in spinal GluA1 were observed in both hemispheres. 69This suggests that distinct forms of plasticity can contribute to overall plasticity in different brain regions and neuronal populations during motor learning.
Further investigations into which population of neurons undergo plasticity while acquiring a new motor task have found that inputs onto neurons that are behaviorally relevant are specifically potentiated.For example, miniature EPSC (mEPSC) amplitude of M1 neurons projecting to parts of the spinal cord that are associated with the control of distal forelimb movements rather than proximal forelimb movements is increased after learning a pellet grasping task. 70More recently, Hwang et al. showed that motor learning increases spine density, spine dynamics, and new spine survival selectively on M1 neurons that were active and expressing the IEG c-Fos during task execution. 65n the presynaptic input side, it has been shown that while newly formed transient spines are targeted by corticocortical projections, newly formed spines that become stabilized are targeted by thalamocortical projections in young adult mice. 67ewly formed spines are also more likely to form synapses with axonal projections that previously did not target their nearby dendritic segment.Yet, potentiated and newly formed persistent spines receive synaptic input activity related to the learned movement. 87Together, these findings suggest that within M1, motor learning selectively drives the formation of new synapses on neurons that are active during behavior from behaviorally relevant thalamic input projections.
Despite this range of strong evidence for spine dynamics and plasticity playing a crucial role during motor learning and the fact that spine formation, stabilization, 62,77 and AMPAR insertion correlate with task performance, 63,69 direct causal evidence supporting the necessity of such plasticity for the learning process remains scarce.One important study, however, addressed this question by expressing an engineered protein that can target recently potentiated dendritic spines and induce their shrinkage by light exposure.Using this approach to artificially weaken new spines potentiated during learning of a rotarod task, this study found significant impairment of the learned behavior, 52 thus suggesting a causal relationship between spine potentiation in the motor cortex and motor learning.
9][130] In the laboratory, a variety of motor tasks are used to study motor function and motor learning in human and non-human primates, such as reaching tasks involving the control of a joystick (Figures 6A and 6B), 131 sequential finger tapping or button pressing tasks, or pinching tasks, 132 which are similar and involve largely the same human equivalent brain areas as forelimb reaching tasks in rodents (Figure 3).Electrophysiological recordings using Utah Arrays, Michigan-style planar electrodes, or high-density probes, such as Neuropixels, in non-human primates have in particular been used to understand the neural population dynamics underlying movement parameters.Given the more complex and sophisticated behavior measurements possible in primates compared with rodents, these studies can provide a more precise picture of how different regions in the brain encode muscle activity and body movements.
0][141][142] Specifically, strengthening of motor cortex outputs in an LTP-like fashion can be induced using transcranial magnetic stimulation (TMS) or transcranial direct current stimulation (tDCS), which results in a long-lasting increase in the amplitude of a motor-evoked potential (MEP) at the muscle. 143One way to test whether LTP-like mechanisms are also engaged during motor learning is to take advantage of the fact that neuronal circuits have a ceiling for the amount of potentiation they can undergo.By measuring how much LTP can still be artificially induced following motor learning, i.e., measuring the amount of LTP-occlusion, studies have shown that motor cortex circuits undergo LTP in humans following motor learning, with larger degrees of LTP induction correlating with higher levels of skill retention. 132 major goal of research on the neural mechanisms encoding human motor control is to enable the decoding of brain activity to help brain-computer interfaces (BCIs) read movement intentions to control prostheses.For example, recording the neural activity from the human motor cortex using multichannel silicon microelectrodes can be used to decode the hand movement intentions and control a computer cursor on a screen (Figure 6C) 135,144 or control a robotic arm. 145In a similar approach, the neural signals encoding handwriting intentions or speech intentions have been recorded and used to create reliable BCIs through which human subjects can produce text. 146,147verall, systems and behavioral studies of motor control and motor learning in human and non-human primates have yielded a rich and high-level understanding of the population-level encoding of movement parameters that can be exploited to control BCIs in patients with lost movement or speech abilities.Importantly, many of the same brain regions have been found to be involved in motor learning across species, and synaptic plasticity mechanisms first discovered in rodents have been observed to play a role in human motor learning as well.Yet, gaining an understanding of the intricacies of circuit adaptation and synaptic mechanisms involved in motor learning is challenging in higher-order organisms.This highlights the value of rodents in movement research to study basic circuit and synaptic mechanisms that can further inform human studies and therapies.

ROLE OF SYNAPTIC PLASTICITY IN MOTOR DISORDERS
Movement disorders encompass a diverse array of conditions that disrupt the finely tuned activity patterns in brain motor regions, leading to impairments in movement control.In many motor disorders, dysfunctions and alterations in synaptic function and plasticity precede more overt neuropathological hallmarks, such as cell death, and play an important role in the manifestation of movement deficits.Understanding the synaptic and  133 ).(B) Schematic of human performing a joystick reaching task (left).A divergent force field is applied such that any deviation from a straight reaching trajectory will generate a force that acts in the same direction.Humans learn to make straight movements in this task setting (right) (B adapted from Davidson and Wolpert 134 ).(C) Motor cortex recordings are used to control neuroprosthetic devices, such as moving a cursor on a screen in a center-out movement trajectory (C adapted from Wilson et al. 135 ).cellular underpinnings of movement disorders is thus crucial for developing targeted therapeutic interventions.One prominent neurodegenerative movement disorder is PD, which primarily affects the basal ganglia circuitry.In PD, there is a gradual loss of dopaminergic neurons in the substantia nigra, leading to a reduction in dopamine levels.PD manifests with significant motor symptoms, such as resting tremors, bradykinesia, rigidity, and postural instability.Notably, PD is associated with significant disruptions in striatal and cortical synaptic function and plasticity.

Impaired synaptic function and plasticity in PD
In rodent models of PD, such as with dopaminergic neuron lesions induced by the neurotoxin 6-hydroxydopamine (6-OHDA), 148 there is a drastic reduction in the number of dendritic spines on striatal SPNs. 149This reduction reflects a loss of glutamatergic input to SPNs and is specific to D2R-expressing iSPNs in rodents. 150However, both D1R-expressing dSPNs and D2Rexpressing iSPNs have reduced dendritic arborization in PD, leading to an overall decrease in the number of spines. 151uch a reduction in spine density has also been observed in post-mortem tissue of human PD patients 152,153 and primate models of PD, 154 where dSPNs and iSPNs are both affected.The implications of this dendritic spine loss on overall synaptic inputs to SPNs are less clear because these effects can be compensated by increased expression of AMPARs 155 and increased size of synaptic terminals. 156Notably, dopamine signaling plays a critical role in proper synaptic plasticity in the striatum 122,125,157 and loss of dopaminergic innervation in PD can lead to changes in plasticity.Specifically, loss of dopamine in a mouse model of PD leads to a lower threshold for inducing LTP in iSPNs and a lower threshold for inducing LTD in dSPNs. 157These striatal impairments of synaptic function in PD may contribute to disrupted neuronal activity patterns in Parkinsonian mice. 158he M1 is also targeted by midbrain dopaminergic projections (Figure 1), 159,160 and loss of dopamine in M1 of PD model mice leads to aberrant dendritic spine dynamics and synaptic plasticity. 66Specifically, both spine formation and spine elimination are increased and LTP is impaired in 6-OHDA or 1-methyl-4phenyl-1,2,3,6-tetrahydropyridine (MPTP)-treated mice.This impaired motor cortex plasticity might play an important role in motor learning deficits observed in PD, as MPTP-treated mice are significantly impaired in their learning of a forelimb reaching task, and spine dynamics normally observed during learning are masked by enhanced spine turnover in PD mice. 66naptic deficits in other movement-related diseases Abnormal synaptic function and plasticity are also key mechanisms underlying a range of other disorders involving movement deficits.For example, in Huntington's disease, characterized by a hallmark degeneration of striatal SPNs 161 and the development of involuntary movements and impaired motor coordination, 162 cortico-striatal synaptic function and plasticity are disrupted and contribute to motor deficits. 163,164n addition, several neurodevelopmental disorders affect synaptic function and plasticity in the cortico-striatal circuit and manifest with deficits in motor control and motor learning.
Autism spectrum disorders, for example, are associated with delayed motor learning. 165,166Studies using Fragile-X or 16p11.2deletion mouse models have found that abnormal cortical circuit activity 167,168 and deficits in spine dynamics and synaptic plasticity 63,71,168 can lead to delays or impairment of motor learning.][172][173] Overall, these deficits in synaptic and circuit function observed in a variety of neurodegenerative and neurodevelopmental movement disorders underline the importance of understanding how circuit and synaptic mechanisms function in the healthy brain to promote movement and motor learning.Furthermore, synaptic plasticity mechanisms, especially within the motor cortex, play an important role in the recovery from brain injuries, such as strokes. 174,175A greater understanding of how synaptic plasticity enables motor learning in healthy individuals can inform how to take advantage of such mechanisms for behavioral recovery.

FUTURE DIRECTIONS AND CHALLENGES
Overall, past and current studies of cortico-basal ganglia plasticity during motor learning have provided valuable insights into the intricate circuit and synaptic mechanisms governing adaptive neuronal activity and movement control.Yet, the field is still far from having a comprehensive understanding of how the brain generates movement signals and what plasticity mechanisms are engaged in forming new motor skills.There are a few key challenges and open questions that future research will need to address, which we want to highlight here.

Bridging the gap between population dynamics, individual neuronal activity, and synaptic plasticity
In broad terms, motor control and motor learning have been studied on three different levels.On the system level, we are beginning to understand how neural population dynamics (Figure 7A) encode movement parameters and how population activity trajectories can adapt with learning.On the cellular level, we are learning how the activity of individual neurons adapts over the course of motor learning with circuit and cell-type specificity (Figure 4).Lastly, on the subcellular and molecular level, we are starting to identify the synaptic mechanisms involved in motor learning (Figure 5).Yet, how the processes on these three levels are linked and influence each other remains an open question.Although knowing the neural population dynamics might be more valuable than trying to interpret individual neuron activity for decoding movement, 48 understanding the neuronal activity of specific cell types, circuits, and projections is critical for understanding how movement representations are formed or adapted during learning.This is particularly important in the context of motor disorders, where restoring movement representations by manipulating specific cell types or circuits might represent a viable therapeutic approach.In addition, understanding how the activity of specific neurons adapts during learning is important for linking overall activity changes to synaptic plasticity mechanisms.Although it is likely that the emergence of stereotyped neuronal activity sequences in the motor cortex and striatum (Figure 4) 81,84 is accompanied by synaptic plasticity between these neurons, such that the activity of one neuron increases the likelihood of the next neuron in the sequence being activated (Figure 7B), this has not been experimentally shown.It also remains unclear whether such synaptic plasticity is the cause for the emergence of neuronal activity sequences or whether patterned neuronal activity leads to synaptic plasticity, engaging mechanisms like spike-timing-dependent plasticity (STDP-LTP and STDP-LTD).Synaptic plasticity can, in turn, reorganize the circuit via mechanisms like clustered potentiation of specific projections (e.g., task-related cortico-striatal projections; Figure 7C). 65Such clustering has been shown during motor learning, both in the striatum 65 and the motor cortex. 64,87,69tudies mostly done in cortical pyramidal neurons have found that such clustering of synaptic connections can have a supralinear effect on neuronal activity through local dendritic computations. 113Additional studies will be required to clarify the link between population dynamics, individual neuronal activity, and synaptic plasticity.
Despite rapid technological progress in the past decade, there are still key technical hurdles preventing the understanding of these interactions.Computational models of population dynamics and the dataset they are based upon need to take celltype-and circuit-specific neural activity information into account so that these models can start parsing which subpopulations of neurons are particularly important for specific movement parameters.To link neuronal activity to synaptic plasticity, new tools need to be developed that allow for the recording of synaptic changes and neuronal activity simultaneously within the same animal and same cells.Most state-of-the-art fluorescencebased sensors used for imaging neuronal activity, such as the GCaMP series of calcium sensors, or for imaging synaptic structure and function-such as the Thy1-YFP mouse, 77 fluorescently tagged receptors (e.g., SEP-GluA1), 69 and glutamate sen-sors 87 -occupy a similar spectral color range, thereby limiting their simultaneous use.Thus, future developments will need to focus on generating new high-performing fluorescence sensors in a variety of colors or enabling the combinatory use of different recording modalities, such as combining imaging with electrophysiological recordings.

Linking contributions and interactions of multiple distributed brain motor regions
Motor control and motor learning involve a set of distributed motor areas in the brain (Figure 1) that act in concert to generate movement.To date, most studies have been focused on understanding how neurons within individual brain regions encode movement and adapt during motor learning, again mostly due to technical limitations in recording from distributed brain regions.New approaches, such as mesoscopic imaging, 176 multi-site two-photon imaging, 177 or electrophysiological recording techniques, 178,179 are now enabling the simultaneous co-recording of neuronal activity in multiple brain motor regions.Given the inevitable large amount of data generated by such studies, new theoretical models for making sense of multiregional interactions will be needed as well. 179 particularly important question that needs to be addressed is what specific circuit connections undergo plasticity during motor learning and, molecularly, an understanding of postsynaptic intracellular signaling mechanisms regulating spatiotemporal aspects of plasticity, which will require further studies linking synaptic plasticity and neuronal activity. 65,67,87Such studies will provide novel insights into mechanisms in motor diseases and allow for the identification of possible new treatment approaches.
Several additional factors, such as neuromodulation and nonneuronal brain cells, can further regulate the interaction between different brain motor circuits.Many neuromodulators, such as acetylcholine, noradrenaline, dopamine, and serotonin, are critically involved in synaptic plasticity, 180 play a role in motor learning, and can be disrupted in movement disorders.Thus, it is important to understand how such neuromodulatory signaling can affect neuronal circuit interactions and neuronal plasticity during motor learning.Lastly, non-neuronal cells also play an important role in regulating motor circuit interactions.For example, microglia and astrocytes are shown to play an active role in modulating synaptic plasticity. 181In addition, oligodendrocytes myelinate axonal projections, regulating their signal conductance, and have recently been shown to be critical for motor learning. 54Interestingly, plasticity in brain myelination patterns has also been observed during motor learning, 68 which potentially has a significant impact on neuronal circuit function. 182Thus, studying how glial cells regulate neuronal circuits and their interaction will be critical for understanding motor learning.
Over the past decades, the field has made significant progress in understanding how the brain controls movement and supports the acquisition of new motor skills.With the continued development of new tools and concepts to tackle these questions, the next years will undoubtedly provide further insights into brain motor functions and offer a more comprehensive understanding of circuit and synaptic functions and interactions underlying motor control and learning.

Figure 3 .
Figure 3. Common behavioral tasks used to study circuit and synaptic plasticity during motor learning in rodents (A) Schematic of mouse performing freely moving forelimb reaching task.(B) Mice learn to perform successful reaches over the course of 8 days of training and maintain proficiency for >30 days.(C) Reach trajectory kinematics analysis.Yellow, 30 individual reach trajectories; red, average reach trajectory.(D) Variance of reaching trajectories decreases over time of learning.(B-D adapted from Albarran et al. 62 ).(E) Schematic of mice performing head-fixed lever pushing task.(F) Mice learn to perform successful pushes over the course of training.Grey, individual mouse succes rate; black, average succes rate.(G) Example traces of the lever-pushing movement in one mouse on days 1 and 12 of training, showing the development of stereotyped movements.Gray, individual lever trajectories; black, average trajectory.(H) Pairwise correlation of lever-pushing trajectories for all trial pairs between adjacent days increases over the course of learning.Grey, individual mouse data; black, average data.(E-H adapted from Sheng et al. 84 ).(I) Schematic of head-fixed treadmill running in mice.(J) Mice increase running velocity over course of motor learning.Mice with and without a gradient refractive index (GRIN) lens implantation exhibit comparable running adaptation.(I and J adapted from Ma et al. 59 ).(K) Example gait patterns of mice running on treadmill.(L) Running gait patterns become more structured after training.(K and L adapted from Adler et al. 57 ).Error bars represent SEM.

Figure 4 .
Figure 4. Circuit adaptations in motor cortex and striatum (A) Schematic of mouse performing head-fixed lever pushing task and task structure.(B) Activity onsets of excitatory neurons in L2/3 of motor cortex become more refined and shifted toward the beginning of movements over the course of motor learning.(C) Correlation between the neuronal activity and the learned activity pattern increases with increasing correlation between trial movement and the learned movement pattern in expert sessions but not in naive mice.(D) A stronger correlation between population activity and movement emerges during learning.(A-D adapted from Peters et al. 81 ).(E) Average neuronal activity of dSPNs in the striatum become more refined and shifted toward the beginning of movements over the course of motor learning.(F) Average neuronal activity of iSPNs in the striatum become more refined and shifted toward the beginning of movements.(G and H) Movements with similar trajectories have higher-correlated striatal activity following motor learning.(E-H adapted from Sheng et al. 84 ).Error bars represent SEM.**p < 0.01, ***p < 0.001.

Figure 5 .
Figure 5. Dendritic spine plasticity in motor cortex (A) Schematic of mouse performing forelimb reaching task.(B) Longitudinal imaging of the same apical dendritic branches of motor cortex layer 5 neurons in Thy1-YFP mice over 1-day intervals reveals spine elimination (arrows), formation (arrowheads), and filopodia (asterisks) in trained and control mice.Scale bar, 2 mm (A and B adapted from Xu et al. 77 ).(C) Motor learning results in a transient increase in spine density.(D) Motor learning induces an early increase in new spine additions (top) and delayed increase in spine elimination (bottom).(E) Motor learning stabilizes newly formed spines.(F)Schematic summary of spine dynamics and plasticity during motor learning.During early phases of learning new spines are formed and existing synapses are strengthened.In late phases, initially potentiated synapses remain potentiated and spines formed in spatial clusters are selectively stabilized.64,77,69

Figure 6 .
Figure 6.Common behavioral tasks used to study motor learning in primates (A) Schematic of monkey performing a center-out joystick reaching task (left).Reaching trajectories of initial and late sessions with clockwise and counterclockwise curl force fields applied.Monkeys learn to adapt their trajectories to with applied force (right).(A adapted from Perich et al.133 ).(B) Schematic of human performing a joystick reaching task (left).A divergent force field is applied such that any deviation from a straight reaching trajectory will generate a force that acts in the same direction.Humans learn to make straight movements in this task setting (right) (B adapted from Davidson and Wolpert134 ).(C) Motor cortex recordings are used to control neuroprosthetic devices, such as moving a cursor on a screen in a center-out movement trajectory (C adapted from Wilson et al.135 ).

Figure 7 .
Figure 7. Relationship between neural population dynamics and cortico-striatal plasticity (A) Neural state space representing the spiking activity from three neurons across time (colors) (A adapted from Vyas et al. 129 ).Such neural population dynamics can encode movement parameters.(B) During motor learning, neurons in the motor cortex form refined temporal activity sequences 81 that are likely mediated by synaptic plasticity mechanisms, such as spike-time-dependent plasticity (STDP-LTP and STDP-LTD), in the cortex.(C) Cortico-striatal projections also undergo synaptic plasticity with likely postsynaptic spatial and temporal clustering, 65 which can have a supra-linear effect on neuronal activity through local dendritic computations.

Table 1 .
Behavioral tasks used to study circuit and synaptic plasticity during motor learning in rodents