Motor learning mechanisms

[rcoreilly]

Motor Learning Mechanisms

This discussion started with some initial notes and ideas in the first few entries, and then developed into a full-scale lit review, starting at: #294 (comment).

Original starting idea

Per Tom Hazy's recent discussion of dorsal-lateral BG DA signals.

Hypothesis: There are mechanisms for localized phasic DA learning signals in dorsal striatum to train BG action gating under the guidance of an overall maintained Goal / Plan (plan in dlPFC / ALM). To the extent that an action selected in SMA / M1 is unexpected relative to the maintained plan (e.g., due to specific sensory / timing circumstances as the plan unfolds), it could drive phasic DA to update the BG gating next time around.

• BG (matrisomal) gating of action directly disinhibits SNc DA, driving phasic DA bursts at action / plan onset. Evidence of such bursts in MarkowitzGillisJayEtAI23 -- lots of ongoing phasic DA during spontaneous behavior -- sometimes bursts vs. flat vs. dips..

• paired striosomes can shunt these DA bursts (? BrownSmithGoldbloom98) -- also classic Joel paper.

• RagsdaleGraybiel90 shows hierarchical pattern of connectivity into striosomes vs. matrisomes: "higher" (goal) areas project to lower area striosomes.

• Per Alan: KrokMalteseMistryEtAl23 show that ACh tends to precede DA -- consistent with ACh disinhibiting gating followed by DA reflecting above evaluation of gating.

[rcoreilly]

Notes from Tom's slides:

• BrownSmithGoldbloom98 - Organizing principles of cortical integration in the rat neostriatum: Corticostriate map of the body surface is an ordered lattice of curved laminae and radial points
• DhawaleWollffKoEtAI21 - The basal ganglia control the detailed kinematics of learned motor skills
• EngelhardFinkelsteinCoxEtAI19 - Specialized coding of sensory, motor and cognitive variables in VTA dopamine neurons
• FisherRobertsonBlackEtAI17 - Reinforcement determines the timing dependence of corticostriatal synaptic plasticity in vivo
• KhatibMorris23 - Dual role of dopamine in shaping spontaneity
• KincaidWilson96 - Corticostriatal innervation of the patch and matrix in the rat neostriatum
• MarkowitzGillisBeronEtAI18 - The striatum organizes 3D behavior via moment-to-moment action selection
• MarkowitzGillisJayEtAI23 - Spontaneous behaviour is structured by reinforcement without explicit reward
• RagsdaleGraybiel90 - A simple ordering of neocortical areas established by the compartmental organization of their striatal projections
• ReinerHartLeiEtAI10 - Corticostriatal projection neurons - dichotomous types and dichotomous functions
• Watabe-UchidaZhuOgawaEtAI12 - Whole-brain mapping of direct inputs to midbrain dopamine neurons.
• YagishitaHayashi-TakagiEllis-DaviesEtAI14 - A critical time window for dopamine actions on the structural plasticity of dendritic spines

[thazy]

Re: KrokMalteseMistryEtAl23 and the timing of phasic and ACh signals:

• phasic ACh signals are tricky. It seems best to assign the pause as the primary signal, which is only variably preceded by a brief increase in firing, and/or followed by a rebound burst (more on all this at S&T next Wed)
• Re: Krok et al. specifically, and focusing on the ACh pauses, here is a quote:

"During rewards, DA and ACh were negatively correlated (Pearson’s cross-correlation coef- ficient r = –0.48 ± 0.02; n = 13 mice) with a temporal lag of 123 ± 9 ms (Fig. 2a–c), indicating that peaks in DA slightly precede dips in ACh. Notably, DA and ACh showed a similar negative correlation outside of reward, whether mice were actively locomoting or remained immo- bile (Fig. 2a–c and Extended Data Figs. 4h and 5b–d)." (p. 3)

• And, in looking at Figure 1b, third graph (immobility), the intrinsic ~delta-scale oscillations show increasing ACh signals slightly-but-consistently leading the increases in DA signals.
• Similarly, for the Reward case (Fig 1b, 1st graph), it shows the large phaDA bump clearly preceding the sustained ACh trough.

[rcoreilly]

Combining BG Trace and Local BG Critic learning mechanisms

BG Trace learning

The usual delayed reward outcome problem looms over the motor sequence learning space: final up / down DA signal only comes after a sequence of actions over time has resulted in either getting the reward or giving up / hitting a wall etc.

The trace learning mechanism in MatrixPrjn is designed to handle this problem, with the computationally-effective "non-discounted credit assignment logic" associated with LSTM / GRU and the ACT-R production system architecture:

• Every action taken during a "goal pursuit" sequence gets modulated up / down by the final RPE DA signal -- if success (better than expected), all actions increase in probability, and vice-versa for failure. This is a "block" or "square wave" credit assignment function.

A key advantage of the BOA goal engagement mechanisms is that it provides a "biochemically explicit" definition of this goal-engaged window: from the onset of the VS goal engagement gating to the final VS goal-clearing gating.

• TODO: The current BG trace window is defined in terms of ACh pauses^* and it is plausible that the CIN (TAN) cells are playing this role, but they need to have some further modulation in terms of actual VS gating, because raw "LDT" ACh is based only on SC novelty or US input -- but SC novelty also happens all the time without VS gating so ACh itself is imperfect.. Add actual CIN units, and also look into those modulatory midline / ventral thalamic layers that provide a mechanism for VS gating to modulate DS in the necessary way.

• ^* TODO: fix ACh: should just go ahead and invert this in the models so it is clearer -- ACh in the model is disinhibitory but ACh is modeled as a burst instead of a pause, so the sign is reversed to get it to work..

As in Kevin's reservoir-based models, we also need significant input from the M1 and ALM PTp layers, which have dynamic unfolding activation states, to drive novel action sequences etc.

Local DS critic mechanism

Per the initial message in this discussion (OP), we can now consider how the "local critic" mechanism can interact with the overall BG trace credit assignment.

• Computationally, the main contribution is to focus this "block" up / down credit assignment on the novel steps in a sequence that were unexpected relative to the active plan. This could be quite important for improving learning efficiency because the "bock" form of credit assignment otherwise isn't able to isolate specific key steps in a sequence of actions. With some amount of random action sampling, new actions are tried and if they lead to better or worse outcomes, they will get the full credit assignment learning from the actual outcome.

Test Case

Need a simple focal test case ala PCore where we can explore these mechanisms! Something with a simple action space -- maybe just like the BOA task where you have to go L or R some number of times and then forward to get to a goal, then consume... Can vary length of sequence. Perceptual input indicates how far goal is L or R etc. Just fake the whole goal outer-loop on it so we dispense with all that machinery...

[rcoreilly]

Dorsal vs. Ventral Credit Assignment Issues

Per Jed: how can the system "intelligently" assign credit for the nature of the outcome to the ventral BOA goal system, so the dorsal BG action selection system isn't punished because the food was spoiled.

Because OFC has an explicit representation of the target outcome, it seems plausible that this outcome-specific difference can be "absorbed" there somehow.. Need to think about what mechanisms might make that happen. We do have plenty of options in terms of VTA vs. SNc dopamine, and lots of control from ventral striatum -> dorsal.

Conversely, a key credit assignment learning dynamic for dorsal is finding "shortcuts" that result in reduced overall effort / pain etc, producing a greater net effort-discounted DA value.

Thus, overall, dorsal DA should be focused on these effort-related differences, while ventral is focused on outcome-related differences.

[rcoreilly]

Specific task design

Basic BOA ArmMaze

• VSgpi and OFC Input layers provide phasic goal engage and maint signals respectively
• VSgpi resets DLS learning window at start / end
• Action sequence depends on OFC: if engaged (active), approach, consume, if not engaged, L or R.
• Distance contingency drives how many approach steps to take.

Advantages:

• directly applicable to existing BOA

Disadvantages

• how to extend difficulty in a natural way? too low dimensional action space?

Start with BOA and then progress from there.

[rcoreilly]

How does the DLS (dorsolateral striatum) contribute to motor control?

First, there is a substantial and growing literature showing that the DS and VS have very different properties (e.g., MohebiPettiboneHamidEtAl19), so everything we've learned about VS in context of BOA likely does not apply here.

Historically, there are two conflicting theories about the role of the BG in motor control:

• Action Selection: The BG essentially picks from a menu of possible actions, based on a history of bidirectional dopamine reinforcement learning (reinforcing good, punishing bad) that implements Thorndike's law of effect, and the actor / critic model where the BG is an actor that chooses actions. Mink96 articulated this "switchboard" model originally: BG suppresses bad choices and facilitates good ones, at the time of action initiation -- i.e., a discrete, phasic effect. We built on this idea in FrankLoughryOReilly01 and our PBWM model for PFC / BG gating interactions in working memory updating (OReillyFrank06). The other major implemented computational models due to Redgrave, Humphreys, Bogacz, et al, are also based on this model.

• Dynamic Modulation: The BG dynamically modulates motor action parameters in an ongoing, continuous way. DeLong et al (1984) is an early version of this idea, and current work appears to be favoring this model according to a recent review that nicely summarizes these two different theories (ParkCoddingtonDudman20). This framework is more consistent with the predictive learning model of corticothalamic interactions (OReillyRussinZolfagharEtAl21): thalamus provides a training signal into cortex, and whereas lower cortical areas are drivers for pulvinar, the BG is the driver for motor thalamus (RovoUlbertAcsady12).

Problems with Action Selection

There are a number of problems with this model, with some of the major conceptual issues being:

• There is no fixed menu: motor commands are very high dimensional, and recent studies show that BG is key for learning to activate a randomly selected subset of neurons using BCI training paradigms. This doesn't look so much like selection, but rather shaping cortical representations in some way.

• The menu is not simultaneously active for BG to select from: the simple switchboard model requires that cortex present multiple options to choose from -- it is unclear how cortex could generate multiple distinct options in parallel in this way. And a serial selection model seems very inefficient: actions have to happen in the moment -- you don't have time to iterate through several different options. This contrasts with the case for VS and goal-driven decision making, where extra time and consideration appears to be devoted at the start, given that it will occupy the animal for a more significant amount of time.

• If the BG is key for action selection, it must be active prior to initiation, and its activity causally-related to the onset of actions. The evidence on balance does not appear to be consistent with this, per ParkCoddingtonDudman20.

Key "design" questions

• Is the BG performing online shaping of motor actions in the moment, or is it more important for shaping learning so subsequent actions are better? This is perhaps the most important distinction: the action selection model is strongly oriented toward an online "actor" role, whereas the dynamic modulation role can shade almost entirely into a learning role. The predictive learning model is entirely learning based, for one precedent example.

• How much does the BG's contribution (to online control and / or learning) come from descending direct motor projections, vs. the loops back through thalamus up into motor cortex itself? One idea is that BG exerts online control to descending subcortical motor areas, while primarily driving learning in cortex (GrillnerThompson21), which neatly addresses both of these first two questions.

• Much of the classic Go / NoGo Direct vs. Indirect pathway conception of BG function is tied up with the action selection paradigm: there are lots of additional pathways in the circuitry, and recent comprehensive anatomical mapping of the full loop pathways (FosterBarryKorobkovaEtAl21) -- how do we understand all this circuitry and DA modulation effects in the context of the dynamic modulation framework? KlausAlvesdaSilvaCosta19 address this question.

• Biologically, how precise vs broadly modulatory are the projections back from BG via thalamus to cortex? Is it plausible that these projections could shape neural activity in motor cortex in a fine-grained, high-dimensional manner, or is it more likely that it is only providing broad modulatory signals that are better suited for shaping learning, or a global threshold-like modulation?

• What kind of learning takes place within the DLS and associated BG nuclei, with respect to what kinds of DA and other signals, over what time frame of delayed credit assignment? And how does it influence learning within the cortex?

Overall, a first pass take is that the GrillnerThompson21 model sounds promising, within a broader view of the DLS as a kind of editor of motor signals: it edits descending signals (inhibiting and disinhibiting the original signals to tune the motor commands), and these edits shape learning in the cortex to improve the quality of the signals in the first place. Overall, such a functional role is very similar to the widely-accepted understanding of cerebellar function, raising the important question as to how these two systems might differ? A simple distinction is that cerebellum fixes finer-grained errors while DLS helps to learn broader time-scale sequencing. But it is very tempting to also say "action selection" here: perhaps the dynamic modulation and learning influences end up shaping action selection over iterations of learning?

Sections below explore recent literature with above issues in mind.

[rcoreilly]

Synthesis of lit review

• VL (cerebellum, ascending DCN) drives learning of specific motor actions, (i.e., what specifically is done at any given point); VM (DBG) drives learning of motor action sequences (i.e., when to perform a given action).

  • VL is focal, phasic, conveys detailed info about specific actions (Primary thalamus PhillipsSchulmannHaraEtAl19), while VM modulates all of motor cortex broadly (FosterBarryKorobkovaEtAl21, KuramotoOhnoFurutaEtAl15).
  • VM acts like a "metronome" or a "condition invariant signal" KaufmanSeelySussilloEtAl16 ZimnikChurchland21 that initiates motor action relative to a prior preparatory state, by gating layer 5 PT neurons via inputs on their layer 1 tufts, 150 msec prior overt action (in monkey). BG also has direct disinhibitory effects on subcortical motor areas, which are sufficient for most behavior in absence of cortex. This is consistent with CalderonVergutsFrank21 BG timing model (though not in terms of detailed properties).
  • VL "plus phase" efferent copy signals of actions may also drive learning of specific action reps in DLS -- it is still not clear if any ascending efferent copy signals hit VM (definitely hits VL). KuramotoFujiyamaNakamuraEtAl11 and RovoUlbertAcsady12 suggest no strong drivers -- maybe just a few weak ones according to former ref.

• PPN, PF (parafasicular intralaminar thalamic) is a major pathway for BOA goal system to modulate DLS: drives CINs in DLS that delineate goal-engaged window for credit assignment learning. PF recv from LDT and other BOA areas, and has closed-loop specificity for different categories of action (trunk, lower limb, upper limb, mouth, eyes) and targets CINs. PPN is driven primarily from SNr and provides key "someone gated" signal for credit assignment, projects to motor cortex too to modulate learning there.

  • Also direct VS -> DS SNr projections allow for direct modulation of output pathways: AokiSmithLiEtAl19 -- this can affect learning too indirectly, but probably the PPN, PF are more important for that.

• DBG's disinhibition of descending projections is sufficient for expression but not learning of detailed precise motor timing and sequencing KawaiMarkmanPoddarEtAl15. Still unclear how much BG key for expression -- TurnerDesmurget10 argue BG not needed for expression based on GPi lesions (but not necc. SNr).

• "Action selection" vs. "action sequencing" vs. "action timing" are all really kind of the same thing: the bottom line is that the DBG has direct disinhibitory gating control over when a motor action happens, via both descending and thalamocortical ascending projections. It seems like it has a very broad, gating-like impact in both directions (descending and ascending), and is thus much more about "when" than "what" -- so in that sense it is not actually doing "action selection" in the sense of selecting one among many actions that have been prepared. But perhaps if two potential actions are being considered and they are oscillating in strength, a well-timed "Go" signal will end up choosing one over the other.

• How exactly phasic DA drives learning in DBG which in turn drives learning in motor cortex remains somewhat unclear, under the above scenario. Most models use supervised learning as an admitted "cheat" in practice.

[rcoreilly]

Learning dynamics: efficient parallel search for sequence learning

First, the "action selection" framework suggests an overly-narrow "one action at a time" perspective on the general problem of learning entire sequences of actions that result in good outcomes. This is the full RL (reinforcement learning) problem, which is typically formulated in the following terms (SuttonBarto18):

• State $S$: representation of the environment and current body state, which can be discretized as $S_t$ at time $t$. This is very high dimensional in the real world / animal.
• Policy $\pi$: determines the action to perform, e.g., $A_t = \pi(S_t)$ returns the action $A_t$ as a function of current state.
• Reward $R$: is the externally-provided reward signal, which the agent is attempting to optimize.
• Value $V$: is introduced as a way of estimating the discounted expected future reward associated with either a State $V(S)$ or a State x Action pair $Q(S,A)$ -- i.e., Q learning. One can then determine the policy as a function of these values, e.g., by using a softmax function over values of possible next states or actions per state. This makes the policy an implicit function of Value, thus simplifying the problem conceptually. Value estimates can be updated by the TD (temporal difference) algorithm that iteratively updates the value of a state in terms of the value of the states that it can lead to.

The DLS receives significant state-like sensory input across multiple modalities, including somatosensory, visual, etc, and individual MSN neurons clearly represent different actions that could be taken in a given state. Thus, the basic $(S,A)$ framework is reasonable (specifically, $S \rightarrow A$). The massive DA modulation of DLS provides the R signal, and a history of DA-modulated learning could reasonably be construed as learning a $Q(S,A)$ map.

This formulation puts all the onus on the State (and Action) representation, which is the key problem that must be solved. For example, if you have a standard simple table-lookup ("localist") state representation of locations in a classic grid world, the resulting $V(S)$ mapping doesn't generally make sense, because locations per se do not have reliable values in any kind of dynamic real-world environment. Also, $Q(S,A)$ does not directly allow arbitrary sequences of actions to be learned unless each action results in a distinct state, and in general the "outer product" nature of a function of $(S x A)$ tends to obscure potential low-dimensional, systematic sub-spaces within that very high dimensional space.

The use of function approximation to learn the $Q(S,A)$ mapping, e.g., deep Q learning (DQL) enables more efficient and abstract functions to be learned, but this does not automatically overcome limitations in the raw S and A representations. Also, the DLS is just a single large, flat layer: all of the state abstraction has to be performed in the cortical inputs, not in the DLS itself. However, the major advantage of DQL is that it enables backpropagated credit assignment to effectively update the estimated value of many states and actions in parallel based on a single learning event, by virtue of shared internal hidden representations activated by many $(S x A)$ inputs. By contrast, the standard tabular Q learning only updates one discrete S,A cell at a time, which is horribly inefficient.

Parallel graded search of high-dimensional spaces is essential

The power of backpropagation neural networks derives from the ability to do massively parallel search through a high-dimensional problem space, with graded credit assignment applying learning with broad brushstrokes across wide swaths of the high-dimensional space. This requires individual neurons (across multiple deeply-stacked layers) that respond in a graded, randomly-mixed way to large subsets of the high-dimensional space.

This graded, parallel search contrasts with any kind of discrete, localized search process that can only pick its way slowly down singular paths.

By way of analogy, the Mastermind game (or Wordle or other similar games) likewise depends on performing a much broader credit-assignment updating process based on feedback from each round of guessing. If you instead did simple sequential trial-and-error guessing without considering the implications for the broader space of options, it would take thousands of guesses to randomly happen upon the right answer. Interestingly, these games abide by the inexorable constraint that the actual behavior of a single agent is always sequential and unitary (one action at a time) -- despite this strong constraint on behavior, learning can still be much smarter!

Thus, the key question at hand is how to get as much of this graded, parallel search within the context of the biological structure of the DLS? Is there some kind of special organization and learning dynamic here that might even be better than DQL? The key is to create an optimized State and Action representational space:

• Each MSN encodes a (parameterized) Action output, with many such neurons redundantly coding for each Action, with different overlapping parameterizations for strength or other relevant continuous variables (i.e., a coarse-coded distributed representation of action space).

• State must be systematically factorized, with many state input neurons active at any time, each projecting broadly and randomly to the MSN action neurons, such that each MSN coding for a given action samples a different subset of the input state. Thus, the $(S x A)$ space is densely, randomly sampled, and each MSN action neuron can learn in parallel about the relative value of actions in different regions of the state space.

• The actual behavioral output is determined by a stochastic choice driven by current MSN activity votes. As learning progresses, this should get better.

• Critically, "paths not taken" can still be learned about as possible alternatives to paths actually taken: if the current action sequence results in a sub-optimal outcome, then other possible paths can be up-regulated as future options. This credit assignment to alternate possibilities makes learning much more efficient and parallelizable.

• In general, egocentric State representations have direct mappings onto actions, and therefore greatly simplify and systematize the $S \rightarrow A$ mapping. If something is directly in front, and you want to approach, the "move forward" action is appropriate, etc. By contrast, allocentric States lack such direct affordances. This comports with the dominant role of egocentric representations throughout the sensory cortex, across species. Nevertheless, a variety of different reference frames and different ways of encoding stimuli is also valuable, to ensure full sampling of the relevant possible contingencies.

• The somatosensory State input implicitly encodes prior actions, by virtue of their impact on muscle spindle fibers, enabling sequences of actions to be encoded directly through the $S \rightarrow A$ mapping. It is also likely that dynamic temporally-evolving state representations across the motor cortex, at different levels in the anterior-posterior gradient, encode relevant state information for learning action sequences across different duration scales.

• In effect, MSNs can effectively encode chunks of actions as random subsets of small n-state transitions (n grams) -- it is likely that sparse random sampling can cover the space reasonably efficiently, based on similar such results in other domains. With reasonable variation in small n and overlapping points of engagement, arbitrary long sequences should be efficiently encoded (could do some math around that). Similar principles apply to PCR reconstruction of genetic sequences, etc. The chunks include sensory and motor sequences (not just motor-motor), and are determined by weights into individual MSNs.

• Requires learning to be modulated by the relationship between actual action taken and chunk's individual vote: if your action was taken, learning follows outcome (RPE); if your action was not taken, you go in opposite direction from outcome (negative outcome increases chances for alternative action plans). This is what the PF feedback can provide: which action actually taken. This algorithm is already in use in BG (pcore) model, but not optimized for these sequence chunks etc.

• Use trace learning: outcome comes at some later point after potentially long sequence, drives learning across all chunks within sequence.

• This is like any kind of basis decomposition (e.g., Fourier) of a function: differentially weight the contribution of lots of simpler basis elements to compose a more complex overall function. Solving for the basis elements in parallel is much faster (is that what the FFT algorithm does in effect: factorizes the problem into smaller pieces and then reassembles the whole much more efficiently).

• It is also possible that reservoir models implement much the same mechanism: the reservoir needs to be a decent simpler basis space, and the decoder learns to weight the basis elements appropriately.

• Striatum is not doing the WTA: it must maintain the full parallel set of chunks, but also needs to get the feedback signal about what actually happened. Instead, WTA happens in downstream funnel into SNr and then sends selection back out to striatum (via PF).

• MSNs need to "act out the sequence chunks" they represent, and also vote for the relevant action at each step (in this sense, at each step, they are actually doing action selection) -- thus the random sampling of MSN profiles needs to include significant temporal dynamics -- keyed by gating event tempo (per ACDC model), but able to distinguish different orders of events etc. SI input probably a major factor here -- encodes effect of last action etc. Also STN, Arkypallidal, etc maybe useful? Some kind of discretized sine wave oscillation vs. something that just stays static...

• Basic activity * DA 3-factor model well supported by YttriDudman16 and other such studies, with opto specific manipulations. Key thing is sign reversal for gating -- does YD16 actually implicitly show that with slow vs. fast selection??

[rcoreilly]

ParkCoddingtonDudman20

• Park, J., Coddington, L. T., & Dudman, J. T. (2020). Basal ganglia circuits for action specification. Annual Review of Neuroscience, 43, 485–507. https://doi.org/10.1146/annurev-neuro-070918-050452

Overall nice review of action selection vs dynamic modulation frameworks, supporting the latter.

Evidence that BG encodes continuous movement parameters:

Activity of dMSNs measured with fiber photometry displayed a significant correlation with three-dimensional (3D) velocity of movement, while the activity of iMSNs was anticorrelated, thus the difference in activity between the pathways was a better predictor of 3D ve- locity than either pathway alone (Markowitz et al. 2018, Meng et al. 2018, Yttri & Dudman 2016). This observation, that distinct striatal ensembles are recruited as a function of movement direc- tion or specific behavioral repertoire, has often been considered to provide evidence that the BG implement a discrete, switchboard-like selection of action (Alexander & Crutcher 1990b, Frank 2011, Graybiel 1998, Humphries et al. 2006, Klaus et al. 2019, Markowitz et al. 2018, Surmeier et al. 2009).

Action selection models have successfully explained abstractions of behavior such as the choice between two levers to press (e.g., see discussion in Collins & Frank 2014). However, the presence of action-specific ensembles of activity in BG does not necessarily imply a switchboard-like se- lection function.

An interpretation that BG were primarily involved in controlling movement execution was in- formed by the fact that inactivation or lesion experiments done around this period failed to pro- duce any effect on initiation of an action (putatively selection) and generally only resulted in execution deficits (Horak & Anderson 1984a,b; Mink & Thach 1991a).

One additional persuasive observa- tion at that time appears to have been the fact that a substantial majority of BG output neurons are modulated primarily or only during movement execution (Alexander 1987, Alexander & Crutcher 1990c, Anderson & Horak 1985, Hikosaka & Wurtz 1985, Liles 1985, Neafsey et al. 1978, Turner & Anderson 1997), especially when contrasted with cortical areas (Alexander & Crutcher 1990a).

However, many authors have tended to view initiation deficits in Parkinson’s disease as primary and deficits in continuous con- trol of movement kinematics as secondary (Albin et al. 1989, Redgrave et al. 2010), while others have argued the converse (Desmurget & Turner 2010, Dudman & Krakauer 2016, Kornhuber 1971).

Moreover, while kinematic deficits have frequently been observed in the absence of initiation deficits, the converse dissociation has not, to our knowledge (Yttri & Dudman 2018), ever been demonstrated. For such reasons, here we focus on the role of BG in shaping voluntary action through their role in the continuous specification of movement execution.

For discrete movement of a limb in mice, it was observed that some striatal cell activity mirrors the acceleration and deceleration of the limb with transient activity at the early phase of movement initiation or at the time of peak amplitude (Panigrahi et al. 2015) [as expected from control models (Baraduc et al. 2013)]. Similarly, subsets of cells in dorsal striatum also preferentially reflect the accelera- tion and deceleration phases of locomotion in mice running on a treadmill (Sales-Carbonell et al. 2018). Other subsets of neurons appear to encode other aspects of kinematics, reflecting more delayed components corresponding to amplitude or possibly sensory-mediated feedback during movement (Coffey et al. 2016, Panigrahi et al. 2015, Sales-Carbonell et al. 2018, West et al. 1990).

There are a few key features of these formulations that span some four decades. First, activity in BG is related to kinematics without reflecting the involved muscle forces (kinetics) per se. Second, the representation must be action specific in the sense that at least partially distinct ensembles of BG neurons represent kinematics of actions implemented with distinct muscle effec- tors. As reviewed above, there are now substantial experimental data from diverse measurement techniques to support both of these key features in multiple vertebrate species.

Thus, contrary to a recent proposal (Klaus et al. 2017, 2019), the observation of action-specific ensembles of neurons does not foreclose that BG populations regulate continuous kinematics of action; indeed, quite the opposite. As is clear from the analogy to a graphic equalizer (Figure 1a), the fact that a given channel exhibits distinct waveforms does not imply that it is nec- essarily selecting the particular music as opposed to, for instance, selectively amplifying distinct components of the source signal.

Population recordings in vivo are consistent with feedback gain control (Brown et al. 2014), and reducing intranigral feedback via chemogenetic perturbation significantly increased variance of movement kinematics without alteration of the mean (Brown et al. 2016), consistent with a role in regulating the gain of BG output rather than action selection.

In summary, the evidence reviewed here indicates that inhibitory connectivity within the BG does not enforce switchboard-like competition between action-specific ensembles. Instead, action specificity of striatal ensembles is predominantly inherited from afferents rather than first emerg- ing through local computations in the BG (Figure 2). This is consistent with (a) observations that BG action encoding follows, rather than leads, cortical (Park et al. 2019, Seo et al. 2012) and tha- lamic (Schwab et al. 2019) encoding of movement kinematics; (b) the preferential disruptions to movement kinematics following BG perturbation (Dudman & Krakauer 2016); and (c) conceptual points about selective modulation of kinematics reviewed above (Figures 1 and 2).

In comparison to categorical selection, continuous modulation can seem overly subtle: Can it really account for proposed BG roles in action control and adaptation? ... large gain modulations can have the practical effect of turning a given equalizer channel on or off.

Meanwhile, most BG output projects via descending pathways to midbrain and brainstem premotor areas, which also receive corticofugal input from deep-layer cortical neurons (Dudman & Gerfen 2015). Thus, the descending projection forms a convergent, feedforward pathway that targets premotor nuclei of the extrapyramidal motor system (tectum, red nucleus, pontine reticular areas) and integrates with descending corticofugal output that targets these same nuclei (Dudman & Krakauer 2016, Yttri & Dudman 2018). A smaller but substantial projection from BG output nuclei targets higher-order thalamic nuclei (Phillips et al. 2019), including the ventroanterior and ventromedial aspects of motor thalamus, as well as intralaminar nuclei (Alexander et al. 1986). These pathways provide indirect feedback to neocortex and are thought to modulate cortical output (and thus also BG inputs) (Figure 3).

These observations led to proposals that SNr activity switched between basal high firing rates and complete silence to function as a categorical gate for movement (Chevalier & Deniau 1990; but see Basso & Sommer 2011, Bayer et al. 2004). However, subsequent work has found high variability in individual SNr neuron responses to even similarly targeted saccades, including some excitation (Handel & Glimcher 1999, Sato & Hikosaka 2002, Shin & Sommer 2010). Even in subpopulations of SNr neurons inhibited around saccades, average responses appear as graded decreases in activity rather than pauses (Basso & Wurtz 2002).

We propose that BG output may thus regulate relatively open-loop components of motor control (e.g., movement vigor), whereas cerebellar output may regulate closed-loop components of motor control (e.g., end point accuracy) that are unaffected by disruption of BG function (Dudman & Krakauer 2016).

Thus, exogenous or aberrant synchronization of BG output can entrain corticothalamic neu- rons. However, physiological function may not involve such overt control. Rather, neurons in the BG output nuclei and BG-recipient motor thalamus respond as though they receive common, partially temporally overlapping inputs (Nakamura et al. 2012). Over separate recordings of the same monkeys making controlled reaches, movement-related activity appeared earlier with re- spect to movement in the motor thalamus than it did in BG output neurons in the GPi (Anderson & Turner 1991a,b). Furthermore, pharmacological inactivation of BG output had no effect on phasic movement-related activity in the motor thalamus (Inase et al. 1996). Experiments in avian (Goldberg & Fee 2012) and rodent (Guo et al. 2017) models have indicated that corticothalamic inputs in fact drive the majority of movement-related activity in BG-recipient motor thalamus. Inhibitory afferents from the BG may play more of a modulatory role, shaping responses to ex- citation rather than gating them or driving disinhibitory bursting (Goldberg et al. 2012, 2013).

Thus, contrary to some predictions from anatomy (Alexander et al. 1986) and exogeneous manipulations, physiological BG output may not gate motor thalamus activity, which helps to explain some perplexing observations from lesion studies (DeLong & Wichmann 2010). The de- tailed timing of integration will likely be critical to understand the modulation of thalamic nuclei by BG output.

Motor thalamus can be divided into BG-recipient and cerebellar-recipient zones distinguished by differences in molecularly defined cell types in recipient areas (Kuramoto et al. 2011, Phillips et al. 2019), and this division is clearly reflected in the structure of ascending projections in single- cell reconstruction studies (Kuramoto et al. 2009, 2015; Phillips et al. 2019). Cerebellum is an interesting complement to BG (Bostan & Strick 2010, Kornhuber 1971), sending excitatory output projections to many of the same regions that are inhibited by BG output, including the motor thalamus (Anderson & Turner 1991a, Phillips et al. 2019), SC (May 2006), and midbrain locomotor region (Semba & Fibiger 1992). Within multiple motor-associated areas of cortex, a majority of BG-recipient TC axons terminate in layer 1 [>50% from the ventral anterior and ventral lateral nuclei (Kuramoto et al. 2009), >70% from the ventromedial nucleus (Kuramoto et al. 2015)], while cerebellar-recipient TC axons terminate in deeper layers (Phillips et al. 2019).

Motor thalamus may thus be separated into first-order, cerebellar-recipient neurons that strongly drive cortical activity and are modulated by layer 6 corticothalamic feedback, and higher-order, BG-recipient neurons that are strongly driven by layer 5 cortical input and return modulatory feedback to the cortex (Phillips et al. 2019, Sherman 2016).

Papers to follow up on:

• Phillips et al. 2019, Kuramoto et al. 2015, Dudman & Krakauer 2016, Klaus et al. 2017, 2019

[rcoreilly]

KlausAlvesdaSilvaCosta19

Klaus, A., Alves da Silva, J., & Costa, R. M. (2019). What, if, and when to move: Basal ganglia circuits and self-paced action initiation. Annual Review of Neuroscience, 42, 459–483. https://doi.org/10.1146/annurev-neuro-072116-031033

Existing models of basal ganglia motor function have been strongly influenced by attributing movement-facilitating and movement-suppressing roles to striatonigral and striatopallidal pro- jections, respectively (Albin et al. 1995). However, recent experimental findings challenge this prokinetic versus antikinetic dichotomy (Cui et al. 2013; Jin et al. 2014; Tecuapetla et al. 2014, 2016; Yttri & Dudman 2016) and suggest that more detailed models are required that consider, for example, ensemble-level descriptions of movement-related activity (Barbera et al. 2016, Klaus et al. 2017, Parker et al. 2018), their fine-scale temporal modulations (Markowitz et al. 2018), and how these ensembles jointly regulate downstream circuits and behavior (Alexander & Crutcher 1990, Turner & Desmurget 2010).

That is, striatal spiny projection neurons (SPNs) gather information about planned movements from the cortex, as well as sensory and other contextual information, and translate it into specific changes in the basal ganglia outputs. Accordingly, many striatal neurons increase activity at movement onset ( Jog et al. 1999, Hikosaka et al. 2000, Jin & Costa 2010, Cui et al. 2013, Jin et al. 2014) with a substantial fraction of SPNs being selectively active during self-initiated movements (Schultz & Romo 1992, Jin et al. 2014).

Thus, these studies show increased activity in D1- and D2-SPNs during movement, and importantly, they show very low activity in both populations during immobility, indicating that immobility is not associated with high activity in striatopallidal SPNs. Furthermore, they show that during normal movement the activity of both neuron types increases with a similar temporal profile ( Jin et al. 2014).

For example, certain models suggest a so-called support/suppress scenario in which the two pathways work in concert to facilitate the desired movements while simultaneously suppressing competing actions (Mink 1996, Hikosaka et al. 2000). While the classical rate model postulates that the inhibitory and disinhibitory effects of direct and indirect pathways are exerted on the same basal ganglia output neurons in SNr, this model suggests that coactive direct and indirect pathways exert their effect on different basal ganglia output neurons, with the direct pathway inhibiting specific neurons to release the desired movement, and the indirect pathway disinhibiting other neurons to suppress competing movements. Other models suggest that both pathways can be concomitantly active but that their relative activity levels and relative timing in- fluence activity in basal ganglia output nuclei (Figure 1c).

While the coactivation of both pathways might still be compatible with variations of the movement-facilitating versus movement-inhibiting views, an elegant experiment by Yttri & Dudman (2016) showed that the stimulation of either pathway can increase or decrease movement speed in subsequent trials depending on whether neurons during fast or slow movement trials were stimulated (Figure 1d). Furthermore, the optogenetic manipulation of either D1- or D2-SPN activity before movement increases the latency for action initiation (Tecuapetla et al. 2016). These results do not support the simple pro- versus antikinetic dichotomy of D1- and D2-SPN function, but they are compatible with the idea that D1-SPNs support the desired action whereas D2-SPNs suppress alternative actions. However, detailed measurements of the spatiotemporal activity of SPN ensembles during natural behavior demonstrate that neurons in both pathways are rather action-specific, and consequently, the patterns of activity of neurons in each pathway are differ- ent for each action (Klaus et al. 2017, Markowitz et al. 2018, Parker et al. 2018)

Consistently, the similarity between neuronal pat- terns is related to the similarity between the actions that are being performed (Figure 2c) rather than the similarity between the speeds at which different movements are performed (Klaus et al. 2017). This suggests that when an organism initiates a specific action, the indirect pathway is not a general suppressor of all other competing actions, but rather, it can have a more specific role (Tecuapetla et al. 2016, Klaus et al. 2017)

SPNs encode action value (Samejima et al. 2005), and pathway-specific manipulations and recordings suggest that D1- and D2-SPNs may relay relative positive and negative values, respectively (Tai et al. 2012, Shin et al. 2018, Nonomura et al. 2018). Furthermore, perturbations of D2-SPN activity during action initiation increase the likelihood of switching to a different action, whereas comparable changes in D1-SPN activity lead to a delay in initiation but not a switch in behavior (Tecuapetla et al. 2016).

Evidence for this hypothesis comes from experiments that recorded activity in the primary motor cortex while stimulating the two striatal pathways in a cued lever-pressing task (Oldenburg & Sabatini 2015). That is, the stimulation of D1-SPNs can increase cortical activity across all layers, predominantly in lever press–related neurons. In contrast, D2-SPN stimulation transiently excites neurons in more superficial layers (involved in thalamocortical and corticocortical processing) that show little or no lever press–related modulation of activity. These results suggest that there are still aspects to be uncovered in how direct and indirect projections connect to basal ganglia outputs and thalamocortical circuits (see, e.g., Saunders et al. 2015).

A characteristic feature of SPNs is their distinctively low firing rate even during cortical activation (Berke et al. 2004, Sippy et al. 2015). This is a result of inward-rectifying conductances (Nisenbaum & Wilson 1995) and inhibitory synaptic inputs that counterbalance glutamatergic excitation (Blackwell et al. 2003, Pidoux et al. 2011, Reig & Silberberg 2014)

Interestingly, in a two-arm bandit task, the optogenetic activation of dopamine neurons promoted the initiation of a trial when rats were not engaged but did not change the probability that the same action (left or right) was repeated in the next trial. However, transiently activating dopamine neurons after the rats had chosen right or left increased the probability of them repeating the same choice in the following trial (Hamid et al. 2016). This exemplifies how both learning and action initiation can be produced by fast and transient changes in dopamine activity.

This activity preceding movement initiation seems to have a causal role because brief optogenetic activation of SNc dopamine neurons (Barter et al. 2015, da Silva et al. 2018) or their terminals in the dorsal striatum (Howe & Dombeck 2016) promotes movement initiation, while the inhibition of SNc dopamine neurons decreases the probability of movement initiation in mice (da Silva et al. 2018) (Figure 3d). Furthermore, movements initiated after the brief optogenetic activation of SNc dopamine neurons are more vigorous, while movements initiated during optogenetic inhibition are less vigorous (da Silva et al. 2018). Importantly, precisely timed optogenetic manipulations of SNc dopamine neurons during movement did not change ongoing movements, perhaps indicating a more specific role for their activity for action transitions than previously acknowledged (da Silva et al. 2018).

SPNs are relatively hyperpolarized and require substantial glutamatergic inputs to fire, which suggests that groups of striatal neurons active during the same action receive common input (i.e., the same individual cortical neurons or connected cortical neurons project to coactive striatal neurons dur- ing the same action; a similar argument could be made for thalamic inputs). Thus, groups of cortical and striatal neurons may form preferentially connected modules that could be the basis for the action-specific patterns of activity (Figure 4b).

Interestingly, self-initiated actions are preceded by neuronal activities in the cortex and striatum by hundreds of milliseconds to several seconds (Romo et al. 1992, Schultz & Romo 1992, London et al. 2018). This implies that premovement activities can reverberate through cortex–basal ganglia–thalamocortical loops, which have propagation latencies of about 200– 250 ms (Oldenburg & Sabatini 2015, Klaus & Plenz 2016).

It is important to note that cortical inputs to the striatum can arrive from different popula- tions, for example, IT and PT neurons (Reiner et al. 2010, Shepherd 2013). While both types are active during movement planning, a specific motor command arises mainly in PT neurons before movement onset (Svoboda & Li 2018). Interestingly, it is not necessarily the average firing rate but rather the ensemble dynamics that have a profound influence on movement initiation (Kaufman et al. 2014). Thus, understanding IT and PT connectivity to downstream circuits (Hooks et al. 2018) and their dynamical organization in relation to basal ganglia circuits will be important avenues for future studies.

This is consistent with studies showing that cortical dynamics that prepare movement can emerge in the absence of movement execution (see also Kaufman et al. 2014) and also that the basal ganglia may be needed to initiate movements that are selected somewhere else (Thura & Cisek 2017).

It is important to note that D1- and D2-type receptors have different affinities, with the high- affinity D2 receptors likely being occupied at lower concentrations. Thus, transient increases in dopamine release might mostly recruit the lower-affinity D1 receptors, promoting the activation of D1-SPNs by coordinated cortical or thalamic inputs (Surmeier et al. 2007).

• Followup: Tecuapetla et al. 2016, Hamid et al, 2016; da Silva et al, 2018;

YttriDudman16

Yttri, E. A., & Dudman, J. T. (2016). Opponent and bidirectional control of movement velocity in the basal ganglia. Nature, 533(7603), Article 7603. https://doi.org/10.1038/nature17639

We first asked whether photostimulation of dMSNs during the fastest third of movements could alter the velocity of subsequent movements. Indeed, brief dMSN stimulation was sufficient to produce a signifi- cant increase in the peak velocity (1.4 cm s−1 increase from 29.7 cm s−1; P < 7 × 10−5; Fig. 2 and Extended Data Fig. 2) of all limb movements. Other movement parameters that were not targeted for closed-loop stimulation such as the amplitude, duration, and tortuosity remained unaltered (P > 0.7). This is despite the fact that mice were capable of rapidly adjusting movement parameters to changing reward contin- gencies (Extended Data Fig. 3). By contrast, iMSNs stimulation during the fastest third of limb movements produced a significant reduction in peak velocity (−1.1 cm s−1; P < 7 × 10−4). The effect of iMSN stim- ulation had its maximal effect on velocity; movement duration and tortuosity were not significantly altered (P > 0.3). Prolonged tonic acti- vation of dMSNs tends to be pro-kinetic in that it evokes generalized increases in voluntary movement (‘response vigor’15), whereas tonic activation of iMSNs tends to decrease voluntary movement11. However, we found that neither brief dMSN nor iMSN stimulation during the fastest movements produced a change in the rate of trial initiation or the rate of licking during reward anticipation and consumption (Fig. 2b and Extended Data Table 1). These results thus demonstrate that closed- loop activation of MSNs is sufficient to produce sustained changes in movement parameters without generalized changes in movement or motivation.

• This suggests a simple activity-dependent learning rule: if active, and reward happens later, do more (D1). Actual learning would rely on lots of natural variation and reinforcing in proportion to effects on outcomes. Results show dependent on DA.

Our results suggest that stimulation engages a dopamine-dependent, bidirectional plasticity in striatal MSNs. While a learning rule that acts directly on a parameter specifying the velocity distribution can account for our behavioural results, MeSH is abstract and it is unclear how it could be implemented in corticostriatal circuits critical for goal-di- rected, instrumental behaviour17. Thus, we implemented a simplified corticostriatal circuit model with the following key features (Extended Data Fig. 6).

There is little empirical evidence for representation of specific velocity ranges in cortical activity18. By contrast, there is substantial evidence that cortical19,20 and striatal21 representations of forelimb movements are monotonically tuned to speed. Consistent with the anatomy of the corticostriatal pathway1, we assume that the speed of a movement is determined by both cortical and basal ganglia output (Extended Data Fig. 6). In combination with monotonic tuning this implies that the mean movement velocity is proportional to the average weight of corticostriatal synapses (Methods).

Here we provide the first demonstration, to our knowledge, that the innate biases in the direct and indirect pathway to increase or decrease the frequency of movement, respectively, do not extend to fixed biases in the control of movement parameters. The direct and indirect path- ways engage opponent, activity-dependent plasticity mechanisms that can produce sustained biases in future behaviour. Each pathway is sufficient to produce bidirectional changes and, to some extent, is innervated by distinct cortical populations27, suggesting that bidirectional control by each pathway could allow for adaptive con- trol of goal-directed actions in different contexts or under different demands. These data argue that phasic activity in the striatum during specific movements is sufficient to selectively reinforce changes in a movement parameter independent of a generalized change in motiva- tion consistent with a role for dopamine-dependent signalling in dorsal striatum in the control of movement vigour21,26,28.

[rcoreilly]

ArberCosta22

Arber, S., & Costa, R. M. (2022). Networking brainstem and basal ganglia circuits for movement. Nature Reviews Neuroscience, 23(6), Article 6. https://doi.org/10.1038/s41583-022-00581-w

This is a great recent review covering the full scope of motor control in an integrated manner.

The brainstem is a large centre that spans the hind- brain and the midbrain and is involved in many physi- ological functions, including movement. The hindbrain is the most caudal extension of the brainstem immedi- ately upstream of the spinal cord, further divided into the pons and medulla (Fig. 1a). The hindbrain transitions rostrally into the midbrain, which encompasses many subregions, including the colliculi (Fig. 1a). To regulate body movements by engaging limbs through spinal motor neurons, the brainstem is conveniently located at the junction between higher-level motor centres and spi- nal circuits involved in their execution and sensory feed- back. The brainstem also engages in the control of other behaviours, such as orofacial movements (for example, whisking and licking) and coordination of head and eye movements, enacted through cranial motor neurons located within the brainstem proper1–3.

However, the spatial intermingling of brainstem neuronal populations involved in differ- ent functions has made it difficult to understand the roles of and information flow in specific brainstem circuits3–6.

• Anything subcortical is, generally, much more intermixed than the nice topography present in cortex. Presumably, genetic wiring processes operate on local chemical diversification, so having neighbors of a certain type causes you to become a different type, and that means that everything is intermixed.

These find- ings suggest that a behaviour perceived as one overall action is constructed by collaboration between differ- ent information streams carried by distinct neuronal populations that influence the exact manifestation of the chosen behaviour. Information from these distinct brainstem circuits is likely to be combined in the spi- nal cord, where local circuits and sensory feedback also contribute further processing towards execution.

They revealed a region in the mid- brain that fulfilled these criteria, since then referred to as the ‘mesencephalic locomotor region’ (MLR), which turned out to be evolutionarily highly conserved9.

Glutamatergic MLR neurons reside in three subdivisions, namely the pedun- culopontine nucleus (PPN), cuneiform nucleus (CnF) and mesencephalic reticular formation (mRT) (Fig. 1b).

In line with a role in the regulation of locomotion, only a proportion of MLR neurons encode locomotor state and/or speed10,11,13,15; however, others are preferentially recruited during distinct spontane- ous behaviours, including rearing, grooming or food handling15, suggesting a more diversified role of excit- atory MLR neurons in movement regulation.

• key: PPN / MLR not a homogenous system: seems to be important for licking behavior too..

Specifically, glutamatergic neurons in the lateral para- gigantocellular nucleus (LPGi), a caudal medulla sub- region, but not in neighbouring magnocellular nucleus

Together, these findings demonstrate that high-speed locomotion is brought about by excitatory interactions between the MLR subregion CnF and the medulla subregion LPGi, giving rise to a CnF–LPGi circuit needed for high-speed locomotion.

Still, the identity of an analogous brainstem circuit module for the implementation of exploratory locomotion, which occurs usually at lower speeds, is less clear at present (Fig. 1b).

Optogenetic stimulation of excitatory MLR or LPGi neurons has the striking behavioural property that one-sided stimulation elicits completely bilateralized, for- ward locomotion without directionality changes, unless natural barriers such as walls are encountered10–14,17. Therefore, unilateral locomotion signals are likely to be distributed symmetrically, leading to non-distinguishable motor programme execution of both body halves. Furthermore, the observations that natural barriers can impede locomotion even when locomotion is induced optogenetically suggest that either parallel circuits enact this additional information to avoid barriers or circuits downstream of the MLR–LPGi network (for example, in the spinal cord) can still integrate signals to influence behaviour. Thus, ongoing motor programmes can be stopped or ‘vetoed’ depending on the circumstances, even when movements are induced optogenetically and relatively close to motor execution.

• Lots of built-in support for avoiding walls etc.

A model suggests that turning towards the ipsilateral body side is achieved by unilateral activation of a stopping pro- gramme, while the unperturbed side continues to move, thereby resulting in orienting curvature towards the ipsilateral side during translocation.

Thus, an alternative model for changes in locomotion directionality could be that short-latency head turning combined with a following speed decrease allows body reorientation to align with head position.

CHX10-expressing Gi neurons, and more generally neu- rons in the reticular formation, receive synaptic input from deep-layer neurons of the contralateral superior colliculus18,19,23, which are a likely source for influencing orientation-regulating brainstem neurons (Fig. 1c). Of note, superior colliculus neurons expressing the tran- scription factor PITX2 are engaged during head turning in mice24. Input by midbrain sources that signal postural states may also have a role in this process15.

• deep SC is generalized motor control for orienting; SC is a "full service" sensory -> motor system capable of doing all manner of complex instinctive motor programs.

The lateral rostral medulla (latRM), first identified as a possible neuronal substrate for the regu- lation of forelimbs through its preferential connectivity to forelimb-innervating but not hindlimb-innervating motor neurons25, plays an important role in different forelimb movements, including reaching and food handling26. Notably, although forelimbs are used dur- ing locomotion, latRM neurons are recruited in a task-specific manner during forelimb behaviours but not during locomotion,

• Key point: these are the "knobs" that BG controls -- they are high-level, not muscle-level..

These popula- tions receive excitatory input from upstream structures, including the cortex and superior colliculus, probably contributing to their activation14,18,19,25. However, brain- stem neurons also receive direct and/or indirect output from the basal ganglia, most notably from GABAergic neurons of the substantia nigra pars reticulata (SNr)30,31,

• A classical model suggests that for selection and execution of specific movements, brain- stem populations need coincident activation from excit- atory inputs and disinhibition from basal ganglia output structures32.

• 32 = Hikosaka07. Nice simple story: Cortex, SC provide detailed parameters for action, and BG gating says when to do it, via broader disinhibition.

• In agreement with this model, execution of saccadic eye movements in a well-trained and rewarded context is accompanied by inhibition of globus palli- dus internus and SNr inputs to the superior colliculus33 (Fig. 3b). However, this model of SNr neurons providing tonic inhibition with a pausing during movement so as to disinhibit downstream brainstem populations has not been widely tested.

Strikingly, although many SNr neurons are indeed inhibited during movement, others are excited34–36.

Another possibility is that this phenomenon reflects a centre–surround circuit mechanism (Fig. 3c).

Indeed, at the population level, SNr neurons inhibited during loco- motion are predictive of locomotion initiation, whereas excited neurons are not35. Several conditions should be met for this model to be generally applicable to SNr out- put signalling. The first is that to convey action specific- ity, different SNr populations should project to distinct brainstem areas associated with the control of these movements, and not broadcast generally to all down- stream areas. The second is that inputs to SNr neurons should also exhibit specificity. For example, SNr neurons inhibited during a particular movement should get spe- cific inputs from the striatum related to the execution of that particular movement, but surrounding SNr neurons should get a different input and not be inhibited. Next, we describe experimental progress on both topics.

... suggest- ing that parallel SNr channels target different brainstem regions38. Within the SNr, these different populations are found in striking spatially organized patterns.

Whereas more anterior-lateral SNr areas project to different areas of the superior colliculus, more posterior-medial areas pro- ject to caudally located brainstem regions in the pons and medulla (Fig. 3d). SNr subpopulations also exhibit different electrophysiological properties, depending on the projec- tion target. More lateral SNr neurons are large, fire fast and have fast membrane time constants, whereas more-medial SNr neurons have slow firing rates and decay constants38.

• 38 = McElvainChenMooreEtAl21 -- summarized below.

These findings demonstrate that SNr subpopulations targeting distinct brainstem regions collateralize to specific thalamic areas. This is likely to lead to concomitant regulation of brain- stem and thalamocortical circuits involved in the same movements, thus making these output channels specific to brainstem targets and specific regions of the thalamus.

two recent studies found that the striatum connects to basal ganglia structures through a highly organized matrix using highly local- ized injection strategies43,44. Specifically, the dorsome- dial striatum (DMS), dorsolateral striatum (DLS) and ventrolateral striatum (VLS) target different SNr and GPe subregions, and through this specific projec- tion pattern, they maintain fine topography with respect to brainstem and thalamic areas43 (Fig. 3e).

• 43 = LeeWangSabatini20 (discussed below), 44 = FosterBarryKorobkovaEtAl21 (also below)

even in cases in which the dorsomedial striatum and DLS provide input to SNr neurons projecting to the same area of the superior colliculus, these populations tar- get different layers, suggesting functional differences in the targeted downstream network43.

Interestingly, segregated basal ganglia cir- cuits also form segregated closed loops with the thala- mus and cortex43,49, indicating that basal ganglia output pathways can modulate their own input and suggesting that specific movement-relevant information is propa- gated within wider networks and not just transmitted to brainstem centres for execution.

Recent work imaging hundreds of striatal projection neurons (SPNs) in the DLS while mice performed different body movements, including locomotion, directional changes and rearing, revealed that distinct neuronal ensembles were active during different movement patterns, and that single SPNs exhibit a high degree of movement specificity50 (Fig. 4a). Furthermore, neurons in the DLS encode detailed task-related movement kinematics51.

• 50 = KlausMartinsPaixaoEtAl17, 51 = DhawaleWolffKoEtAl21

Interestingly, both dSPNs and iSPNs in the DLS are active dur- ing movement56,57, and their activities are movement specific50,58. ... and show similar movement specificity 50.

However, the timing of their activation may be important because in vivo patch clamp membrane potential recordings from dSPNs versus iSPNs revealed that these neuron types have different timings of depolarization on a milli- second timescale59.

Moreover, recent studies suggested that fine topography between striatal regions and the GPe appears to be maintained, whereas more convergence seems to be evident in direct striatal input to the SNr44.

In many situations, it is important to simul- taneously select or disinhibit two or more movement components to perform an action, for example loco- moting and reorienting at the same time.

performing simultaneous recordings from the cortex and striatum reveal that striatal activity topographically reflects deep-layer cortical activity75.

On the basis of these studies, we posit that striatal action specificity reflects highly organized excitatory input into the striatum.

Mammals contain three DCN arranged bilaterally along the medi- olateral axis, with the medial DCN representing the evo- lutionarily oldest structure, followed by the interposed and lateral DCN positioned progressively more laterally (Fig. 5d). Additional subdivisions along the three ana- tomical axes are discernible within each of these regions, making anatomical and functional studies challenging.

For all DCN, 125 ipsilateral and 140 contralateral cen- tral nervous system targets were identified88 (Fig. 5d), whereas the medial DCN projects to 60 targets89. ... In addition, unlike the medial and interposed DCN, the evolutionarily younger lateral DCN does not provide input to the spinal cord88,89.

• 88 = KebschullRichmanRingachEtAl20; 89 = FujitaKodamaduLac20

Work using polysynaptic ret- rograde tracing experiments with rabies viruses in monkeys suggests that indeed highly organized loops exist that bind subsystems into coordinated networks in very specific ways99.

For learned fine forelimb movements, continued thalamic input is required for cortical dynamics to unfold dur- ing execution100.

• 100 = SauerbreiGuoCohenEtAl19

Indeed, medial DCN neurons also exhibit preparatory activity and are required for correct behavioural choice from the late sampling period throughout the delay period102. Moreover, perturbation of ALM activity con- sistently affected medial DCN activity102.

• 102 = GaoDavisThomasEtAl18 -- reviewed below.

Moreover, lateral and interposed DCN input to specific thalamus subdivisions is passed on to the caudal forelimb area motor cortex, and this pathway conveys a movement initiation signal for a specific trained lever pull task104.

• 104 = DacreColliganClarkeEtAl21

But although learning usually starts with a fast increase in the probability of performing the same gross movement, it is followed by a slower phase of refinement of the movement111,112. The time course of emergence of coordinated activity between cortical and striatal cir- cuits matches the learning of the gross movements111

• 111 = LemkeRamanathanGuoEtAl19

These findings are consistent with models proposing that dopamine-dependent cortico- striatal plasticity rapidly increases the probability of re-entering a specific neuronal pattern in the brainstem or cortex via thalamic loops (Fig. 6b), but that the gradual refinement of the pattern selected also requires corti- cal plasticity112 (Fig. 6c). It will be interesting to consider a role of cerebellar–motor cortex loops, via the thala- mus, in this gradual cortical plasticity (Fig. 6c). Loops via thalamic neurons receiving input from basal ganglia or cerebellum116 could mediate different aspects of learning.

[rcoreilly]

GrillnerThompson 21

Grillner, S., & Thompson, W. S. (2021). Basal ganglia reign through downstream control of motor centers in midbrain and brain stem while updating cortex with efference copy information. Neuron, 109(10), 1587–1589. https://doi.org/10.1016/j.neuron.2021.04.015

Note: doesn't actually make point about efferent copy driving learning in cortex -- talk more in terms of information:

The efference copy information for- warded to the cortex and striatum will be important for the planning of upcoming movements or the next phase of a move- ment sequence, rather than the actual movement being executed at that partic- ular moment. With regard to the informa- tion fed back to the cortex, it is noteworthy that lesions of the motor thalamus per- formed clinically for reducing Parkinso- nian symptoms, particularly in the pre- DOPA era (e.g., Duval et al., 2006), had no effects on deficits of gait and posture, but only on hand tremor. No cognitive or emotional symptoms were reported, despite the extensive projections back from the parts of the striatum concerned with emotional and cognitive processing located in the dorsomedial and ventral striatum (Alexander et al., 1986). These findings suggest that the information car- ried back via the thalamus is of the effer- ence copy character, whereas motor symptoms of Parkinson’s disease with re- gard to gait and posture are due primarily to the immediate downstream effects ex- erted from the basal ganglia.

Also, this does not exclude a strong learning role, which would happen very early in development -- this is a general issue with most of the DLS literature on shaping learning in cortex, and why the novel BCI learning is so important.

In addition to the specific upstream projection from the SNr via thalamus or downstream to the many different tar- gets, the authors identify a third category of SNr projections, comprising the pe- dunculopontine nucleus (PPN) and the midbrain reticular formation. In contrast to the orderly arranged input to the thalamic nuclei from SNr, the axonal branches to PPN appear to converge from across the different SNr subpopula- tions. If activity is reduced in any of the SNr subpopulations, PPN will receive less inhibition, and its activity will thus be enhanced. Therefore, increases in PPN activity will reflect the activation of any of the motor centers that are acti- vated from the SNr. The PPN is subdi- vided into two parts, one mainly ascending and another descending. Each part consists of a glutamatergic and a cholinergic portion, and both have been implied in the control of loco- motion (see Grillner et al., 2020).

McElvain et al.’s results were obtained in the mouse. In rodents, the SNr is responsible for around 89% of the output of the basal ganglia, while the smaller globus pallidus interna (GPi; in rodents often called the entopeduncular nucleus) is responsible for the remainder. The latter has not been studied in detail, but some of these neurons are known to have both an ascending and a descending branch, as in SNr. Whether most have multiple axonal branches is not known. In cats, a parcella- tion of the SNr into separate parts control- ling eye, postural, and locomotor movements have been documented (Ta- kakusaki et al., 2004). In primates, the SNr has a similar downstream projection pattern. Kim and Hikosaka (2013) have, for instance, shown that different types of eye movements are handled by different parts of the primate SNr. In the macaque, SNr represents but 45% of the output of the basal ganglia and GPi the remainder (Hardman et al., 2002). The latter has thus increased in relative size. The GPi and SNr are present in all vertebrates (Grillner and Robertson 2016). The SNr is clearly concerned with the control of different midbrain and brainstem motor structures, but the role of GPi is less clear, and it has been suggested that it may be concerned also with the strength or intensity of the movement.

TODO: look into GPi and higher order functionality.

GrillnerRobertson16

Grillner, S., & Robertson, B. (2016). The Basal Ganglia Over 500 Million Years. Current Biology, 26(20), R1088–R1100. https://doi.org/10.1016/j.cub.2016.06.041

Cyclostomes have evolved separately from mammals over more than 500 million years. It follows that when detailed similarities are demonstrated between forebrain circuits in the lampreys of today and those of mammals, these circuits were most likely already present at the dawn of vertebrate evolution (Figure 1)... At this time, many of the molecular components of nerve cells had been designed (through evolution), including most ion chan- nels, transmitters, and ionotropic and metabotropic receptors.

The net effect of this arrangement is that during resting conditions the motor centres are under tonic inhibition (Figure 2), and it is only when subpop- ulations of neurons in the GPi/SNr are inhibited that the corre- sponding motor centres will be disinhibited and free to become active [3,12,18,21–26].

The striatum, the input stage of the basal ganglia, can be subdi- vided into the ventral striatum or nucleus accumbens in mam- mals, which has input from the limbic areas and hippocampus in particular, and the dorsal striatum. In rodents, the dorsal stria- tum, also referred to as neostriatum, can be subdivided into a dorsomedial and a dorsolateral part, and in primates and hu- mans into caudate nucleus and putamen. Finally, in lamprey the striatum forms only one entity. All parts of the striatum are further subdivided in a mosaic of compartments referred to as striosomes and matrisomes, in both lamprey and mammals [13,37,38].

The SPNs of the striosomes inhibit the activity of the dopamine neurons, whereas the matrisomes take part in the control of movement via the direct and indirect pathways [37]. The striosomes can be regarded as related to a circuit of value-based decisions [39–41], as they influence the level of ac- tivity in the dopamine neurons in contrast to the matrisomes, which influence movements (see also below). However, collat- erals of the GABAergic SNr neurons have also been reported to affect the activity of the dopaminergic SNc neurons [42]. Both compartments contain SPNs characterised by a large den- dritic tree with numerous spines.

The thalamic input equals that of the cortical input and repre- sents some 45% of the glutamatergic input to striatum in rodents and originates in particular from the intralaminar nuclei [52]. The central lateral nucleus targets the spines, while the parafas- cicular nucleus targets mainly the dendritic shafts [53]. Both in rodents and lamprey, the thalamic synapses display activity- dependent depression, so that their synaptic potentials decrease progressively in amplitude, whereas corticostriatal synapses are of the facilitating type and the synaptic potentials instead increase in amplitude [54–56]. One may speculate that the fast information via the thalamic route to the striatum pro- vides an initial response for fast action, and that it is subse- quently decreased, while the response via the longer cortical/ pallial route would lead to a more elaborated response and hence take over through the facilitating synapses.

Or, perhaps the thalamic input sets up the action-specific coding: central lateral hitting spines vs. PF on shafts -- one of these is likely to be the "driver". PF is more from ventral / modulatory?? need to figure this part out.

The arkypallidal neurons were recently shown to provide a stop signal to activity in the striatum [59]. These neurons have been characterised in rodents, and whether they exist in other vertebrates needs to be explored.

The dendritic processing seems to be the target of this synaptic interaction [between D1 / D2], rather than regulating the frequency of action potentials, which instead is the role of fast-spiking interneurons targeting the soma level [64].

The membrane properties of SPNs are characterised by a subtype of potassium channels (the inward rectifiers, Kir), which are open under resting conditions and hyperpolarise the cells [5,65–67]. If, however, a cell is depolarised to levels close to generating action potentials, the Kir channels will be closed due to their voltage dependence. This leads to an increase in excitability, which is a hallmark of SPNs, defining their cellular properties. They thus represent the converse of the spontane- ously active SNr/GPi neurons.

TODO: must add Kir to model!

It should be recalled that mammals like rodents, rabbits and cats have a seemingly normal basic movement repertoire when devoid of the entire neocortex, provided that the basal ganglia remain intact [98]. Cats, for example, can move around, search for food and explore different routes in a graceful way.

In other non-mammalian vertebrates, from lamprey to birds, the GPe cells that receive input from D2R SPNs are mixed with the GPi cells, which receive input from D1R SPNs [3,35,100,101]. The output targets of GPe and GPi neurons are, however, similar in the different classes of verte- brates. Functionally, the basis for the D1R and the D2R pathways to GPi and GPe thus remain the same [3,7,13,31,92].

STN is also present in lamprey, with input from both identified GPe neurons and from the pallium, and thus both the indirect and the hyperdirect pathways have been present from very early on in vertebrate evolution [3,44].

Depletion of dopamine as caused by the toxins MPTP or 6-OH-dopamine gives rises to hy- pokinesias, and the same types of symptoms are observed in humans, rodents, amphibians, teleosts and lamprey [116–122].

As discussed above, GPh receives inhibi- tory input from striosomes [13] and excitatory input from pal- lium/cortex orbitofrontal, the limbic system and thalamus [139]. GPh will in turn control the lateral habenulae and thereby the level of activity in the dopamine neurons. This circuit is apparently well-suited to contribute to the control of the level of activity in the dopamine neurons. When it comes to evaluation of whether an action has been successful or not, GPh neurons are known to be modulated in relation to aversive behaviour in both primates and rodents [140,148,149]. In lamprey, GPh neurons are also excitatory and have the same connectivity [129].

This is the key VSGpi pathway in PVLV model for driving LHb phasic dipping!!

So perhaps the GPh -> Hab -> RMTg pathway is the entirety of the original phasic DA, and the Amygdala part came in later??

In all these cases, one most likely utilises elements of the intrinsic motor repertoire, but sequenced in a new way for a given individual.

Since both the direct and indirect pathways are conserved throughout vertebrate phylogeny, it would appear likely that the basal ganglia circuit for selection of behaviour provides a good solution that has not been possible to improve further during evolution, at least with regard to these central and basic features.

GrillnerRobertsonKotaleski20

Grillner, S., Robertson, B., & Kotaleski, J. H. (2020). Basal Ganglia—A Motion Perspective. In Comprehensive Physiology (pp. 1241–1275). John Wiley & Sons, Ltd. https://doi.org/10.1002/cphy.c190045

NOTE: this is a great review paper with all manner of useful references and ideas.

BjurstenNorrsellNorrsell76 -- decorticate cats.. They move around in a seemingly normal way, explore the environment, find their way out of simple labyrinths, get hungry, locate food and feed themselves, go through periods of sleep and maintain a diurnal rhythm. They are thus able to perform very important aspects of their normal behavioral repertoire.

In contrast, if a lesion is made leaving only the mid- brain and caudal brainstem structures intact, these animals are unable to adapt to the environment, although specific motor programs can still be elicited but now performed in a stereotypic manner (136, 137). They include locomotion, eye movements, swallowing, and respiration. Thus, generation of coordinated movements adapted to the environment is entirely dependent on the integrity of the basal ganglia, as well as its input from the thalamus and other sources. The relation between the cortex and basal ganglia has been for- mulated by Arber and Costa (13) as if the cortex broadcasts a wish, the basal ganglia will decide, whether the wish should be fulfilled or not, that is if a movement should be initiated and executed or not.

The rostral part of the caudate nucleus and putamen corresponds approximately to the rodent dorsomedial striatum (DMS), while the dorsolat- eral striatum (DLS) corresponds to the caudal putamen and caudate tail part (184).
+ @thazy -- 184 = Kim HF, Hikosaka O. Parallel basal ganglia circuits for voluntary and automatic behaviour to reach rewards. Brain 138: 1776-1800, 2015.

The importance of the thalamic input is borne out by exper- iments showing that when a rodent has learnt a specific task (e.g., lever-pressing with a certain interval), the motor areas in the cortex can be removed and the performance is unchanged. If, however, after learning the task, the thalamic input from PF and CL is incapacitated the task cannot be performed or learnt (179, 355).

The PF nucleus is the thalamic nucleus with the highest number of neurons projecting to the striatum. In mice, it is subdivided into three parts with transcriptionally and physiologically different neuronal properties (223). They project to the dorsolateral, medial, and central parts of the striatum, representing the somatosensory, associative, and limbic circuits, respectively. All three parts project to SPNs and ChINs, whereas FS interneurons are only activated in the dorsolateral part (107). The PF nucleus projects only to the matrix portion of the striatum. Its axons branch profusely in the striatum and form several dense plexuses (202). This suggests that PF axons may have an overall effect on the excitability of the striatal circuitry rather than targeting particular modules.

• Key point: DLS PF inputs only on inhibition. And only on Matrix. Broad modulation.

The PF neurons have long dendritic branches with spines and relatively few branch points. They can, therefore, be described as being of a reticular type rather than a classic thalamic relay neuron, in contrast to the neurons in CL [see below; (99)].

The PF neurons receive broad synaptic inputs from the SC, PPN, brainstem reticular formation, cortex, cerebellum, and GPi/SNr (311, 363). The PF neurons are activated by unexpected salient auditory stimuli, like beeps and clicks, and by visual and somatosensory stimuli (229).

In nonhuman primates, the rodent PF nucleus is subdivided into a CM part and a PF part, however, often treated together (118, 311, 363). Acute inactivation of the PF part disrupts attention processing and the role of CM/PF is thought to relate to attention shifting, behavioral switching, and reinforcement processes (234–236, 311, 363). The CM part receives input from the motor cortex and the other sources referred to above. Neurons in CM/PF respond with short latency to behaviorally salient stimuli (visual, somatosensory, or auditory). The CM projects primarily to the putamen, corresponding to the rodent DLS and can have a powerful effect on ChINs. If salient stim- uli has been associated with reward, ChINs respond with a brief activation followed by inhibition and a postinhibitory rebound activation.

In addition to short-latency activated neurons, the CM contains neurons activated by salient stimuli with a long latency facilitation preceded by inhibition (234–236). In a situation in which the monkey under certain experimental conditions receives a large or small reward (one or several drops of juice), such long latency CM neurons respond pri- marily when the monkey realizes that it will receive a smaller reward than it had hoped for. Thus, when accepting this fact, these long-latency neurons were activated, and this is thought to be a behaviorally important signal for associative learning.

Intralaminar central lateral nucleus (CL) The CL nucleus is adjacent to the PF nucleus but contains fewer neurons. CL neurons have short bushy dendrites with many branch points described as classic thalamic dendrites studded with spines (202). They display more burst firing than PF neurons. Their axon collaterals are long and smooth with synapses en passant. In contrast to PF, CL neurons synapse almost exclusively on the spines of SPNs and provide stronger depolarization, and drive dSPNs more efficiently than iSPNs (99). The NMDA/AMPA ratio is around 0.5 for CL synapses, but 2.5 for PF synapses, suggesting that PF synapses are involved in long-term plasticity and facilitating plateau potentials, more so than CL synapses.

The rat SNc has around 12,000 neurons, and it has been estimated that each neuron has 100,000 to 370,000 synapses. For comparison, each SPN makes 300 to 500 synapses and a motoneuron even less (39).

In primates, the head and tail of the caudate nucleus are supplied by different sets of dopamine neurons from the SNc, which may relate to their involvement in habitual and goal-directed aspects of behavior, respectively (184). The heterogeneity of midbrain dopamine neurons is also manifested through transcriptomics, and they can be sub- divided into seven subpopulations (341).

An unusual mechanism is provided by cholinergic axons (ChINs) that can activate dopaminergic terminals directly via nicotinic receptors, thereby causing a release of dopamine (52, 340), as shown with synchronous optogenetic activation of ChIN populations.

• Note: above is key for local da signals.

The cholinergic PPN and LDT neurons are activated in reward situations and provide part of the excitatory drive to dopamine neurons (135, 263). They also have extensive striatal projec- tions, with the rostral part targeting the DLS, while the caudal part targets the dorsomedial and ventral striatum (77, 78). PPN and LDT neurons are also known to provide input to thalam- ostriatal neurons. PPN/LDT neurons can thus affect the stria- tum via three avenues, directly and indirectly via dopamine neurons in SNc or VTA and via the thalamus. In the stria- tum, the PPN targets both spines and dendritic shafts of SPNs and perhaps also interneurons, but only in the matrix region, avoiding the striosomes (77, 78).

Within the PPN, there is in addition to the cholinergic popu- lation, a glutamatergic group of PPN cells that represents an independent set of neurons and constitutes around 60% of the total number of cells (14). The glutamatergic PPN projects to the entire striatum (Figure 5), but in contrast to the cholin- ergic PPN neurons they do not target SPNs, only the striatal interneurons, prominently ChINs, FS interneurons and also in a smaller number of LTS interneurons and THINs recorded [(14); Figure 5]. When glutamatergic PPNs are activated opto- genetically, they provide a strong excitation to these different groups of striatal interneurons, comparable to thalamic exci- tation. Both types of SPNs receive disynaptic inhibition due to the activation of GABAergic interneurons, without a trace of excitation. This is thus a prominent way of shutting down the entire output of the striatum. In the freely moving mouse, optogenetic activation of glutamatergic PPN neurons during movement results in an ipsiversive rotation and thus asym- metric movements. This rotating behavior stops as soon as the stimulation is terminated. It resembles the one-sided rota- tion produced by unilateral inactivation of the dopamine sys- tem (344).

• Note: PPN glut is major shutdown pathway

Traditionally, the reciprocal relation between the cerebellum and the cerebral cortex has received considerable attention over many years (171). Recently, however, it has become evident that the different output nuclei of the cerebellum all target neurons in the intralaminar nuclei of the thalamus that in turn project to the dorsal striatum (53, 59, 165, 169, 358). Both types of SPNs and ChINs become activated. It will, therefore, be important to consider the cerebellum as a major thalamic input to the striatum, recalling that the thalamus is responsible for around 40% to 45% of the glutamatergic input to the striatum.

• TODO: follow up on these refs -- key avenue for DCN to train or at least modulate DLS.

Also, in the GPi/entopeduncular nucleus (352), there are projections to both brainstem centers and thalamus, but less detail is available (329).

The STN consists primarily of projection neurons with few collaterals and limited interactions between them within the nucleus (315). There is in addition a smaller subpopu- lation with extensive collateralization (127), also clear from a transcriptomic analysis (266)... The hyperdirect cortico- STN-SNr/GPi pathway can thus effectively terminate a given pattern of motor behavior. In addition to the prominent input from the cerebral cortex and thalamus, the STN receives additional inputs from the SC, cholinergic and glutamatergic PPN neurons, and dopaminergic SNc neurons (66, 186).

The STN is in addition part of the indirect pathway and receives GABAergic input from the prototypical cells of the GPe, while the STN provides excitation to the same cells (8, 186). The prototypical cells represent 70% to 80% of the GPe neurons (219), but there is specificity so that only a limited number of GPe neurons interacts with a select group of STN neurons (22). The sequence in the indirect pathway is as follows: iSPNs inhibit spontaneously active prototypical GPe cells, which disinhibit STN neurons and thereby enhance their activity (Figure 7). In addition, the decreased activity of the GPe neurons directly disinhibit SNr/GPi and enhance their activity.

The arkypallidal cells, on the other hand, express the tran- scription factor FoxP2 and instead have a massive inhibitory projection back to the striatum to both SPNs and interneu- rons (Figure 7) with an estimate of 10,000 varicosities. The somatodendritic morphology is similar between the two types except that the prototypical cells are described as aspiny, whereas the arkypallidal cells have spines (185, 187, 219). Both types have a local collateral network within the GPe.

The input from iSPNs and STN to the prototypical GPe cells is well documented. During resting conditions, the prototypical cells are active at a higher rate than the arky- pallidal cells. The latter receive inhibitory input from the prototypical cells, and excitatory from STN and tend to be rhythmically active in a reciprocal relation to the prototypical cells (1, 88, 219). Moreover, Karube et al. (178) recently showed that the arkypallidal neurons also receive strong monosynaptic input from PT-neurons in the cortical motor areas (M1 and M2) and inhibitory input from dSPNs (180). However, during movements, the arkypallidal neurons tend to be active at a higher rate. An activation of a population of arkypallidal neurons, such as from the motor areas, would very efficiently inhibit striatal neurons within their striatal target area, and they have been referred to as potential “stop cells” (219). Mallet et al. (222) showed that arkypallidal neurons were activated specifically at a stop task, in contrast to the prototypical neurons, and may thus contribute to this function together with the STN. Conversely, activity in the direct pathway inhibits the arkypallidal stop cells (180), which seems purposeful, since direct pathway activity is related to initiation of movement.

• Todo: check up on more recent updates and behavior in pcore model.

The Costa laboratory (73, 333) showed that dSPNs and iSPNs were coactivated during the initiation and execution of locomotion and lever-pressing. Both were activated just prior to the movement and dSPNs remained active throughout the movement, while the activity in iSPNs tended to decrease earlier. This finding was further corroborated by Parker et al. (269) showing simultaneous recordings from thousands of dSPNs and iSPNs in the DMS. The two types of SPNs were coactivated in freely moving mice (73). Moreover, specific clusters of neurons within the striatum, both in the dorsomedial (269) and dorsolateral part (189, 349) were activated in relation to different aspects of the movement. This means that the striatum is subdivided into subpopulations of neurons engaged in the initiation and exe- cution of different aspects of motor behavior. When recording the moment-to-moment changes in the activity of dSPNs and iSPNs of the DLS in freely moving animals (225), different subpopulations were activated in succession and the two types of SPNs were sometimes decorrelated. Lesioning the DLS abolished the ability to generate appropriate sequences.

Furthermore, they reported that optogenetically and electrophysiologically identified dSPNs were activated during the execution of a pull movement, and when the reward signal occurred dSPNs continued to discharge, while a no-reward signal led to instantaneous cessation of activity (Figure 8). The iSPNs conversely were strongly activated when the no-reward signal appeared, but not when a reward occurred.

• todo: follow up on this paper!

The dSPNs thus encoded reward outcomes, whereas indi- rect pathway neurons encoded a no-reward outcome and next-action selection. After a series of mostly no rewards, the rodent is assumed to change strategy from push to pull or vice versa. A continued activity in the dSPNs thus signals that a switch will not occur, while a high level of activity in the iSPNs predicts that a switch may occur.

In rodents, the habitual and flexible parts corre- spond to the DLS and DMS, respectively.

To learn to combine these individual movements into a well-programmed sequence of events is referred to as chunking and is also an important contribution of the basal ganglia (130, 174, 175). In rodents, habit formation is dependent on the DLS and in primates on putamen and the tail of the caudate nucleus. Patients with stroke affecting the basal ganglia have marked difficulties in learning to perform new sequences compared to healthy controls (41). Similarly, Parkinsonian patients have difficulty combining different movements into a well-coordinated sequence [(176); and see below].

If we elicit these three motor programs in a sequence, the dopamine activity will be combined with the concurrent inhibition of ChINs (decreased activation of M4 receptors), which would promote the development of LTP (251, 252) in the cortico-striatal synapses (dSPNs) active at this very moment. If the next movement occurs close in time when there is still residual dopamine activity related to the first movement, this would add to the dopamine activity linked to the second movement and subsequently to the third. If the dif- ferent movements appear in rapid succession combined with the dopamine drive, it would thereby gradually link the three different commands into one combined sequence—chunking.

• DA carryover forges links in the chain.. hmmm...

The role of the striosome compartment appears to be to contribute to the evaluation of whether an action has been contextually appropriate for the behavior or if it should be suppressed [see above; (182–184, 261)].

and the cortical input to striosomes originates mainly in the limbic areas of cortex and the pregenual anterior cingulate cortex (7, 95), while the matrix area primarily receives input from neocortex. The thalamic input to striosomes originates from a restricted part of the intralaminar nucleus and does not include the PF nucleus (120).

• Note: direct BOA -> striosome story here.

In contrast to the matrix area, dSPNs in the striosomes target dopamine neurons in the SNc/VTA (120, 133, 353), whereas iSPNs target a subpopulation of spontaneously active glutamatergic cells adjacent to GPi, referred to as globus pal- lidus cells projecting to the habenulae (GPh), or border cells in primates (Figure 14), which as the name implies, project to the lateral habenulae (18, 47, 155, 163, 318, 320). This is an intricate organization conserved throughout vertebrate phylogeny. Recent data suggest that some SPNs in the matrix also target dopamine neurons (310).

In addition to striosomes, the GPh receives input from the cerebral cortex and thalamus and provides excitation to the lateral habenula (Figure 14), which in turn projects directly via a set of GABAergic interneurons to dopamine neurons. When the GPh increases in activity, the lateral habenula follows, and inhibits dopamine neurons. Conversely, if the spontaneously active GPh is inhibited, the drive to the lateral habenula decreases and dopamine neurons are disinhibited and their activity enhanced, as in a reward situation.

In the mouse, it was shown by using expression of light-sensitive channel rhodopsins that a depression of GPh activity leads to a reward situation behaviorally, while an enhancement leads to a reduced motivation (320). This has also been demon- strated in primates (162). The GPh is present from lamprey to primates—in lamprey as a separate entity, in rodents as one part of the GPi (entopeduncular nucleus) and in primates, it surrounds the outer rim of the GPi with some neurons located in the center (Figure 14). In addition to the GPh, the lateral habenula receives input from the lateral hypothalamus (206) and limbic structures. If a motor action leads to an enhanced dopamine release, this may lead to a long-term facilitation of the specific cortico-striatal axons that triggered the action, reinforcing and strengthening this particular set of commands. This is the basis of reinforcement learning (see below)

Ca2+ enhancement is, however, not sufficient for LTP to occur and requires, in addition, a parallel activation of cAMP-PKA signaling resulting from the activation of D1 receptors in the direct pathway and adenosine A2A receptor activation in the indirect pathway (253, 302)

Likewise, LTD signaling cascades seem to counteract LTP initiation (302). The specific interactions are, however, not well understood, although calcineurin- dependent dephosphorylation of DARPP-32 might be one contributing factor and calcium-dependent activation of phosphodiesterases (PDEs) another (31).

Thus, the probability to induce LTP in the corticostriatal synapses is enhanced significantly by a transient reduction in the respective Gi-coupled branch (M4 and D2 recep- tors, respectively) together with a transient increase in the Golf coupled branch [activating adenylyl cyclase; (49, 252, 253, 257)].

Thus, a local FS network with gap junctions can function as a detector of transient increases in synchronous cortical inputs. When embedding such gap junction-coupled FS networks into a striatal net- work consisting of several 1000 SPNs, it turned out that the presence of gap junctions was important for controlling the balanced activity between the dSPNs and iSPNs (76). When a dopamine-depleted condition was simulated in the network, the presence of gap junction-coupled FS interneurons con- tributed significantly to conveying cortical β-oscillations into the striatal network, which are enhanced in PD (75). These observations suggest that the FS network gave rise to an entrainment effect of the SPNs when electrically coupled, in line with a study (247) showing that wide-spread FS interneu- ron inhibition has a stronger inhibitory effect onto SPNs than slightly synchronized FS interneuron populations. That FS interneurons convey frequency components within the striatal network is also supported by simplified spiking models (23).

allow large-scale in silico reconstructions of the (mouse) striatum (161) considering the cellular properties and soma-dendritic morphology of large populations of iSPNs, dSPNs, ChINs, FS interneurons, and their synaptic interactions.

[rcoreilly]

McElvainChenMooreEtAl21

McElvain, L. E., Chen, Y., Moore, J. D., Brigidi, G. S., Bloodgood, B. L., Lim, B. K., Costa, R. M., & Kleinfeld, D. (2021). Specific populations of basal ganglia output neurons target distinct brain stem areas while collateralizing throughout the diencephalon. Neuron, 109(10), 1721-1738.e4. https://doi.org/10.1016/j.neuron.2021.03.017

Most critical, SNr firing responses to depolarizing inputs are highly linear (R2 = 0.99 ± 0.01) (Figure 3C). The limited adaptation and linearity of the neuronal input-output relation is consistent with open-loop drive for a motor system (Astro ̈ m and Murray, 2008), akin to linear neuronal and synaptic properties common in brain stem sensori- motor networks (Kolkman et al., 2011; McElvain et al., 2010) and unlike forebrain projection neurons.

Neurons in the SNr that project to brain stem areas span a range of post-inhibitory rebound firing capabilities. Robust non-linear rebound firing responses to hyperpolarization are greatest in neurons projecting to the Med and IC, which ex- hibited bursting up to hundreds of spikes per second and aver- aged 69 ± 15 Hz and 58 ± 11 Hz, respectively (Figure 4E). In contrast, transient and sustained rebound firing of other projec- tion populations is modest, and neurons projecting to the DR exhibit the lowest capacity for rebound (Figures 4E).

Projection neurons that exhibit slower active and passive properties are concentrated in the medial portion of the nucleus (Figure S6C), which overlaps with terminal fields from associative striatal regions. Slow neurons are additionally located in a small cluster at the dorsal-most extreme.

To verify that the PPN is indeed a broad collateral hub of the SNr, the PPN was additionally targeted with the retrograde lentivirus. Consistent with data showing broad collaterals to the PPN, SNr neurons labeled retrogradely from the PPN collateralized to all other SNr targets (Figures 6E, top).

todo: need to learn more about PPN!

This study demonstrates direct and extensive connections from basal ganglia to diverse components of the thalamocortical and brain stem motor systems. Rather than a single monolithic output, the SNr contains distinct projection neuronal pools that are electrophysiologically specialized and topographically orga- nized and differentially target brain stem effector circuits. Each projection pool emits ascending, spatially organized collateral branches to several thalamic nuclei and broadly collateralizes to the PPN and MidRF.

[rcoreilly]

TurnerDesmurget10

Turner, R. S., & Desmurget, M. (2010). Basal ganglia contributions to motor control: A vigorous tutor. Current Opinion in Neurobiology, 20(6), 704–716. https://doi.org/10.1016/j.conb.2010.08.022

Key, widely-cited paper: slow build to central hypothesis: BG is for learning new motor skills. Once that is done, it retains a role in motivationally-driven modulation of motor gain, but little else. From model perspective, motivational modulation is what drives learning so it all makes sense.

Among the most influential hypotheses, one may cite: selection of action and suppression of potentially competing actions and reflexes [1–3], control of the scale of movement and related cost functions [4�,5��,6��], on-line correction of motor error [7,8], motor learning [9,10,11��], and the reten- tion and recall of well-learned or natural motor skills [10,12,13,14��].

Interestingly, a very different outcome is observed following discrete lesions of the main output regions of the BG [the globus pallidus internus, GPi, or substantia nigra pars reticulata, SNr (Box 1)]. In that case, behavioral effects are typically subtle or imperceptible [4�,19], con- sistent with the fact that surgical ablation of large portions of the GPi (‘pallidotomy’) is an effective treatment for striatal-associated disorders such as PD and dystonia [20,21,22�].

Together, these observations can seem paradoxical. BG- associated disorders arise primarily from pathology in the principal input nucleus, the striatum, and can be alleviated by lesions of a BG output nucleus. The seeming contra- diction can be explained by the concept that it is better to block BG output completely than allow faulty signals from the BG to pervert the normal operations of motor areas that receive BG output [15]. Abnormalities in striatal function, whether from frank lesions [23,24] or neurotransmitter imbalance [25–27], induce grossly abnormal ‘pathologic’ patterns of neuronal activity in the inhibitory output neurons of the BG. These abnormal firing patterns are thought to disrupt the normal operations of BG-recipient brain regions.

First, movement-related changes in firing in GPi are almost always influenced by specific characteristics of a movement such as its direction, amplitude, and speed (i.e. movement kinematics) ([36], and references therein). However, motor activity in GPi neurons is also often influenced by the context of the behavioral task being performed. Single-cell responses in GPi can differ depending on the memory requirements of a task [37], whether the movement is discrete or part of a movement sequence [38], the reward contingencies of the task (i.e. whether or not a primary reward is expected to follow the movement [39]), and the learning context [40]. Similar influences of behavioral context have been observed in the oculomotor circuit in animals performing eye movement tasks [3].

These observations suggest that the BG motor circuit is not involved directly in movement execution, but rather that it brings cognitive and motivation-related signals together with signals related to movement kinematics [37].

In GPi, for example, onset latencies of peri-movement changes in neural firing are typically clustered around the time of earliest agonist muscle activity (‘EMG’; 50–80 ms before movement onset; see [36] for references) and after the activation of primary motor cortex (�120 ms before movement). Interestingly, peri-movement increases in GPi firing have later onset times than peri-movement decreases [36].. The timing of movement-related activity in GPi makes it impossible for GPi output to contribute to processes that are completed before the initial activation of a movement’s prime moving muscles (e.g. selecting which agonist muscles to activate or trig- gering their activation). Based on timing, GPi activity may modulate the ongoing commands issued by BG-recipient motor control centers.

Given that increases in GPi firing inhibit activity in recipient motor control circuits, this observation has been cited as evidence that an important function of output from the BG motor circuit is to suppress or inhibit patterns of motor activity and reflexes that would be inappropriate or in conflict with the movement being performed [1–3]. The late timing of peri-movement GPi activity, particularly that of increases in firing [36], appears to conflict with that concept. To be more specific, the rest activity of antagonist muscles [42,43] and the gain of reflexes that might interfere with a desired movement [44,45] are suppressed tens of millisecond before activation of a movement’s prime moving muscle. At the cortical level, suppression of potentially competing activity pat- terns also begins before the initial activation of agonist muscles [46,47]. Thus, the known inhibitory processes that contribute to movement selection begin too early to be mediated by output from the BG. Cortical mechanisms may mediate most aspect of movement selection [6��,48]. A potential role for GPi movement-related activity in the control of movement vigor is discussed below.

Several studies over the past three decades have investigated the effects of GPi inactivation on motor performance in neurologically normal animals [4�,49–54,55��]. Although very different motor tasks were used and minor disparities were sometimes noted [4�], results from these studies are surprisingly consistent. Overall, they reveal a relatively discrete group of deficits and a wide range of preserved functions. Five points are particularly noteworthy.

First, GPi inactivation does not lengthen reaction times (RTs; Figure 1e [4�,50–52]), consistent with the frequent clinical observation that pallidotomy, if anything, speeds movement initiation [21,22�]. These observations are not consistent with the idea that the BG contributes to the selection or initiation of movement.

Second, GPi inactivation does not perturb on-line error correction processes [4�] or the generation of discrete corrective submovements in a single-joint movement task [52]. These findings are consistent with the observation that rapid hand-path corrections are preserved in PD patients [56], but present challenges for the idea that the BG mediates the on-line correction of motor error [7,8].

Key question: are these the things that are impaired with cerebellar lesions? presumably. This represents a clear dissociation if so.

Third, GPi inactivation does not affect the execution of overlearned or externally cued sequences of movements. This was shown in two recent studies in monkeys [4�,55��] (Figure 1b–e).

Transient inactivations in the skeletomotor territory of GPi using muscimol, a GABAA agonist, impaired specific facets of motor per- formance (see below), but had absolutely no effect on sequencing-related aspects of task performance. In particular, GPi inactivation did not affect an animal’s ability to chain independent movements together in quick succession (under the Random condition) or to reproduce an OverLearned sequence as a fluid predictive arpeggio. Importantly, GPi inactivation did not alter the animals’ habit-like tendency to reproduce the Over- Learned sequence as a predictive whole when, by coinci- dence, the initial targets of a Random trial matched the OverLearned sequence. These results are consistent with a previous study showing that GPi blockade does not impair the reach-to-retrieval transition in a simple reach- grasp-and-retrieve task [54]. They also agree with reports that pallidotomy does not impair the execution of well- learned motor skills in patient populations [21,22�] and the consistent observation that lesions of the BG homolog in the song-bird have little impact on the execution of already-learned song sequences [9]. By contrast, these finding contradict claims, based on neuroimaging and clinical evidence, that the BG is involved in the long- term storage of overlearned motor sequences [13] or the ability to string together successive motor acts [56].

Fourth, GPi inactivation reduces movement velocity and acceleration. This is, without a doubt, the most consistent impairment found across studies [4�,49,51,53,54,55��] (Figure 1e).

Fifth, for fast targeted movements, GPi inactivation causes a marked hypometria (undershooting of the desired movement extent, Figure 1e) that is a consistent across directions of movement but is not accompanied by changes in movement linearity or directional accuracy [4�,53]. The degree of hypometria induced by an inacti- vation correlates closely with early markers of movement slowing (peak velocity, acceleration, and agonist EMG) [4�]. A similar form of global hypometria with no direc- tional bias is observed in PD patients [61,62]. These results present challenges for the suppression hypothesis in which movement-related increases in GPi activity are proposed to inhibit competing motor commands and reflexes [51,54]. It is difficult to conceive of a general disturbance in motor command selection or muscle ago- nist/antagonist balance that would affect movement extent and speed equally for all directions of movement, but have no effect on the initial direction of movement, hand-path curvature, or final directional accuracy [4�,53].

Many of the observations summarized above can be explained by a classic and seemingly simplistic concept that the BG regulates the speed and size of movement (i.e. ‘movement gain’ [49,63]).

so boring!? what's the point? key missing perspective: early development is when BG does its job -- rest of the time it is just waiting for something new to learn, and regulating things accordig to motivational variables??

An independent line of evidence regarding the gain hypothesis originates from behavioral observations that, at certain stages of motor planning, ‘movement gain’ (the extent and speed of movement in a given workspace) is controlled independently of movement direction ([72,73], and references therein). Consistent with those observations, current models of motor control recognize the need for a mechanism that identifies optimal balances between the ‘costs’ of movement (e.g. physical work, elapsed time, and control complexity) and the rewards available in a given behavioral setting [6��,74]. Motor cost terms, which scale with velocity, amplitude, and other aspect of motor performance, may link an animal’s previous experience of the cost/benefit contingencies of a task [75] to its current allocation of energy to meet the demands of a specific task [57��,66]. We and others have conjectured that a breakdown in that link would yield motor impairments similar to those observed follow- ing GPi inactivation [4�,6��]. In essence, the BG motor circuit may compute and store cost functions that modu- late motor performance based on an animal’s previous experience of the requirements of a task and the rewards available.

This role for the BG motor circuit is consistent with an emerging view that the BG as a whole, including its dopaminergic innervation, regulates action motivation or response ‘vigor’ [66,75]. Limbic circuits of the BG, for example, have been implicated in the appropriate scaling of a subject’s rate of responding or choice of effortful responses to match the cost/benefit tradeoff of a task [76,77�]. This idea is supported further by obser- vations that focal damage in the BG is often accompanied by abulia or ‘auto-activation deficit’ in which patients suffer from a marked deficit in motivation to perform spontaneous acts despite an absence of overt motor impairment [78]. Schmidt et al. [5��] provided a striking demonstration of this disorder by showing that patients with bilateral lesions of the putamen or pallidum are able to control grip forces normally in response to explicit sensory instructions, but do not increase grip force spon- taneously despite full understanding that higher forces will earn them more money. The authors concluded that BG lesions specifically block the influence of task incen- tives on movement vigor.

The slowing and hypometria induced by experimental inactivation of GPi are similarly immune to many aspects of behavioral context (e.g. differences in memory con- tingency and sensory cueing [4�,53,55��]).

The hypothesis that BG modulates movement gain has been challenged on the grounds that movement-related activity in the striatum and globus pallidus begins later than activity in motor regions of cortex [15]. Both psy- chophysical [81,82] and electrophysiological [83,84] evi- dence suggests, however, that movement gain can be modulated after the earliest stages of movement initiation. Furthermore, electrical stimulation of the GPi can modify the speed of reaching movements even when stimulation is delivered solely at the time of agonist EMG onset (i.e. at latencies similar to those of move- ment-related GP activity [49]). Thus, the latencies of movement-related activity in the GPi may be appropriate for a role in modulating movement gain. BG output may exert its scaling influence both at the cortical level (via thalamo-cortical pathways) and at brainstem and spinal motor centers via descending outputs from GPi (Box 1).

BG and learning, but not retention

Growing evidence suggests that the connectivity and physiology of the BG is ideally suited for fast ‘directed’ formation of reward-relevant associations, which over the course of practice train slower Hebbian learning in tha- lamocortical circuits [7,85��]. This view predicts that BG circuits are intimately involved in and necessary for new skill learning, but are of far less importance in the reten- tion and recall of well-learned motor skills. This view constitutes a major revision of the long-standing and highly influential theory that memory traces underlying motor skills are stored long-term in the BG [14��,86]. The general concept that the BG and its dopaminergic inner- vation play central roles in many different forms of learning is noncontroversial and supported by a vast literature (for recent reviews, see [10,14��,87]). This sec- tion focuses on evidence related to the concept that BG circuits may be involved selectively in reward-driven acquisition, but not in long-term retention or recall of well-learned motor skills.

Single unit recording studies have demonstrated major changes in neuronal activity in the BG as animals learn procedural tasks [88–90], and a few of these studies provided evidence that learning-related activity appears earlier in the course of learning in the striatum than in connected regions of cortex [89,90]. Importantly, several reports have indicated that, after a motor skill becomes well-learned, the prevalence of task-related activity in the motor striatum declines and neuronal response latencies shift to follow movement onset (see [91�] for references). These studies suggest that the BG motor circuit is acti- vated preferentially during the learning process. Note that some task-related activity is still present in the BG in overtrained animals [38,88,91] thereby suggesting either that BG activity is not involved solely in learning or that a certain degree of learning persists even in overtrained animals.

91. Tang CC, Root DH, Duke DC, Zhu Y, Teixeria K, Ma S, Barker DJ, � West MO: Decreased firing of striatal neurons related to licking during acquisition and overtraining of a licking task. J Neurosci
    2009, 29:13952-13961.
    This is one of several studies from the West group showing that neurons in the skeletomotor region of the rat striatum are far more likely to show task- related activity during initial stages of training on a task than after extensive training. These studies, as a whole, stand out for their care in recording from identified single body part-related neurons. Results from this specific study: (1) suggest that the decline in prevalence of task-related activity is not simply a correlate of habit formation and (2) corroborate the observation that, in a minority subpopulation of striatal neurons, task-related activity persists and is even accentuated with overtraining.

One of the sequelae most consistently associated with pallidotomy is an impaired ability to learn new motor sequences [22�,92] and arbitrary stimulus–response associations (e.g. [93]). An important but often overlooked point is that pallidot- omy does not degrade, but typically improves, a patient’s ability to perform overlearned motor skills such as groom- ing, dressing, and handwriting (many of which are assayed by the ‘activities of daily living’ scale, see pallidotomy references above). Given that pallidotomy produces large well-localized lesions centered on the motor territory of GPi, these studies provide some of the best evidence that BG output is necessary for motor skill learning, but not for the retention and recall or well-practiced skills.

Combined with the observation that transient inacti- vation of the GPi does not degrade the performance of well-learned skills in neurologically normal animals [4�,55��], these observations suggest that the BG func- tions as a kind of tutor, being important for learning, but not for storage or recall of already-learned information. It is likely that the motor cortices play a central role in the long-term retention and recall of skills based on the evidence for slow synaptic modification [94], the emergence of task-specific activity [95], and even macro- scale reorganization at the cortical level [96] in response to long-term training on a skill. This concept fits well with the idea that the responses of nigrostriatal dopa- mine neurons mediate fast reinforcement-driven synap- tic plasticity in the BG [97��]. Cortical plasticity appears to be inherently slower than striatal plasticity because it is insensitive to phasic dopaminergic training signals and thus governed by Hebbian learning rules [98]. Long-term retention at the cortical level may bring advantages, however, owing to greater processing effi- ciency (i.e. lower conduction times and numbers of synaptic delays [85��]).

[rcoreilly]

PhillipsSchulmannHaraEtAl19

Phillips, J. W., Schulmann, A., Hara, E., Winnubst, J., Liu, C., Valakh, V., Wang, L., Shields, B. C., Korff, W., Chandrashekar, J., Lemire, A. L., Mensh, B., Dudman, J. T., Nelson, S. B., & Hantman, A. W. (2019). A repeated molecular architecture across thalamic pathways. Nature Neuroscience, 22(11), 1925–1935. https://doi.org/10.1038/s41593-019-0483-3

• AD = anterior dorsal, RE = reunions -- both live in their own cluster (with clear relationships to others nonetheless)

• VL = cerebellum, VA = SNr / BG -- live in different groups.

• Primary = sensory (LGd, VB = vis, somato) / VL, cerebellar motor -- fast, narrow spikes -- L4 and NOT L1

• Secondary = BG, higher-order -- wider, slower -- L1 very strongly

• Tertiary = intralaminar, including CM, -- widest, slowest -- L4 but broadly distributed

[rcoreilly]

FosterBarryKorobkovaEtAl21

Foster, N. N., Barry, J., Korobkova, L., Garcia, L., Gao, L., Becerra, M., Sherafat, Y., Peng, B., Li, X., Choi, J.-H., Gou, L., Zingg, B., Azam, S., Lo, D., Khanjani, N., Zhang, B., Stanis, J., Bowman, I., Cotter, K., … Dong, H.-W. (2021). The mouse cortico–basal ganglia–thalamic network. Nature, 598(7879), Article 7879. https://doi.org/10.1038/s41586-021-03993-3

Definitive anatomical tracing through the whole BG loop: it really is a loop! Also, is it really legal to have a paper so full of acronyms, that does not appear to define many of them, and has no central loop table?

Here we map the multi-synaptic output pathways of these striatal domains through the globus pallidus external part (GPe), substantia nigra reticular part (SNr), thalamic nuclei and cortex. Accordingly, we identify 14 SNr and 36 GPe domains and a direct cortico-SNr projection. The striatonigral direct pathway displays a greater convergence of striatal inputs than the more parallel striatopallidal indirect pathway, although direct and indirect pathways originating from the same striatal domain ultimately converge onto the same postsynaptic SNr neurons. Following the SNr outputs, we delineate six domains in the parafascicular and ventromedial thalamic nuclei. Subsequently, we identify six parallel cortico–basal ganglia–thalamic subnetworks that sequentially transduce specific subsets of cortical information through every elemental node of the cortico–basal ganglia–thalamic loop. Thalamic domains relay this output back to the originating corticostriatal neurons of each subnetwork in a bona fide closed loop.

PF thalamus

This remains more segregated throughout the loops, as shown in above figure. Note that ECT / TEa (lateral IT areas) project into striatum but do not get a PF return prjn (open loop), but Perirhinal does have a PF loop.

Notably, the "sensory associative" areas also seem open loop, where ORBital gets the return. The striatal projections from some of these hit both the closed loop and an open loop through trunk.

VM thalamus

f. The ventromedial thalamic nucleus (VM) also has 6 output channels, but in VM the motor output pathways to cortex are not organized somatotopically by body subregion, as they are in PF, but rather the VM.primary (VM.p) projects to all body regions of primary motor and primary somatosensory cortex, and the VM.secondary (VM.s) projects to all body regions of secondary motor cortex. The limbic (lim), ventral striatal (vs), associative (a), and mouth (m) output channels are similar to the corresponding channels in PF. g, Nigrothalamic and thalamocortical tracing data support the 6 channel model of VM. Note that the same transsynaptic tracing experiment described for d also resulted in the ventral striatal data in g (i.e., the image labeled ‘SNr.dm’). h, A summary diagram illustrating the thalamocortical projection differences and similarities between VM (shapes) and PF (colors). The tr and ll zones are separated in this PF model because tr and ll exhibit a partially overlapping, partially separable thalamocortical connectivity pattern (data not shown), and emphasizing them separately underscores the somatotopic features of this pathway. All photomicrographs are coronal, medial is left, and scalebars = 200μm.

[rcoreilly]

PF Thalamus

FallonHughesSeverinoEtAl23

The parafascicular (Pf) nucleus is an intralaminar thalamic nu- cleus that sends glutamatergic projections to the basal ganglia, a set of subcortical nuclei implicated in the initiation and execu- tion of actions.1,2 Traditionally considered a thalamic component of the reticular activating system, the Pf receives afferent inputs from both neuromodulatory brainstem circuits as well as sensory pathways.3

Studies have implicated the Pf in arousal and attention.4–6 In primates, many Pf neurons respond with short latency to salient stimuli in multiple modalities, especially when they are unex- pected. More recent work also showed that the Pf plays a role in action initiation. In mice, many Pf neurons increase activity at the time of action initiation in an operant lever-pressing task, and inhibition of these neurons could delay action initiation.7 Although the Pf is not required for instrumental learning (e.g., lever pressing for reward), it is critical for the regulation of behav- ioral flexibility, to adapt to changes in action-outcome contin- gencies.8,9 Moreover, activation of Pf projections to the subtha- lamic nucleus (STN) can restore movement during dopamine depletion.10 These results suggest the Pf plays a key role in ac- tion generation, but its contribution remains poorly understood.

AllowaySmithWatson14

Many thalamic nuclei innervate the striatum (Castle et al. 2005; Deschenes et al. 1995, 1996; Smith et al. 2004; Van der Werf et al. 2002), and whisker-sensitive regions in DLS receive most of their thalamic inputs from the medial posterior (POm) and parafascicular (Pf) nuclei (Smith et al. 2012).

The Pf nucleus projects to the striatum, where it innervates cholinergic interneurons (Berendse and Groenewegen 1991; Lapper and Bolam 1992). When activated by thalamostriatal inputs from Pf and other intralaminar nuclei, the cholinergic neurons initiate a pause in the activity of other striatal neurons (Ding et al. 2010). Presumably, this mechanism enables Pf and other intralaminar nuclei to interrupt ongoing striatal process- ing so that another behavior can be selected (Matsumoto et al. 2001; Thorn and Graybiel 2010; Van der Werf et al. 2002).

Consistent with this view, the Pf receives multimodal sen- sory inputs from the superior colliculus (Grunwerg and Krau- thamer 1992; Krout et al. 2001; Schultz et al. 2009), a brain region instrumental in directing attention to unexpected stimuli (Yamasaki and Krauthamer 1990). Although these collicular projections are known, whisker-evoked responses have never been recorded in the Pf nucleus, and the full extent of the thalamic targets of whisker-sensitive sites in the superior col- liculus are not well established.

The POm complex receives ascending inputs from the tri- geminal nuclei, and it responds to whisker stimulation (Chiaia et al. 1991; Masri et al. 2008; Sosnik et al. 2001; Veinante et al. 2000). Although POm is known for innervating sensorimo- tor cortex (Lu and Lin 1993; Ohno et al. 2012; Smith et al. 2012; Viaene et al. 2011; Wimmer et al. 2010), neurons in the posterior POm also send collateral projections to the striatum (Deschenes et al. 1995, 1998; Ohno et al. 2012; Smith et al. 2012). The POm could be important for controlling striatal processing, but no study has characterized the topographic specificity of its projections to the striatum to determine how it compares with thalamostriatal projections from Pf.

Our results demonstrate that whisker-sensitive sites in POm project to the DLS, but not to other striatal regions. This is significant because DLS is involved in the execution of highly repetitive behaviors that do not depend on rewarded outcomes (Balleine and O’Doherty 2010; Cromwell and Berridge 1996; Yin et al. 2004, 2006). Whisking behavior is consistent with this classification scheme, and the DLS receives dense inputs from the whisker representations in SI and primary motor (MI) cortex (Alloway et al. 2006; Hoffer and Alloway 2001).

By comparison, the lateral Pf projects to both the DLS and the dorsomedial striatum. This is significant because the dor- somedial striatum mediates goal-directed behaviors (Yin et al. 2005a, 2005b). Hence, compared with the relatively focused thalamostriatal projections from POm, the projections from Pf should have a more general function.

Rostrocaudal extent of POm and Pf projections. We found that thalamostriatal projections from Pf and the posterior POm differed in their rostrocaudal extent. Although projections from both thalamic regions overlapped substantially in DLS, the projections from the whisker-sensitive parts of POm were concentrated in the caudal DLS. By comparison, projections from Pf had a more extensive rostrocaudal distribution and were densest in striatal regions rostral to bregma.

The distribution of projections to the DLS from whisker- sensitive regions in POm matches the topography of corticos- triatal projections from SI cortex. Whereas SI barrel cortex innervates the caudal DLS, the SI forelimb region innervates more rostral parts of the DLS (Hoffer and Alloway 2001; Hoover et al. 2003). This topography is consistent with the fact that lesions in the rostral DLS interfere with behaviors that involve forelimb movements such as habitual bar-pressing and grooming behaviors (Cromwell and Berridge 1996; Yin et al. 2004, 2006). Given that the POm whisker area projects to caudal DLS regions that receive inputs from SI barrel cortex, it is parsimonious to expect that POm forelimb sites probably project to more rostral parts of DLS that receive inputs from the SI forelimb area (Hoover et al. 2003).

In primates, for example, the intralaminar nuclei respond to unexpected sensory stimuli but habituate when the same stimulus is repeated (Matsumoto et al. 2001).

Axon collaterals of individual Pf neurons form dense plex- uses of terminals in local parts of the striatal neuropil (De- schenes et al. 1995; Lacey et al. 2007). Our bulk tracer injections in Pf extend these findings by revealing an extensive network of Pf projections throughout the dorsomedial and dorsolateral striatum, which are associated with the execu- tion of goal-directed and stimulus-controlled behaviors, re- spectively.

Ultrastructural studies have shown that cholinergic interneu- rons in the striatum are innervated by the intralaminar thalamic nuclei, but not by corticostriatal projections (Lapper and Bo- lam 1992; Meredith and Wouterlood 1990). Activation of the intralaminar thalamus evokes a pause among tonically active cholinergic neurons that resembles the behavioral pause evoked by salient stimuli in awake, behaving monkeys (Api- cella 2007; Ding et al. 2010). Synchronous activation of cholinergic interneurons facilitates the indirect, but not the direct, striatal pathways (Ding et al. 2010; Threlfell et al. 2012), and this suggests that Pf activation could interrupt an ongoing behavior so that a new behavior can be selected that is more appropriate in the context of an unexpected stimulus (Thorn and Graybiel 2010).

We hypothesize that POm projections to the DLS transmit “ex- pected” somesthetic signals that accompany the motor ele- ments that are kinematically linked during the execution of highly repetitive behaviors such as grooming, whisking, or other fixed action patterns. In contrast to thalamostriatal pro- jections from Pf, which probably interrupt ongoing behavior, the projections from POm cooperate with corticostriatal signals from SI and MI to facilitate and maintain the automatic sequence of movements that comprise an ongoing, highly repetitive sensorimotor activity.

[rcoreilly]

JohanssonSilberberg20

Johansson, Y., & Silberberg, G. (2020). The functional organization of cortical and thalamic inputs onto five types of striatal neurons is determined by source and target cell identities. Cell Reports, 30(4), 1178-1194.e3. https://doi.org/10.1016/j.celrep.2019.12.095

The principal structures giving rise to thalamostriatal fibers in rodents are the centrolat- eral/centromedian (CL/CM) nucleus and the parafascicular (PF) nucleus (Ellender et al., 2013; Lacey et al., 2007; Mandelbaum et al., 2019; Sadikot et al., 1992; Smith and Wichmann, 2015). Both thalamic nuclei target striatal medium spiny neurons (MSNs), but only the PF additionally targets striatal interneurons (Lapper and Bolam, 1992; Rudkin and Sadikot, 1999; Sidibe ́ and Smith, 1999). Both thalamic nuclei target striatal medium spiny neurons (MSNs), but only the PF additionally targets striatal interneurons (Lapper and Bolam, 1992; Rudkin and Sadikot, 1999; Sidibe ́ and Smith, 1999).

Whole-cell recordings of MSNs revealed that optogenetic stimulation of PF afferents induced glutamatergic, monosynaptic EPSPs (Figures S2A, S2C, and S2F). Striatal re- sponses to thalamic activation differed in several aspects from those evoked by cortex: Thalamostriatal inputs evoked smaller EPSPs in D1-MSNs than in D2-MSNs (2.30 ± 0.37 mV in D1- MSNs versus 3.51 ± 0.43 mV in D2-MSNs, p = 0.0032, n = 23 pairs, N = 5; Figures 3E and 3F). EPSPs in FSIs were larger than in adjacent MSNs, but the difference was not as large as for cortical inputs (3.83 ± 0.81 mV in FSIs versus 2.20 ± 0.3 mV in MSNs, p = 0.0329, n = 19 pairs, N = 9; Figures 3E and 3F

Interestingly, our recordings revealed that thalamus constitutes the most reliable input to ChINs (100%, n = 20/20 ChINs), whereas LTSIs failed to respond (n = 0/21 LTSIs, 3.29 ± 0.60 mV in MSNs, n = 21 pairs, N = 7; Figures 3E and 3F). Nonetheless, the EPSP amplitudes in ChINs were significantly smaller

According to this, PF projects equally strongly to MSN, CHIn, but uniquely to CHIn

Our data show that ipsilaterally projecting M1 and S1 inputs differ both anatomically and functionally: while S1 afferents densely target the dorsolateral striatum and rarely contact LTSIs and ChINs, M1I afferents innervate not only the dorsolateral areas but also the dorsomedial striatum and frequently form functional synapses onto LTSIs and ChINs.

These anatomical findings support behavioral studies identifying the dorsolateral striatum as the neural substrate for sensorimotor integration, whereas the dorsomedial striatum acts as an associative terri- tory central to instrumental conditioning and behavioral flexibility (Devan and White, 1999; Sabol et al., 1985).

The short-term plas- ticity of S1I and M1I inputs to striatal neurons also varied in an input-specific manner. EPSPs evoked by stimulation of S1I depressed strongly in a use-dependent manner in MSNs, whereas M1I inputs exhibited slower use-dependent depres- sion. The overall depressing short-term plasticity of corticostria- tal inputs to MSNs is surprising given the rich literature reporting facilitatory corticostriatal synapses (Ding et al., 2008; Sciamanna et al., 2015). In these previous studies, however, it is less clear which excitatory inputs were activated because of the use of extracellular stimulation electrodes or optogenetic stimulation, without the spatial specificity obtained by local expression of ChR2.

Both cortical input structures excited FSIs more than MSNs, whereas LTSIs and ChINs received weaker and sparser inputs than MSNs. Our re- sults are in accordance with in vivo results from Sharott et al. (2012), in which the responses to electrical stimulation in motor cortex were most pronounced in PV-expressing interneurons, followed by MSNs and lastly by LTSIs and ChINs.

• FF inhib via FSIs is strong -- could increase that param presumably in the models.

Yet we found clear differences between M1I and M1C inputs, including a lack of synaptic inputs from M1C to ChINs.

In contrast, LTSIs were most reliably innervated by M1 afferents from both hemispheres, with M1C providing the strongest input to LTSIs of all inputs tested.

This is in line with a previous study that showed that co-activation of M1 and secondary motor cortex can strongly activate LTSIs (Iba ́ n ̃ ez-Sandoval et al., 2011). LTSIs release GABA and co-release several neuromodulators during high-frequency activity, including SOM, nitric oxide, and neuropeptide Y (Kawaguchi et al., 1995). These neuromodula- tors are key controllers of striatal excitability, synaptic plasticity, and dopamine release, which puts LTSIs in an excellent position for assisting motor learning (Adewale et al., 2007; Calabresi et al., 1999; Galarraga et al., 2007).

• TODO: need to investigate LTSIs more!

Thalamostriatal inputs were found to excite D2-MSNs more than adjacent D1-MSNs and to evoke facilitating responses in both types of projection neurons. The responses involved a strong NMDA component that was frequently activated at rest, indi- cating strong dendritic excitation. The high NMDA to AMPA ratio expressed by MSNs specifically at thalamostriatal synapses, but not corticostriatal synapses, is consistent with the comparatively slow kinetics we observed for thalamic-evoked EPSPs. Further- more, these results are in agreement with several previous studies, some of which used the same genetic approach for tar- geting PF (Ellender et al., 2013; Hunnicutt et al., 2016; Mandel- baum et al., 2019; Spruston et al., 1994).

Moreover, Mandelbaum and colleagues provided evidence for the heterogeneity of PF by showing that this nucleus consists of at least three subpopulations that differ on genetic, anatomical, and physiological levels (Mandelbaum et al., 2019).

Compared with cortical inputs, the relative strength of PF input to FSIs was smaller, indicating that PF is less effective in activating the FSI-MSN feedforward circuitry. This is consistent with anatomical studies that have shown that PF inputs to FSIs are sparse, whereas cortical affer- ents target FSIs massively (Kita, 1993).

• FS is clearly modulatory / learning related via NMDA, etc

The discovery of novel glutamatergic inputs arising in the pedunculopontine nu- cleus raises the question whether these inputs follow similar rules of functional organization (Assous et al., 2019; Klug et al., 2018).

[rcoreilly]

MandelbaumTarandaHaynesEtAl19

Mandelbaum, G., Taranda, J., Haynes, T. M., Hochbaum, D. R., Huang, K. W., Hyun, M., Venkataraju, K. U., Straub, C., Wang, W., Robertson, K., Osten, P., & Sabatini, B. L. (2019). Distinct Cortical-Thalamic-Striatal Circuits Through the Parafascicular Nucleus. Neuron, 102(3), 636-652.e7. https://doi.org/10.1016/j.neuron.2019.02.035

Within TH, the neighboring parafascicular (PF) and centrome- dian (CM) nuclei project heavily to STR (Smith and Parent, 1986; Berendse and Groenewegen, 1990; Wall et al., 2013), unlike typical thalamic nuclei that primarily interact with CTX (Sherman and Guillery, 2013).

In primates, projections from PF/CM to STR are anatomically organized into multiple functionally distinct output channels (Steiner and Tseng, 2016; Sadikot and Rymar, 2009), a finding that seems to grossly hold in cats and rats (Gime ́nez-Amaya et al., 2000; Jones, 2007).

Conversely, in rodents, the lack of histological demarcations within and between PF and CM as well as the small size and close packing of TH nuclei have led researchers to treat mouse PF as anatomically uniform and cellularly homogeneous (Parker et al., 2016; Kato et al., 2011; Aceves Buendia et al., 2017; Assous et al., 2017; Choi et al., 2018).

Here, we deconstruct the mouse PF. Using quantitative anatomical approaches, we reveal that PF has the highest density of striatum-projecting neurons among sub-cortical structures. We find that PF contains anatomically, transcriptionally, and physiologically distinct neuronal populations with topographi- cally organized projections to STR and to and from CTX. Each PF subregion and neuron class influences distinct striatal regions and cell classes through independent and parallel channels that carry information principally from limbic, associative, or somato- sensory cortical regions.

Thus, despite its small size and lack of cytoarchitecturally evident substructure, mouse PF contains distinct and topographically organized projections to STR that share similar organizational features of PF and CM in larger species (Jones, 2007; Gime ́ nez-Amaya et al., 2000).

In primates, PF neurons inner- vate prefrontal CTX, whereas the histologically distinct CM neu- rons target motor and premotor areas (Parent and Parent, 2005). In rats, reconstructed PF cells project to STR and several cortical regions (Descheˆ nes et al., 1996). However, studies of mouse PF, albeit using manipulations that could not specifically target this small nucleus, suggest that it does not project to CTX (Oh et al., 2014).

Thus, STR- projecting PF neurons differentially innervate cortical regions: mPF and cPF innervate mainly limbic structures, whereas cPF also targets associative areas, such as MOs, and lPF selectively targets somatosensory cortical areas in the posterior CTX.

In summary, projections from PF to STR and CTX are orga- nized into parallel CTX-PF-CTX and CTX-PF-STR motifs: subre- gions of PF and CTX are reciprocally connected and these linked subregions each target the same STR domain.

Summary:

• mPF neurons expressed Pdyn, the precursor protein for the K-opioid receptor agonist dynorphin, project to matrix compartments of mSTR and to limbic CTX (e.g., agranular insular, infra-, and pre-limbic CTX), and receive bilateral input from layer 5 in PFC. mPF/STR projection neurons have higher input resistance, lower capacitance, and higher resting potential relative to those in central and lateral aspects of the PF. In addi- tion, these neurons have prominent sag potentials, a feature predicted from the single-cell transcriptional analysis showing expression of the Hcn1 gene.

• In cPF, neurons express Tnc, proj- ect to dmSTR and to limbic and associative regions of CTX (e.g., to infralimbic, secondary motor, and gustatory CTX), and receive input from layer 5 of associative areas (secondary motor).

• Lastly, in lPF, neurons express Spon1, project to dlSTR and predomi- nantly somatosensory regions of CTX (e.g., primary and second- ary somatosensory CTX), and receive input from layer 5 of SSp.

• All cell classes in PF have high connectivity to SPNs but differ in their innervation of STR interneurons. Neurons do not intercon- nect across regions in PF, suggesting that PF subregions do not intermix their incoming cortical and midbrain signals through local inhibitory or excitatory connectivity.

The anatomical organization that we describe for mouse PF/ STR is present in other species (Gime ́ nez-Amaya et al., 2000; Jones, 2007). Primate PF/CM is also subdivided into 3 regions that preferentially innervate motor CTX (lateral CM), sensory mo-tor STR (medial CM), and associative limbic STR (PF; Sadikot and Rymar, 2009).

In primates, CM and PF are also distinguished based on cell density and size (Jones, 2007), vulnerability to disease (Henderson et al., 2000), as well as in vivo firing patterns (Matsumoto et al., 2001) and have been proposed to have different functions (Glimcher and Lau, 2005; Smith et al., 2014).

In rats, lPF projects to dlSTR and mPF projects to dmSTR (Be- rendse and Groenewegen, 1990), which is a similar, albeit a simplified version of the relationship observed between CM, PF, and STR in primates. Nevertheless, no region in the rat TH has been defined as CM due to the lack of a clear histological boundary.

• On this basis, we can skip tracking down details about CM, and just think in terms of these 3 parallel loops in a generalized PF / CM thalamus.

Thalamic nuclei typically form reciprocal connections with CTX by receiving input from and projecting to a single cortical region or receiving input from one region and project to another (Sher- man, 2016). Thalamic nuclei receive modulatory inputs from layer 6, whereas higher-order thalamic nuclei also receive inputs from layer 5 (Harris and Shepherd, 2015; Sherman, 2016; Jeong et al., 2016). These CTX-TH-CTX circuits have been proposed to have two functions. First, via recurrent excitation, they maintain persistent activity in CTX, as has been shown for projections from motor TH to the anterior lateral motor region of CTX (Guo et al., 2017), which is thought to be necessary for working mem- ory (Bolkan et al., 2017; Halassa and Kastner, 2017). Second, other CTX-TH-CTX circuits have a triangular motif, in which a cortical region targets a second cortical area and a thalamic nu- cleus that also projects to the second cortical region. This motif, for example, is seen in Pulvinar outputs to visual cortices and has been proposed to transmit and synchronize signals about atten- tional priorities between directly connected cortical regions (Saalmann et al., 2012). These canonical principals of organiza- tion have not been examined fully for ILM TH and its interactions with CTX.

PF -> CTX projections align with CTX -> PF projections, sug- gesting a corticothalamic recurrent network through PF. Howev- er, unlike typical TH nuclei, the main output of PF is to STR and not to CTX. For this reason, and analogous to the triangle atten- tional motif described above, we propose that CTX-PF-CTX circuits facilitate and shape the striatal output of the associated cortical region. PF can integrate information from CTX with that from sub-cortical nuclei (such as SC and SNr) to facilitate correct action selection in an ongoing sensorimotor context. Because we find that PF neuron classes are not interconnected in PF, these networks of activity can act relatively independently of each other, although other potentially diffuse external inputs, such as from the thalamic reticular nucleus, may coordinate ac- tivity across PF subregions

• Key points: CTX -> PF -> CTX is closed loop, not open. Not doing predictive learning. Rather, doing maintenance / gating. And PF -> CTX is not major -- main output of PF is to striatum!

We find that PFC projects bilaterally to mPF, whereas SSP/ lPF or MOs/cPF projections are strictly IPSI, highlighting the potential different functions of the PF subclasses characterized here—laterality may be important to maintain sensory and asso- ciative circuits although perhaps a global limbic signal may need to be dispersed across both hemispheres. PT cortical neurons do not project to contralateral STR (Harris and Shepherd, 2015); thus, the bilateral PFC/mPF projections may transmit PT- related activity signal near synchronously to both STR via PF without recruitment of the bilateral intratelencephalic (IT)-type cortical projections.

[rcoreilly]

LeeWangSabatini20

Lee, J., Wang, W., & Sabatini, B. L. (2020). Anatomically segregated basal ganglia pathways allow parallel behavioral modulation. Nature Neuroscience, 23(11), Article 11. https://doi.org/10.1038/s41593-020-00712-5

Two models of BG function predict radically different architec- tures and topographical organization of BG circuits and outputs. One model posits that BG circuits mediate learning of arbitrary stimulus–response associations4,5. In this model, stimulus identity is encoded in the input to striatum which activates specific striatal neurons that generate a specific behavior. If the behavior leads to a reward, dopamine is thought to reinforce that stimulus–response association6,7. In order for BG to be able to reinforce arbitrary stim- ulus–response associations, projections from striatum to BG output nuclei need to form an all-to-all connected network, allowing cross talk between each striatum pathway as necessary for contextual information to control any potential motor output. In contrast, a different model posits that each BG pathway controls only a specific set of behaviors8. In this model, each pathway remains anatomi- cally segregated and innervates a unique set of target brain regions, endowing each BG pathway with exclusive control over specific aspects of behavior.

we show that the outputs of the BG are subdivided into anatomically and functionally segregated pathways. Anatomically, SNr projections to the parafascicular nucleus (Pf), ventromedial thalamus (VM) and SC are topographically organized. These three nuclei, in turn, project back to the striatal region from which they receive input, suggesting the existence of at least three closed BG loops.

Using tapered optical fibers to optogenetically activate focal regions of striatum, we reveal functional maps for licking and locomotor turn- ing, arising from ventrolateral striatum (VLS) and ventromedial/ dorsomedial striatum (VMS and DMS, or MS), respectively.

Our results support a model of BG in which multiple seg- regated pathways project to distinct anatomical regions to control specific behaviors in parallel.

Together, these results indicate that anatomically segregated BG loops via Pf are functionally closed and synaptically segregated.

anterior cingulate cortex (ACA), forelimb motor cortex (fM1), and tongue and jaw motor cortex (tjM1) in the same mouse. These frontal regions project to DMS, DLS and VLS, respectively29,30.

• It is not mentioned here but VLS interconnects principally with the insular cortex (IC) (FudgeBreitbartDanishEtAl05, BerendseGalis-DeGraafGroenewegen92) which is critical for linking sensory cues with consummatory behavior, along with VLS itself (Kusumoto-YoshidaLiuChenEtAl15, NatsuboriTsutsui-KimuraNishidaEtAl17, ParkesBradfieldBalleine15, WittmannFouragnanFolloniEtAl20 -- global reward state).

The striatum is often described as functionally delineated along the medial–lateral axis, with DMS and DLS involved in goal-directed and habitual behaviors, respectively50. Our work suggests that this dichotomy might arise, partially, from distinct anatomical projec- tions that underlie differential behavioral control. Interesting, many previous studies of DMS and DLS used lever-pressing tasks that require forelimb movement and locomotion. Thus, it is possible that goal-directed behavior and habitual behavior are two behavioral modes that specifically engage locomotion (DMS-based behavior) or forelimb motor sequences (DLS-based behavior), respectively, in a lever-press task.

Complex behavioral sequences require brain circuits to ulti- mately integrate information across behavioral space. Our work suggests that BG might not be a site of integration, and that regions outside of the BG, such as cortex or SC, might serve this function.

es- tablishes a principle of circuit organization by which SNr neu- rons concomitantly affect cortical and brain stem systems. Branching ascending-descending collateral projections might ensure coordinated gating of cortical and brain stem circuitry and are a practical consideration for interpretation of experi- ments in which SNr neurons are labeled retrogradely from a single target.

[rcoreilly]

Cerebellum -> Thalamus -> Striatum pathway via central-lateral (CL) thalamus

IchinoheMoriShoumura00

Ichinohe, N., Mori, F., & Shoumura, K. (2000). A di-synaptic projection from the lateral cerebellar nucleus to the laterodorsal part of the striatum via the central lateral nucleus of the thalamus in the rat. Brain Research, 880(1), 191–197. https://doi.org/10.1016/S0006-8993(00)02744-X

"LDS" == DLS

The morphological features of cerebellar terminals in the CL were very similar to those previously reported for the VL, red nucleus (RN), Pf, central medial nucleus (CM) and zona incerta (ZI) in the rat [3]. Cerebellar terminals in these brain regions were large and made multiple asymmetric synaptic contacts with dendritic shafts and spine-like appendages. Aumann et al. [3] also pointed out that cerebellar terminals predominately contacted proximal dendrites in the main target nuclei (the VL and RN) of the cerebellum. In the VL, 55% of these inputs occur on proximal dendrites and only 3% on somata; in the RN, 65% are found on proximal dendrites. By contrast, neurons in the minor target nuclei such as the Pf, CM and ZI receive only a small portion of cerebellar inputs on their proximal dendrites; in the Pf 15% occur on proximal dendrites while in the CM there are 8%. In the ZI, 13% are found on proximal dendrites. This may mean that cerebel- lar synaptic transmission is very secure through such favorably positioned contacts on the VL and RN neurons. In fact, electrophysiological studies have shown that cerebellar input can evoke large unitary excitatory post- synaptic potentials having a fast rising phase on VL and RN neurons [24,25]. We have shown that the proportion of cerebellar boutons contacting proximal dendrites of CL neurons is similar to that of neurons in the cerebellar main target nuclei (see above). Thus, potent excitatory cerebello- thalamic transmission can occur in the CL, suggesting the CL is a significant relay site for the information from the LCN to the LDS.

Cortical fibers synapse mainly with the heads of spines, whereas thalamic terminals preferentially synapse with dendritic shafts [9,23]. Furthermore, cortical or thalamic afferents have complementary input upon subclasses of striatal interneurons.

Thus, the LCN could modulate LDS function simultaneously at least in two different ways, through the cerebello-thalamo (mainly VL)-motor cortico- striatal pathway and the cerebello-thalamo (mainly CL)- striatal pathway including the di-synaptic LCN-CL-LDS link shown in this study. The physiological significance of this LCN-CL-LDS link awaits elucidation.

HoshiTremblayFegerEtAl05

Hoshi, E., Tremblay, L., Féger, J., Carras, P. L., & Strick, P. L. (2005). The cerebellum communicates with the basal ganglia. Nature Neuroscience, 8(11), 1491–1493. http://www.ncbi.nlm.nih.gov/pubmed/16205719

The cerebellar nuclei are known to project to these thalamic regions6, and there is evidence that these thalamic nuclei (especially intralaminar nuclei) project to the striatum (12 = SmithRajuPareEtAl04).

Our findings support a previous study (13 = Ichinohe et al, 2000) that demonstrated a disynaptic connection between the rat cerebellum and the striatum. The present results extend these observa- tions to a nonhuman primate and demonstrate four new findings: (i) the disynaptic projection to the striatum originates from both the motor and nonmotor domains of the dentate, (ii) the striatum also receives less substantial inputs from fastigial and interpositus, (iii) the projection from the dentate to the striatum connects with medium spiny stellate cells that innervate GPe and thus influences the so-called ‘indirect’ pathway of basal ganglia processing14 and (iv) the number of dentate neurons that influence localized portions of GPe is comparable to the number of dentate neurons that influence some areas of posterior parietal and prefrontal cortex9–11.

SmithRajuPareEtAl04

Smith, Y., Raju, D. V., Pare, J.-F., & Sidibe, M. (2004). The thalamostriatal system: A highly specific network of the basal ganglia circuitry. Trends in Neurosciences, 27(9), 520–527. http://www.ncbi.nlm.nih.gov/pubmed/15331233

The main source(s) of thalamostriatal projections are intralaminar thalamic nuclei, but substantial inputs from midline and specific relay nuclei have also been described in rats [7–11], cats [12,13] and monkeys [14–18] (Figure 1). The rostral intralaminar complex comprises the central medial, paracentral and central lateral nuclei, whereas the caudal intralaminar group includes the centromedian (CM)–parafascicular (PF) nuclear complex in primates. Although the CM is not clearly delineated in rodents, the lateral part of the PF is considered the homolog of the primate CM, whereas the medial PF displays strong similarities with the PF proper in monkeys [3,7]. The topography of thalamostriatal projections from these nuclei is depicted in Figure 1.

• Key point: Rostral = CL, Caudal = PF / CM. Thus, cerebellum -> CL -> Striatum is a separate pathway from SNr -> PF / CM -> Striatum.

Figure 1. Rostrocaudal striatum (Str), showing the pattern of distribution of thalamic inputs from caudal intralaminar thalamic nuclei in monkey (a) and from rostral intralaminar, caudal intralaminar and midline thalamic nuclei in rat (b). Striatal territories receive inputs from the corresponding color-coded thalamic nuclei (middle boxes). The lateral centromedian (CMl) nucleus in monkey projects to the primary motor cortex but not to the striatum. In rats, projections from the central medial (CeM) and medial parafascicular (PFm) nuclei terminate along the rostrocaudal extent of the medial third of the caudate (Cd)–putamen (Put) complex, whereas inputs from the lateral parafascicular (PFl) nucleus innervate profusely the lateral sector of the dorsal striatum. Projections from central lateral (CeL) and paracentral (PC) nuclei occupy the central core of the dorsal striatum. Projections from the parataenial nucleus (not shown) provide a substantial innervation of the central portion of the rostral part of the nucleus accumbens (Acc). Inputs from intermediodorsal (Imd) nucleus are concentrated in the core of the nucleus accumbens rostrally, and caudally they invade the ventromedial part of the caudate–putamen complex and a striatal region along the dorsal border of the globus pallidus (GP). Anteroposterior stereotaxic coordinates that these drawings correspond to are indicated below each section. Additional abbreviations: A, anterior; AC, anterior commissure; Breg, bregma; CC, corpus callosum; CMm, medial centromedian nucleus; GPe, globus pallidus, external segment; GPi, globus pallidus, internal segment; IC, internal capsule; LV, lateral ventricle; PF, parafascicular nucleus; PFdl, dorsolateral parafascicular nucleus; PV, paraventricular nucleus; Th, thalamus. Rat data are reproduced, with permission, from Ref. [8]

In monkeys, organization of thalamostriatal projec- tions from the caudal intralaminar nuclei has been studied in detail, but little is known about the topography of the rostral intralaminostriatal pathways.

In rats, the midline nuclear group comprises the paraventricular, parataenial, interanteromedial, intermediodorsal, rhomboid and reuniens nuclei. Projections from these nuclei are mainly confined to the ventral striatum but they also provide significant inputs to dorsal striatal regions [7,8] (Figure 1).

• VS versions of PF/CM/CL are midline!

Overall, retrograde double labeling studies agree that a substantial proportion of neurons in the rostral intra- laminar nuclear group and some specific thalamic nuclei (the ventral anterior and ventral lateral complex, and the mediodorsal nucleus) provide axon collaterals to both the striatum and the cerebral cortex, whereas thalamostriatal and thalamocortical neurons are largely segregated in the caudal intralaminar CM–PF nuclear complex [6,23,24].

However, single- cell-filling injection data suggest that most, if not all, PF neurons that project to the caudate–putamen complex send sparse collaterals to the cerebral cortex in rats [25]. This pattern is opposite from that displayed by projec- tions from the central lateral nucleus, which provide scarce, loosely organized long varicose processes in the striatum but form dense patches of terminals in the rat cortex [11,26].

In contrast to cortical inputs, which terminate predominantly on the heads of dendritic spines [27], the vast majority of thalamic terminals from caudal intra- laminar nuclei form asymmetric synapses on dendritic shafts of medium-sized projection neurons in rats and monkeys [17,27–29] (Figure 2; Table 1). However, studies in rats implanted with striatal grafts demonstrate that inputs from rostral intralaminar nuclei target

Projections from CM form synapses more frequently with striatofugal neurons that project to the internal versus the external globus pallidus in monkeys, which suggests that CM modulates preferen- tially the direct output pathway of the sensorimotor striatum in primates [29] (Figure 2; Table 1).

• Note: this is contradicted by more recent data above showing overall a balanced distribution on Go vs. No.

By contrast, thalamic inputs from CM and dopaminergic terminals are not in close proximity on the same postsynaptic targets in the monkey striatum [28] (Table 1). Together, these anatomical arrangements suggest that dopaminergic afferents are located to subserve more specific modulation of afferent cortical inputs than afferent thalamic inputs in the sensorimotor territory of the primate striatum [28].

• Thalamic projections don't learn! they are plus phase drivers..

Ventral motor nuclei and the CM–PF complex are the main thalamic targets of GPi efferents [6,19]. The majority of GPi cells that project to the ventrolateral nucleus send axon collaterals to the CM–PF complex [6]. In turn, both the ventral motor and CM–PF nuclear groups send highly topographic and specific inputs to different functional territories of the striatum [15,17,20,29] (Figure 1).

Although the ventral anterior and ventral lateral complex is commonly seen as the main source of basal ganglia outflow to the cerebral cortex, it also provides substantial and highly topographic projections to the striatum [4,6,9,15,21]. Therefore, the ventral motor thalamic nuclei should not be seen solely as a relay for basal ganglia outflow to the cerebral cortex but, rather, as a group of thalamic nuclei that also provide a significant positive feedback to the sensorimotor striatum (Figure 4).

• TODO: follow up on VA / VL -- this is not reflected in other recent papers above.

In addition to the striatum, CM–PF projections innervate other basal ganglia nuclei, including the subthalamic nucleus (STN), the globus pallidus and the substantia nigra [10,11,25,26,59]. In monkeys, the thalamosub- thalamic projection is functionally organized so that sensorimotor neurons in CMm terminate preferentially in the ‘motor-related’ dorsolateral part of the STN, whereas limbic and associative neurons in PF project almost exclusively to the medial ‘limbic-related’ region of the STN [23]. A similar topographical organization was also found for PF–STN projections in rats [60].

• TODO: add Thalamus -> STN

Two main functional characteristics of CM–PF neurons were disclosed in monkeys [63]. First, CM–PF neurons have multimodal properties – that is, they respond to a large variety of sensory stimuli (auditory, visual and somatosensory) presented either in or outside sensorimotor conditioning tasks. Second, CM–PF neurons are temporally tuned – that is, they can generate in a timely manner discrete and coherent responses to a wide variety of sensory stimuli. On the basis of their latency and pattern of responses to sensory stimuli, CM–PF neurons have been categorized into two main populations, namely those that display short-latency facilitatory responses (SLF neurons) and those that display long-latency facilitatory responses (LLF neurons) to sensory events [63]. These two populations are largely segregated in the CM–PF complex, SLF neurons being mainly found in PF whereas LLF are particularly abundant in CM.

However, CM–PF inputs are required for the sensory responses of TANs that are acquired through sensorimotor learning. Inactivation of CM–PF decreases the characteristic pause and subsequent rebound facili- tation, but does not affect the early short-latency facilitation, of TANs in response to sensorimotor conditioning [63,67].

• 63 = MatsumotoMinamimotoGraybielEtAl01

BostanStrick18

Bostan, A. C., & Strick, P. L. (2018). The basal ganglia and the cerebellum: Nodes in an integrated network. Nature Reviews Neuroscience, 19(6), Article 6. https://doi.org/10.1038/s41583-018-0002-7

ShepherdYamawaki21

Shepherd, G. M. G., & Yamawaki, N. (2021). Untangling the cortico-thalamo-cortical loop: Cellular pieces of a knotty circuit puzzle. Nature Reviews Neuroscience, 22(7), 389–406. https://doi.org/10.1038/s41583-021-00459-3

Beautiful pics of the thalamocortical circuits! Very nice updated summary of TC circuits.

Key points:

• PO = posterior nucleus of the thalamus -- need to learn more about that.
• M1 <-> VL and S1 <-> VB are classic sensory / pulvinar style (core) predictive learning, top-down attentional modulation: CT prediction, bottom-up efferent-copy / cerebellum VL drivers. notably no PT -> VL.
• M1 <-> VM / PO is matrix modulatory, BG gated, recurrent bidirectional. significant PT -> VM (matrix) input
• PFC <-> VM and MD is notably distinct from M1: core and matrix basically the same -- VM is bg gated maintenance / update, MD has both VS gated and predictive learning; PT -> core MD

• [ALM]: The MD also receives some input from PT (and probably CT) axons, and may project weakly to the ALM. Nevertheless, the ALM appears to differ from M1 and S1 in that its main source of thalamic input is from a matrix­type nucleus (the VM).

• Key point: no motor plus phase drive input from VL -> ALM

GaoDavisThomasEtAl18

Gao, Z., Davis, C., Thomas, A. M., Economo, M. N., Abrego, A. M., Svoboda, K., De Zeeuw, C. I., & Li, N. (2018). A cortico-cerebellar loop for motor planning. Nature, 563(7729), Article 7729. https://doi.org/10.1038/s41586-018-0633-x

The ALM projects to the cerebellum via the basal pontine nucleus23,24.

• CN (DCN) neurons have tonic firing, same kind of disinhibitory dynamic as SNr?

Fastigial lesions impaired the initiation of contralesional licking (Fig. 1d) and delayed the initiation of contralesional licking following the ‘go’ cue (Fig. 1e). After the initiation of movement, licking duration and frequency were unaffected (Fig. 1e, Extended Data Fig. 1b, c). Dentate lesions (around 1.5 mm away from the fastigial nucleus) did not change behavioural choice and movement execution (Fig. 1d).

Fastigial photoactivation during the early sample period had no effect on behavioural choice (P > 0.05 for either trial type, bootstrap; Methods). Photoactivation during the delay and response periods both resulted in incorrect choices (Fig. 1g, h), which were induced even when the photo- stimuli ceased 800 ms before the movement. Photoactivation during the delay period biased future movements to either the contralateral or ipsilateral direction in individual mice (Fig. 1h, Extended Data Fig. 1e). Photoactivation immediately before the onset of movement (response period) produced a contralateral bias (Fig. 1h), which is opposite to the directional bias induced by fastigial lesions. The distinct patterns of bias induced by photoactivation in the delay and in the response periods suggest temporally specific roles for the fastigial nucleus in driving movement planning and initiation.

Photoinhibition [of purkinje] during the late sample and delay periods resulted in incorrect behavioural choices, with a larger effect induced in the delay period (Fig. 1k, Extended Data Fig. 2g, h). Photoinhibition during the delay period biased future movements to either the contralateral or the ipsilateral direction in individual mice (Extended Data Fig. 2g, h). Photoinhibition during movement initiation (response period) pro- duced an ipsilateral bias (Fig. 1k), consistent with that observed after fastigial lesions, and opposite from the bias induced by fastigial photo- activation.

Fastigial projections most overlapped with parts of the thalamus that project to the ALM (VM, VA/VL), whereas dentate projections targeted an adjacent region that was more dorsal and lateral (VA/VL)

Our results show that the cerebellum is critical for the coding of future movement in the frontal cortex. Considering its roles in the timing of movement26, the cerebellum may provide an urgency drive that is necessary for the emergence of selectivity in the ALM thalamo-cortical loop17 in the preparation of movement, which is consistent with attractor models of motor planning30.

EconomoViswanathanTasicEtAl18

Economo, M. N., Viswanathan, S., Tasic, B., Bas, E., Winnubst, J., Menon, V., Graybuck, L. T., Nguyen, T. N., Smith, K. A., Yao, Z., Wang, L., Gerfen, C. R., Chandrashekar, J., Zeng, H., Looger, L. L., & Svoboda, K. (2018). Distinct descending motor cortex pathways and their roles in movement. Nature, 563(7729), Article 7729. https://doi.org/10.1038/s41586-018-0642-9

Intratelencephalic neurons in layers (L) 2–6 receive input from other cortical areas and excite pyramidal tract (PT) neurons5–7. PT neurons, the somata of which define neocortical L5b8, link the motor cortex with premotor centres in the brainstem and spinal cord9,10 and directly influence behaviour10–12. PT neurons also project to the thala- mus6,13,14. Preparatory activity requires reverberations in a thalamocor- tical loop15. Consistent with roles in both the planning and initiation of movement, PT neurons are structurally heterogeneous14,16,17 and show diverse activity patterns, including preparatory activity and movement commands18–20.

Here we show that PT neurons in the mouse motor cortex comprise two cell types with distinct gene expression and axonal projections. We refer to these cell types as PTupper and PTlower neurons, reflecting their distributions in different sublaminae in L5b. PTupper neurons project to the thalamus, which forms a feedback loop with the motor cortex. PTlower cells project to premotor centres in the medulla. Cell-type- specific extracellular recordings in the anterior lateral motor cortex (ALM) during a delayed-response task suggest that PTupper neurons are preferentially involved in motor planning, whereas PTlower neurons have roles in movement execution.

• TODO: PTupper == PTMaint and PTPred; gets VL/VA thal and interconnects with DCN -- for planning details of movement, per Gao et al (2018) above; PTlower = BG gated (VM) for movement initiation. Note that exact VL/VA/VM delineation is unlikely to be super precise.

This pattern was similar across the primary and secondary motor cortex (Extended Data Fig. 3). Retrograde labelling from the superior colliculus and pons labelled PT neurons in both L5b sublaminae (Fig. 2c). Few neurons were co-labelled by injections into the thalamus and medulla (2.2%; 22 out of 984 cells labelled from either location), whereas many neurons were co-labelled by injections into the thalamus and superior colliculus (77.1%; 687 out of 890 thalamus-labelled cells), as expected from axonal reconstructions.

PTlower project to premotor centres in the medulla and spinal cord, with few collaterals in the basal ganglia and thalamus, suggesting a role in movement execu- tion.

A substantial proportion of PTlower neurons dis- played spike bursts (18.8%), which were rare among PTupper cells (3.3%; P = 0.006, Fisher’s exact test; Extended Data Fig. 5c).

In this projection, selectivity was larger and more consistent across trials in the PTupper population compared to the PTlower population (Fig. 5b). Furthermore, selectivity in the PTupper population remained constant throughout the sample and delay epochs until the go cue, implying retention of decision-related information. By contrast, selectivity in the PTlower population decayed rapidly along CDearly (no significant selectivity in 400 ms preceding go cue; P = 0.34, bootstrap). This decay did not reflect a lack of selectivity in the PTlower population: along a different direction in activity space, which maximized selectivity at the end of the delay epoch, CDlate, selec- tivity was substantial in both cell types18 (Extended Data Fig. 7a, b; PTupper: 40 out of 61 neurons selective; PTlower: 44 out of 69).

PTlower neruons were strongly modulated at the go cue, at the offset of movement, or both -- features rarely observed in PTupper cells.

LiChenGuoEtAl15

Li, N., Chen, T.-W., Guo, Z. V., Gerfen, C. R., & Svoboda, K. (2015). A motor cortex circuit for motor planning and movement. Nature, 519(7541), Article 7541. https://doi.org/10.1038/nature14178

• just shows that PTlower brainstem projecting neurons drive lateralized motor commands, while IT (and likely PTupper) are bilateral.

Although we studied ALM in the context of licking, ALM and nearby motor cortical areas also have roles in planning other movements4,37,38 (unpublished data).

Pyramidal tract neurons are downstream of intratelencephalic neurons, which show contralateral and ipsilat- eral selectivity with little contralateral bias. This suggests that during movement planning distributed preparatory activity in intratelence- phalic neuron networks is converted into a movement command in pyramidal tract neurons (‘output-potent’ activity)39, which ultimately triggers directional movements.

DacreColliganClarkeEtAl21

Dacre, J., Colligan, M., Clarke, T., Ammer, J. J., Schiemann, J., Chamosa-Pino, V., Claudi, F., Harston, J. A., Eleftheriou, C., Pakan, J. M. P., Huang, C.-C., Hantman, A. W., Rochefort, N. L., & Duguid, I. (2021). A cerebellar-thalamocortical pathway drives behavioral context-dependent movement initiation. Neuron, 109(14), 2326-2338.e8. https://doi.org/10.1016/j.neuron.2021.05.016

Thus, individual cerebellar nuclei provide differing contributions to movement control, where DN/IPN likely convey motor timing signals via thalamus to cortex in order to initiate and modify ongoing movements (Kurata, 2005; Nashef et al., 2018; Thach, 2014).

In rodents, cerebellar nuclei project to different regions of ventral thalamus. The fastigial nucleus primarily targets ventro- medial (VM), while DN/IPN target the anteromedial (AM), and ventral anterolateral (VAL) subdivisions (Angaut et al., 1985; Gor- nati et al., 2018; Haroian et al., 1981; Kuramoto et al., 2009; Teune et al., 2000). DN/IPN axon terminal fields overlap substan- tially displaying morphological and functional characteristics consistent with strong feedforward driver inputs, such as large synaptic boutons (Aumann and Horne, 1996a, b; Aumann et al., 1994; Gornati et al., 2018) and large unitary responses (Gornati et al., 2018; Sawyer et al., 1994; Scha€fer et al., 2021). Cerebellar input drives short-latency spiking in thalamic neurons that project to superficial and deep layers of motor cortex (Hooks et al., 2013; Kuramoto et al., 2009; Scha€fer et al., 2021), trans- forming output via top-down excitation through layer 2/3 (Weiler et al., 2008) or direct excitation of layer 5 (Hooks et al., 2013; Sauerbrei et al., 2020). Key remaining questions are whether ventral motor thalamus plays a role in movement initiation and whether this is dependent on cerebellar input.

Using imaging, electrophysiology, and gain- and loss-of-function experiments, we investigated how an audi- tory go cue transforms thalamic and motor cortical activity patterns during movement initiation. Population responses in DN/IPN-recipient regions of motor thalamus were dominated by a time-locked increase in activity immediately prior to move- ment initiation, providing a fixed-latency feedforward timing signal to motor cortex. Consistent with this view, membrane po- tential dynamics of layer 5B projection neurons matched pre- movement timing of thalamic activation, while suppressing cerebellar or thalamic output blocked movement initiation. Conversely, photostimulation of DN/IPN or recipient thalamic regions triggered movement initiation, but in a context-depen- dent manner.

Since both DN and IPN are implicated in motor timing and send glutamatergic projections to ventral motor thalamus (Aumann and Horne, 1996b; Bosch-Bouju et al., 2013; Gornati et al., 2018; Kuramoto et al., 2009), we sought to define the region of thalamus that receives input from DN/IPN and projects to the caudal forelimb area (CFA) of motor cortex using a dual labeling strategy (Figure 1F).

• DN, IPN are glut, not gaba -- providing "specific info" (vs. BG gating) consistent with VAL / AM thalamus being core, not matrix.

A region of dense overlap centered on VAL and AM nuclei, with sparse colocalization in the ventral postero- medial nucleus (VPM). We found no overlap in the ventromedial nucleus (VM), which primarily receives input from the fastigial nucleus (Gao et al., 2018; Gornati et al., 2018) (Figures 1G and S1A–S1E).

• fastigial is licking per above papers, whereas they are looking at reaching / forelimb -- but is this part of VM core or matrix like?

First, thalamic activity is uncoupled from the go cue rising immediately before movement onset. In this regard, rapidly increasing thalamic activity dictates the time of movement initiation (model i). ... Response profiles were consistent trial to trial and across mice, indicative of a reliable motor timing signal that is temporally uncoupled from the auditory go cue (i.e., model i) (Figures 2E–2G).

• thalamus activity determines Go

As expected, blocking MThDN/IPN activity reduced layer 5B firing rate changes and the number of successful push trials (Figure 4G).

Together, these data suggest that the DN/IPN thalamocortical pathway conveys motor timing signals that trigger behavioral context-dependent movement initiation.

Direct projections to the brainstem provide an alternate pathway to initiate movement. We found that recruitment of the DN/IPN tha- lamocortical pathway is necessary for learned forelimb move- ment initiation given that photoactivation of DN/IPN axon termi- nals in MThDN/IPN mimics cue-triggered CFA population dynamics and behavior, while silencing each node along the pathway blocked initiation.

In rodents, rapid go cue-evoked changes in activity have been observed in a delayed directional licking task for mice (Catanese and Jaeger, 2021; Gao et al., 2018; Guo et al., 2014; Li et al., 2015),

However, direct photoactivation of MThDN/IPN did not reproducibly evoke short-latency tongue or orofacial movements, suggesting paral- lel, non-overlapping thalamocortical pathways for movement initiation. Directional licking requires channeling of information through VM, VAL, mediodorsal, and intralaminar nuclei for both movement planning and execution (Catanese and Jaeger, 2021; Inagaki et al., 2020), while forelimb movements require ac- tivity in VAL, AM, and VPM nuclei. Together, this suggests that parallel processing of motor timing signals through different ventral motor thalamic nuclei could provide an anatomical sub- strate for initiating complex, multifaceted motor behaviors.

In the absence of thalamic input to MThDN/IPN (i.e., miss trials or MTh inactivation), layer 5B Vm and firing rate changes were signifi- cantly reduced, with residual Vm changes being insufficient to trigger movement, suggesting input convergence is a prerequi- site for learned movement initiation. The origin of the additional input(s) remains unknown, but likely candidates are cortico- cortical interactions between frontal motor areas and CFA (Hooks et al., 2013; Reep et al., 1990), thought to accumulate task-rele- vant information required for motor planning and execution (Gao et al., 2018; Li et al., 2015), or basal ganglia thalamocortical inter- actions involved in selecting, timing, and invigorating different ac- tions (Dudman and Krakauer, 2016; Inase et al., 1996; Klaus et al., 2019; Thura and Cisek, 2017; Williams and Herberg, 1987).

SauerbreiGuoCohenEtAl19

Sauerbrei, B. A., Guo, J.-Z., Cohen, J. D., Mischiati, M., Guo, W., Kabra, M., Verma, N., Mensh, B., Branson, K., & Hantman, A. W. (2019). Cortical pattern generation during dexterous movement is input-driven. Nature, 1–6. https://doi.org/10.1038/s41586-019-1869-9

After we optogenetically set cortex to aberrant initial states, it rapidly produced the appropri- ate pattern for reaching without returning to the initial state observed on control trials, provided the inputs were intact. This suggests that a specific initial state is not required for cortical pattern generation.

Contrary to the predictions of the autonomous model (Fig. 1b, left), silencing or stimulating thalamus severely disrupted cortical pattern generation and arm movement. Thus, temporally patterned inputs are critical for producing the cortical output pattern during movement execution (Fig. 1b, right).

We have shown that the cortical pattern that controls reaching can be produced only if temporally patterned inputs are maintained through- out the entire movement. Thus, while motor cortex is a bottleneck for descending motor commands, its activity patterns during movement are crucially moulded by its embedding within a vast and distributed network, which incorporates muscles, the sensory periphery, subcorti- cal regions and other cortical areas.

[rcoreilly]

PPN glutamatergic neurons and SNr outputs

Mena-SegoviaBolam17

Mena-Segovia, J., & Bolam, J. P. (2017). Rethinking the Pedunculopontine Nucleus: From Cellular Organization to Function. Neuron, 94(1), 7–18. https://doi.org/10.1016/j.neuron.2017.02.027

Major review with lots of refs etc.

The LDT, like the PPN, is composed of GABAergic and glutamatergic neurons, as well as the cholinergic neurons, and similarly innervates a diverse array of structures in the midbrain and forebrain, although LDT targets are predominantly limbic, whereas PPN targets also include motor structures.

• TODO: would be interesting to investigate glutamatergic LDT -- if PPN is any guide, they could be important!

Anatomical studies have identified synaptic connections of LDT neurons in the midline thalamus (Cornwall et al., 1990), the nucleus accumbens (Dautan et al., 2014), the VTA (Oakman et al., 1995), and the hypothalamus (Cornwall et al., 1990). In addition, cholinergic axons originating in the LDT have been identified in the ventral pallidum, the external globus pallidus, amygdala, inferior colliculus, and dorsal raphe (Dautan et al., 2016a). Thus, the targeting of ventral basal ganglia and midline thalamic structures by cholinergic neurons of the LDT mirrors the corresponding set of structures in the dorsal basal ganglia and intralaminar thalamus innervated by the PPN (Figure 2).

In conclusion, we propose that the main role of cholinergic neurons of the brainstem is to facilitate the updating mecha- nisms in the basal ganglia during changing environmental contin- gencies. More experiments are needed to hone and test this hypothesis, but the notion of cholinergic modulation of behav- ioral flexibility has been associated with other cholinergic sys- tems, such as the cholinergic interneurons in the striatum and the basal forebrain projections to the prefrontal cortex.

RoseberryLeeLaliveEtAl16

Roseberry, T. K., Lee, A. M., Lalive, A. L., Wilbrecht, L., Bonci, A., & Kreitzer, A. C. (2016). Cell-type-specific control of brainstem locomotor circuits by basal ganglia. Cell, 164(3), 526–537. https://doi.org/10.1016/j.cell.2015.12.037

The BG locomotor command is thought to be relayed to spinal cord central pattern generators through the mesencephalic loco- motor region (MLR), a brainstem area first described in 1966 by Shik et al. (1966b). The MLR is defined functionally as a mesen- cephalic region in which increasing intensities of electrical stim- ulation induce a transition from a stationary state to walking and then running with short latencies (Shik et al., 1966a, 1966b). In mammals, the MLR overlaps with the cuneiform nucleus (Cun), mesencephalic reticular nucleus (MRN), and pedunculopontine tegmental nucleus (PPTg) (Garcia-Rill et al., 1986; Ryczko and Dubuc, 2013). The MLR is comprised of three neurochemically distinct cell types: glutamatergic, GABAergic, and cholinergic (Martinez-Gonzalez et al., 2011). Although the major cell popula- tions of the MLR give rise to ascending projections into the fore- brain that may be relevant for reward, arousal, and cortical state (Ehrich et al., 2014; Grace, 2010; Lee et al., 2014; Thompson and Felsen, 2013), the control of locomotion appears to be driven through descending outputs because locomotion is intact in decerebrate animals (Bedford et al., 1992; Whelan, 1996).

These results demonstrate that the BG can modulate activity within the MLR, although the responses within the MLR are heterogeneous.

10-ms light pulses delivered at 20 Hz to the glutamatergic population for 5 s elicited robust locomotion (Figures 2F and 2I) at significantly shorter latencies than electrical stimulation (mean movement onset, 211 ± 22 ms; p < 0.01; Wilcoxon rank-sum test). The speed reached at the end of stimulation was graded by stimulation fre- quency (Borgius et al., 2010; Lee et al., 2014), a canonical feature of the MLR (Figure 2K).

In contrast, stimulation of the GABAergic population during running caused deceleration (mean decelera- tion onset, 837 ± 99 ms; Figures 2H and S2F) but no change when the mouse was stationary at the beginning of stimulation (Figures 2E and 2J).

Choline acetyltransferase (ChAT) neuron stimulation resulted in a significant increase in speed during trials when the mouse was running but not when the mouse was stopped (Figures 2D, 2G, and 2J and S2F). Therefore, ChAT neu- rons appear to modulate locomotion but are not sufficient to drive locomotion at short latencies. This modulation did not appear to result from co-release of glutamate or GABA because no EPSCs or IPSCs were observed in the MLR during light stim- ulation in slice (Figure S1E). eYFP controls showed no significant changes in locomotion (Figures 2J and S2F).

Together, these re- sults indicate that increased activity within the glutamatergic population alone is sufficient to drive locomotion.

• Note: this is all relatively slow latency modulations

In contrast, unidentified neurons from a separate cohort of mice displayed significantly lower areas under the curve (AUCs), indicating functional heterogene- ity among MLR neuronal subpopulations (Figure 3F). To further dissect this result, we tested the locomotor-predictive MLR glutamatergic neurons for correlations with speed using a linear regression model. This analysis yielded two distinct populations: one that predicts locomotor state alone and one that predicts state and correlates with speed (Figure 3G).

However, in individual trials, spiking increase onset was highly variable, and an absolute threshold at which spiking would correlate with running onset was not observed (Figure 3H, inset). Therefore, we tested the prediction that spiking increases during the stationary state would result in an increased probability of running onset. Indeed, an increased firing rate during the stationary state was correlated with an increase in the probability of a start occurring within the next second (Figure 3I; Pearson’s correlation coefficient; p < 0.01). This suggests that, at the individual level, glutamatergic neurons do not predict the timing of locomotion onset. Rather, these neurons contribute to an increased probability of locomo- tion that is read out from the population.

Because activity in MLR glutamatergic neurons is required for running (Figure 4), deceleration and stopping because of optogenetic activation of GABAergic neurons is likely in part due to local inhibition of MLR glutamatergic neurons. However, in vivo recordings from optogenetically identified MLR GABA neurons during spontaneous locomo- tion revealed heterogeneous responses, indicating additional complexity in the composition and function of this subpopula- tion (Figure S4D).

We next tested the connection strength between the BG and the two MLR populations displaying the most robust effects on loco- motion. In theory, BG-driven locomotion could be initiated by disinhibition of MLR glutamatergic neurons (Grillner et al., 2008; Hikosaka et al., 2000) or inhibition of GABAergic neurons.

Retrograde trans-synaptic labeling from MLR glutamatergic neurons revealed dense projections from several BG nuclei, whereas few, if any, projections targeted MLR GABAergic neurons (Figures 5C–5F).

• Note: CeA, SC potentially allow direct CS -> US -> motor responding connections. But SNr holds everything under tonic inhibition, and disinhibition there allows action initiation, in the classical story.

• interestingly, mostly dorsal / medial SNr -- crossreferencing with LeeWangSabatini20 shows that this is the "VLS" circuit, which is specifically identified as the "lick zone" and also the area for locomotor turning: Activation of VLS modulates the activ- ity of cortical and collicular regions involved in licking, consistent with the existence of a functional lick-controlling loop involving BG. This is critical for understanding what is going on with the InagakiChenRidderEtAl22 paper below! Somehow licking and locomotion are colocated??

• Also, Kusumoto-YoshidaLiuChenEtAl15 shows that insular cortex (IC), which is the cortical area interconnected with ventral lateral striatum (VLS), is critical for linking sensory cue to food approach behavior.

A brain-wide survey of long-range projections to these cell types revealed another strong projection to MLR glutamatergic neurons from the central amygdala and oval bed nucleus of the stria terminalis (ovBNST), GABAergic nuclei that could play a role in fear- and anxiety- associated behaviors such as freezing (Figure 5F; LeDoux, 2000). Major MLR GABAergic targeting regions included the superior colliculus, dorsal raphe, laterodorsal tegmentum, and ovBNST. Together, these results suggest a specific role for MLR glutamatergic neurons in the control of locomotion by up- stream targets and, notably, the BG.

• Overall, this is classic BG inhibition of SNr resulting in disinhibition of downstream PPN / MLR neurons, resulting in "Go" motor activity.

In addi- tion, our tracing highlighted a number of non-canonical path- ways to the MLR from the BG (entopeduncular nucleus [EP], external globus pallidus [GPe], subthalamic nucleus [STN]) to MLR glutamatergic neurons. The STN projection is of interest because it is predicted to arrive in the MLR prior to the classical indirect pathway signal through the SNr, and it should drive ac- tivity in the opposite direction. Indeed, a small minority of cells displayed a small uptick in firing rate prior to inhibition (Figure 6F), consistent with the idea that this pathway modulates MLR activ- ity, perhaps as a brief arousal signal prior to suppression of locomotion.

Our data suggest that the balance of activity between the direct and indirect pathways is represented in the firing rate of glutamatergic MLR neurons, which is predictive of the initiation of running and sufficient to drive graded locomotion. Further experiments can help clarify the validity of this model for the MLR and other BG output structures.

AssousDautanTepperEtAl19

Assous, M., Dautan, D., Tepper, J. M., & Mena-Segovia, J. (2019). Pedunculopontine Glutamatergic Neurons Provide a Novel Source of Feedforward Inhibition in the Striatum by Selectively Targeting Interneurons. Journal of Neuroscience, 39(24), 4727–4737. https://doi.org/10.1523/JNEUROSCI.2913-18.2019

Glutamater- gic neurons constitute the largest neuronal population in the PPN (Wang and Morales, 2009) and innervate different compo- nents of the basal ganglia, including midbrain dopaminergic neu- rons (Ros et al., 2010; Galtieri et al., 2017) and the striatum (Klug et al., 2018). Furthermore, PPN glutamatergic neurons have re- cently been implicated in the regulation of motor activity as a downstream target of the basal ganglia (Roseberry et al., 2016; Caggiano et al., 2018), suggesting that their bidirectional connec- tivity with basal ganglia nuclei may underlie their role in motor behavior.

Those results demonstrate that PPN glutamatergic neu- rons do not monosynaptically innervate SPNs.

In contrast to the results described above, under these conditions, we ob- served that, in 48.38% of the recorded SPNs, optogenetic stimu- lation of PPN glutamatergic inputs elicited IPSCs (n 15 of 31; 15.16 � 2.661 pA; Fig. 3D–M). These IPSCs were mediated by GABAA receptors as they were blocked by bicuculline (10 �M, n 6; two-tailed paired t test, p 0.0172; time constant: 8.515 � 1.142 ms, n 14; Fig. 3D–M) and exhibited unusually long latencies (8.459 � 0.698 ms; Fig. 3 H, I; n 15) implying media- tion by polysynaptic pathways. The IPSCs were also blocked by bath application of AMPA/NMDA glutamate receptor antago- nists (Fig. 3 L, M; n 6), which confirmed the polysynaptic na- ture of the response. These results suggest that activation of PPN glutamatergic axons in the striatum induced feedforward inhibi- tion of SPNs.

In this study, we also reveal that striatal-projecting midbrain PPN glutamatergic neurons are embedded in a region that has been functionally defined as the mesencephalic locomotor region (for review, see Mena-Segovia and Bolam, 2017). Together with the cuneiform nucleus, activation of mesencephalic locomotor region excitatory pathways has been suggested to modulate loco- motor speed and gait (Roseberry et al., 2016; Caggiano et al., 2018). Recent studies (Josset et al., 2018), including our own unpublished data, however, have reported that the optogenetic activation of PPN glutamatergic neurons, rather than inducing an increase in motor activity, induces a reduction in movement, suggesting that the PPN recruits spinal circuits to halt locomo- tion (Takakusaki et al., 2016).

• umm, Caggiano et al (2018) are saying that it induces exploratory locomotion.. not clear what's up here.

Furthermore, our findings reveal the PPN as the only known excitatory input capa- ble of driving striatal inhibition without concurrent excitation of SPNs.

CaggianoLeirasGoni-ErroEtAl18

Caggiano, V., Leiras, R., Goñi-Erro, H., Masini, D., Bellardita, C., Bouvier, J., Caldeira, V., Fisone, G., & Kiehn, O. (2018). Midbrain circuits that set locomotor speed and gait selection. Nature, 553(7689), Article 7689. https://doi.org/10.1038/nature25448

The most important neuronal struc- ture that has been implicated in these functions is the mesencephalic locomotor region (MLR)7–9, which is located in the midbrain.

The MLR was first defined functionally in cats as a region locali- zed in or around the CnF, in which continuous electrical stimulation evoked persistent locomotion10. Analogues of the MLR have been observed in many vertebrates—including fish, rodents, primates and humans8,9,11,12—but with conflicting results as to their anatomical location. In addition to the CnF, the more ventrally located PPN has also been implicated in locomotor control. Besides being anatomi- cally separated, each of these regions contain neurons with diverse transmitter phenotypes with excitatory long-range projection neurons —glutamatergic in the CnF and both glutamatergic and cholinergic in the PPN—intermingled with local inhibitory interneurons11,12.

Electrical stimulation or lesion studies are therefore unable to distin- guish the contribution from the various intermingled neuronal popu- lations present in these areas12,13. Recently, optogenetic manipulations have shown that stimulation of GlutNs in and around the PPN induces locomotion in mice14.

Our results reveal that the MLR is defined by glutamatergic subpopulations of neurons in both the PPN and the CnF that may act in conjunction to control slower, alternating-gait locomotion. Furthermore, glutamatergic neurons in the PPN promote locomotion for the purpose of explorative behaviour, whereas those in the CnF promote escape locomotion. Our study identifies circuits that have key roles in the appropriate command pathways for selecting locomotor outputs contingent on behavioural contexts.

The main inputs to Vglut2+PPN neurons originate in midbrain struc- tures (Fig. 6c) and sensory-motor and raphe nuclei in the brainstem (Fig. 6d). Furthermore, Vglut2+PPN neurons also receive direct input from the output nuclei in the basal ganglia (Fig. 6e, f). Sparse inputs were found from sensory-motor and frontal cortices or the hypothala- mus (Fig. 6c). Therefore, Vglut2+PPN neurons integrate sensory-motor information from many brain structures. Conversely, Vglut2+CnF neurons receive little input from basal ganglia output nuclei (Fig. 6e, f) or from cortices, but stronger projections from midbrain structures (for example the periaqueductal grey or the inferior colliculus, Fig. 6c, d) that have been assigned a role in escape responses25,26.

Vglut2+PPN neurons have broad—predominantly ipsilateral— projections, including to motor-related nuclei in the pons as well as to modulatory nuclei (Extended Data Fig. 9b, c1–4). Most of these brainstem nuclei project to the spinal cord in mice27.

JossetRousselLemieuxEtAl18

Josset, N., Roussel, M., Lemieux, M., Lafrance-Zoubga, D., Rastqar, A., & Bretzner, F. (2018). Distinct Contributions of Mesencephalic Locomotor Region Nuclei to Locomotor Control in the Freely Behaving Mouse. Current Biology, 28(6), 884-901.e3. https://doi.org/10.1016/j.cub.2018.02.007

Looks like similar overall conclusions to above.

InagakiChenRidderEtAl22

Inagaki, H. K., Chen, S., Ridder, M. C., Sah, P., Li, N., Yang, Z., Hasanbegovic, H., Gao, Z., Gerfen, C. R., & Svoboda, K. (2022). A midbrain-thalamus-cortex circuit reorganizes cortical dynamics to initiate movement. Cell, 185(6), 1065-1081.e23. https://doi.org/10.1016/j.cell.2022.02.006

This paper is amazing and yet rather confusing because it loses the trail of causality at the PPN/ MRN system, and also is so narrowly focused on ALM. Their story is essentially that the PPN is a kind of "homunculus" that detects the Go cue and decides to act. PPN/MRN is also mostly associated with locomotion, and yet they are looking at licking? The mice are head restrained. Also, the role of the VP thalamus is relevant in that the directional cue is whisker-based.

Based on above literature, it is likely that PPN/MRN are being controlled by strong inhibitory SNr inputs, and that SNr disinhibition is necessary to enable PPN activity. The relevant area of SNr is apparently the VLS according to above (LeeWangSabatini20 and Roseberry et al), which makes perfect sense in terms of the behavioral response in this paper. Thus, PPN activation is a correlate of BG Go activity coming from somewhere else outside of ALM. Projections to VLS come from the insular cortex (Berendse et al., 1992; FudgeBreitbartDanishEtAl05) -- NOT the ALM. There are also SC, CeM inputs, but apparently CeM is inhibitory according to another Roseberry paper?

According to AssousDautanTepperEtAl19 (via GrillnerRobertsonKotaleski20), PPN projects to FSI neurons in thalamus (very broadly across thalamic areas), so Go initiation should involve inhibiting the PPN neural activity. This inhibitory dynamic suggests that when moving, the MSNs are broadly suppressed?

BUT, the manipulations in this paper involve exciting PPN to trigger Go, and inhibiting to prevent, so this does not go through. None of the cell types in this paper were systematically identified I think, so it is unclear if perhaps the causal event really is PPN -> FSI and that this inhibition of striatum somehow results in VM gating?

Regardless, the broader point remains that ALMthal and ALM are downstream of actual motor initiation in this kind of classic Go cue task -- if SNr is truly the driver, it suggests that another BG area is making the Go call (based on strong auditory sensory inputs presumably in this case) and that is driving the rest of thalamus which just reports back up to cortex. Cortex has loaded up the details of the planned action, so it gets to influence the action as it happens (it isn't superfluous).

In this case, it looks like anterior insular cortex (AIC) projects to VLS (ventral lateral striatum) which then projects to the part of the SNr that projects to PPN/MRN, which is then somehow disinhibiting "ALMthal" (VM etc) to enable motor actions in ALM to shape the output. This is a version of the classic "ventral leads, dorsal follows" dynamic of goal-driven functionality.

For example, during motor planning, preparatory activity in MCx occupies an activity mode that discriminates future movement types. This activity mode follows attractor dynamics and funnels preparatory activity to an initial condition (a fixed point) appro- priate to trigger accurate and rapid movements (Churchland et al., 2010; Shenoy et al., 2013; Inagaki et al., 2019).

After the Go cue, the dynamics in MCx shows large changes. The motor planning mode collapses (Funahashi et al., 1989; Shadlen and Newsome, 2001; Kaufman et al., 2016, 2014), and a new activity mode with multi-phasic dynamics emerges (Churchland et al., 2012).

This movement-type-specific mode is preferentially repre- sented in the descending MCx neurons that project to premotor neurons in the brainstem and spinal cord (Li et al., 2015; Economo et al., 2018; Duan et al., 2021)... These modes occupy near-orthogonal subspaces, which may explain in part why movements are not triggered dur- ing planning (Kaufman et al., 2014; Elsayed et al., 2016).

• Key point: Action initiation = matrix-mediated VM -> PT gating!

ALM forms reciprocal connections with parts of the thalamus (referred to as ALM-pro- jecting thalamus, or thalALM) to maintain the motor plan (Guo et al., 2017). ThalALM receives input from the basal ganglia, cer- ebellum, and the midbrain, which directly or indirectly receive input from ALM (Guo et al., 2017,2018). Thus, thalALM is a hub linking subcortical structures and ALM, forming multi-regional loops essential for orofacial movements.

Here, we mapped the mechanisms underlying cue-triggered switching of activity modes and the resulting movement initi- ation, in the context of a delayed directional licking task (Guo et al., 2014b). By combining anatomy and large-scale electrophysiological recordings, we established that Go-cue- related information flows from the midbrain to ALM via thalALM, where it initiates Dgo signals and motor-command- like dynamics in ALM, followed by appropriate movements. Altogether, we have identified a multi-regional pathway medi- ating cue-triggered mode switching for the release of planned movements.

The simple notion that preparatory activity is a subthreshold motor command therefore does not explain ALM activity around the time of movement initiation.

In contrast, population selectivity has a low correlation between time points before and after the Go cue (Figure 1D, magenta box), implying that different combinations of neurons show selectivity. Similarly, the population activity of ALM neurons within each trial type (rlick-right, t, or rlick-left, t) shows low correlation before versus after the Go cue (Figure S1B).

Activity projected onto CDdelay and CDresponse is correlated at the level of single trials (Figure S1L). This implies that information carried along CDdelay is transferred to CDresponse following the Go cue (i.e., trials with strong activity along CDdelay have strong activity along CDresponse, and vice versa). This finding is consistent with the observation that fine-scale movement parameters and reaction times are coded in preparatory activity (unpublished observa- tions) (Li et al., 2016; Even-Chen et al., 2019) and implies that ALM preparatory activity (activity along CDdelay) contributes to control of future movements (activity along CDresponse).

Activity along Dgo explains a large proportion of ALM activity after the Go cue (Figure S1I), similar to the ‘‘condition-invariant signal’’ described in a primate reaching task (Kaufman et al., 2016). Activity along Dgo is non-selective (Figure 1E) and cannot decode lick direction (Figure S1J) because activity changes around the Go cue are largely similar across trial types (Figure S1C). The trial-type differences that do exist contribute to CDresponse. These three directions in activity space, together with a fourth direction that captures non-selective ramp- ing activity during the delay epoch (Li et al., 2016; Inagaki et al., 2018), account for nearly all (87.4%, 84.8%–89.6%; mean, 2.5%–97.5% confidence interval) of population activity around the Go cue (±200 ms from the Go cue; Figure S1I, cyan line).

We silenced ALM neurons projecting to the medulla (pyramidal tract neurons in lower layer 5b, ‘‘PTlower’’) which contribute disproportionately to the CDresponse (Economo et al., 2018). Because the medulla contains the motor centers for orofacial movement (Travers et al., 1997; Stanek et al., 2014), de- scending signals from ALM to the medulla may be necessary for movement initiation.

Bilateral optogenetic silencing of PTlower cells in ALM (centered at anterior-posterior [AP] 2.5 mm L 1.5 mm from Bregma; 1 s of photostimulation starting at the onset of Go cue) resulted in a loss of cue-triggered licking (Figures 2B, 2C, S2B, and S2C). Similar bilateral silencing in posterior cortical regions (centered around AP 0 mm L 1.5 mm from Bregma) had a weaker behavioral effect (Figure S2C). Following the end of the photostimulus, mice licked in the correct direction (Figures 2B and S2D), implying that activity of ALM PTlower cells is required to initiate movements, but not to maintain motor plans and not for a memory of the Go cue.

These experiments imply that Dgo develops independent of CDresponse.

Additional support for distinct roles of Dgo and CDresponse comes from analysis of trials in which mice failed to lick after the Go cue (no response trials; mostly near the end of a session when they are satiated). Activity along CDdelay is attenuated dur- ing the delay epoch in these trials likely modulated by the moti- vational state of the animal (Allen et al., 2019) (Figures S2I and S2J). Although activity along Dgo increases, activity along CDresponse does not develop after the Go cue (Figures S2I and S2J). Thus, even when activity along Dgo appears after the Go cue, without properly developed motor planning (CDdelay), activ- ity along CDresponse does not emerge.

In contrast, activity along Dgo precedes movement but it is not instructive on movement type or sufficient to trigger movement by itself (without a change in activity along CDresponse). Thus, ac- tivity along Dgo is not a motor command. Instead, Dgo may trigger the activity along CDresponse following the Go cue. Testing this hypothesis requires manipulations of activity along Dgo by activating or inhibiting neurons that carry this signal. This requires mapping the pathways that transmit Go-cue- related signals to ALM.

• Key point: consistent with Matrix / modulatory nature of VM, Dgo is "pure gate" and not action specific -- BG, by implication, does not communicate specific actions, but just when to initiate.

ALM forms strong reciprocal connections with thalALM, including parts of the ventral medial (VM), ventral anterolateral VAL), mediodorsal (MD), paracentral (PCN), central lateral (CL), central medial (CM), and parafascicular (PF) nuclei of the thal- amus (Guo et al., 2017). The PCN, CL, CM, and PF comprise the so-called intralaminar (IL) nuclei of the thalamus.

We performed extracellular recordings in thalALM and compared responses to the Go cue (i.e., changes in spike rate after the Go cue) with those in ALM (Figures 3A, 3B, S3, and S5). Neurons in ALM and thalALM responded with increases or decreases in spike rate (go-up and go-down cells, respec- tively). The latency was shorter in thalamus (16.5 ± 1.5 ms; mean ± SEM; time when 1% of cells show increase in spike rate) compared with ALM (25.0 ± 0.8 ms; mean ± SEM) (p < 0.001; bootstrap; Figures 3A and 3B). The latency difference between thalALM and ALM is consistent with the action potential propagation speed in thalamocortical axons (Guo et al., 2017). Neurons with short latencies (<20 ms) were widespread in thalALM, and appeared to be partially spatially segregated from delay-selective neurons (Figure 3C; 5.4% of thalALM cells with delay selectivity had short latency) (Gaidica et al., 2018).

• Critical point that is almost entirely brushed over in this paper: The majority of short latency thalamic responses are outside of Thal_alm. In the AP -1.6mm slice, that looks like LD and VP nuclei (LD = anterior nucleus, involved in mammilary body / memory circuits; VP = somatosensory). There is very poor sampling of areas where activity may be strongest. So, the key question remains: how exactly does the PPN drive the Go signal? it is probably indirect relative to most of THALalm?

Thus, ALM is not necessary for the Go cue response in thalALM. Together with the latency anal- ysis, these results indicate that the Go cue activity first arrives in thalALM, then drives ALM (Dacre et al., 2021; Takahashi et al., 2021).

To map the projections from these subcortical areas within thalALM and beyond, we injected AAVs expressing fluorescent proteins in each area (Figures 4A and S4D; Table S2). PPN/ MRN neurons have widespread projections to thalALM, whereas other structures have relatively localized projections (Figure 4B). We focused on PPN/MRN because their output overlaps with the short latency Go cue responses in thalALM (Figure S4F).

Thalamus-projecting PPN/MRN (PPN/MRNTh) neurons are distributed across a region referred to as the ‘‘mesencephalic locomotor region’’ (MLR) (Shik et al., 1966). Stimulation of the MLR produces locomotion, mediated by glutamatergic neurons that descend into the medulla (Shik et al., 1966; Roseberry et al., 2016; Josset et al., 2018; Caggiano et al., 2018). PPN/MRN neu- rons in mice project to both thalamus and the medulla (Fig- ure 4C).

These experiments re- vealed that PPN/MRNTh neurons lack a projection to the me- dulla, and thus constitute a distinct population, intermingled with neurons that project to the medulla (Figure 4D). PPN con- tains a high density of cholinergic (chat+) cells (Sofroniew et al., 1985; Mena-Segovia and Bolam, 2017; Huerta-Ocampo et al., 2020). However, the majority (75%) of PPN/MRNTh neu- rons are glutamatergic (vglut2+), not cholinergic or GABAergic (gad1+), as confirmed by fluorescent in situ hybridization and immunohistochemical staining (Figures S4G–S4J; Table S4).

We compared the latencies and spike rates after the Go cue across brain areas (Figures 5D and S5). All thalALM-projecting subcortical areas contain cells with Go cue latencies shorter than thalALM. ... Among the thalALM-projecting areas, PPN/MRN is one of the first where Dgo emerges after the Go cue (Figure 5E; p = 0.07, the probability that PPN/MRN has the shortest latency, hierarchical bootstrap). Unlike activity along Dgo, the selective component of the Go cue response, i.e., activity along CDresponse in each brain area, emerged almost simultaneously across brain areas, and later than Dgo (Figure 5F). Thus, following the Go cue, non-selective activity rapidly spreads across brain areas, followed by emergence of selective activity.

The fast Go cue responses in PPN/MRN are not simple audi- tory responses. Rather, they constitute a learned response that is specific to the sound used as the Go cue.

Optogenetic perturbation during the Go cue presentation blocked cue-trig- gered licks (Figure 7B; laser was turned on 0.6 s before the Go cue and lasted for 1.2 s; 473 nm 40 Hz sinusoidal). Perturbation of glutamatergic PPN/MRNTh neurons (using AAVretro-CamKII- Cre; Figure S4H) resulted in a similar behavioral effect (Figure 7B; in contrast, perturbation of cholinergic neurons did not affect movement initiation).

• Key: glut PPN, not cholinergic

Our findings are consistent with previous experiments in rats, cats, and monkeys, in which lesioning or silencing of PPN/MRN blocks cue-triggered movements (Wilson, 1973; Conde ́ et al., 1998; Florio et al., 1999). In humans, PPN/MRN and downstream thalamic areas show activity in cued movement tasks (Kinomura et al., 1996).

PPN/MRN receives input from NLL (Reese et al., 1995), which itself receives direct auditory input from the cochlear nucleus (Davis et al., 1982). Therefore, the latency to the Go cue in NLL is shorter than in PPN/MRN (Figure 5). The auditory response in PPN/MRN is specific to the sound associated as the Go cue (Figures 5G and S5G–S5J). PPN/MRN is likely the site of this as- sociation. PPN/MRN also responds to other sensory stimuli (Pan et al., 2005; Okada and Kobayashi, 2009) and may serve to asso- ciate sensory stimuli as a ‘‘Go’’ signal.

For example, VM contains neu- rons that have broad projection patterns to layer 1 (‘‘matrix’’). VAL instead projects in a more focal manner to middle layers (‘‘core’’) (Jones, 1998; Kuramoto et al., 2015). These different thalamocortical projections activate cortical microcircuits in specific ways (Anastasiades et al., 2021). The spatial distribution of thalALM neurons with short latencies to the Go cue appears to differ from those showing delay selectivity (Figure 3C).

We note that the PPN/MRN input to thalALM produces short latency and reliable synaptic currents with paired-pulse depression (Figures S4P and S4Q). Thus, the PPN/MRNTh thalALM projection has the hallmarks of a classic driver input (Guillery and Sherman, 2002).

Here, we emphasized the PPN/MRNTh / thalALM / ALM cir- cuit. However, PPN/MRN neurons likely exert their role in move- ment initiation via additional brain regions. For example, PPN/ MRNTh neurons also project to the SNr and subthalamic nucleus (STN) (Figure 4A) (Martinez-Gonzalez et al., 2011; Vitale et al., 2019; Ferreira-Pinto et al., 2021). STN in turn projects to SNr, and SNr is known to control premotor circuits in the SC and me- dulla (DeLong, 1990; Hikosaka et al., 2000; Klaus et al., 2019). In addition, PPN/MRNTh neurons may locally excite or inhibit PPN/ MRN neurons that descend to motor centers. Indeed, PPN/MRN is ideally positioned to coherently control brain-wide circuits for movement initiation.

Yet, in a cued reaching task, DCN activity is causal for the movement initiation and activity in MCx (Dacre et al., 2021). It remains to be seen whether DCN can serve a similar role as PPN/MRN in terms of switching modes of cortical activity. PPN/MRN and DCN project to partially overlapping thalamic nuclei. Further investigation is required to test whether PPN/ MRN and DCN are redundant, serve as parts of a chain or loop (Hazrati and Parent, 1992; Bostan and Strick, 2018; Judd et al., 2021) or as parallel Go cue pathways for different sectors of MCx, or are recruited differently depending on task requirements.

[rcoreilly]

Kuramoto et al on VM vs. VL

KuramotoOhnoFurutaEtAl15

• Confirms very wide axonal arbor of VM neurons relative to more focal projections of VL neurons (see figs above).

Thus, the present results on VM neurons, which are in good agreement of those previous findings, indicate that VM neurons are classified as matrix-type neurons (Jones 1998, 2001). Recently, a further distinction has been made between matrix-type neurons with widely ramified axons targeting mul- tiple, distant cortical areas (multiareal subtype) and those with an axon that arborizes within a single or few adjacent areas (focal subtype) (Clascá et al. 2012). VM neurons are obviously classified as the multiareal subtype, because all the examined VM neurons projected to multiple distant cortical areas.

Although, from the present sampling of VM neurons, we did not detect a clear spatial grouping of somal locations for the 2 populations within the VM, nonmotor-preferring VM neurons had more abundant dendritic arborization than motor- preferring neurons (Fig. 8A,C).

Thus, the motor-associated areas receive 3 kinds of information: 1) cerebellar information through the EZ to their middle cortical layers, 2) basal ganglia information through the IZ mainly to L1, and 3) basal ganglia information through the VM highly preferentially to L1. Taking the shape of pyramidal neurons into consideration, the basal dendrites of L2/3 and L5 pyramidal neurons receive cerebellar information through the EZ and L4 spiny neurons, and the apical dendritic tufts of those pyramidal neurons admit infor- mation from the basal ganglia through the IZ and VM. As the axonal arborizations of single EZ, IZ, and VM neurons are widely distributed in the motor-associated areas, these 3 kinds of inputs are considered to converge on single pyramidal neurons in the motor-associated areas. Thus, the cerebellar and basal ganglia information would be integrated in those pyramidal neurons and used for an appropriate motor control.

These findings hence raise a question of why the basal ganglia information is transferred by the 2 routes through the IZ and VM. One possible answer is that the 2 routes convey different lines of information, which are produced in the basal ganglia. Actually, the IZ or rostroventral portion of the VA-VL receives afferents mainly from the ventral edge of the SNr, whereas the VM accepts afferents chiefly from the central and dorsal portion of the SNr in the rat brain (Deniau and Chevalier 1992; Nishimura et al. 1997; Sakai et al. 1998). Thus, the basal ganglia information transferred by motor-preferring VM neurons might be different from that conveyed by IZ neurons. However, to our knowledge, no functional difference has been reported between the 2 portions of the rat SNr, except for the proposal that the ventral SNr–IZ pathway is concerned with oculomotor function (Sakai et al. 1998).

The thalamocortical axon fibers of these VM neurons were mostly distributed in L1 of those orbital and cingulate areas. However, the thalamocortical axon terminals, which are specifi- cally characterized by VGluT2 immunoreactivity in the cerebral cortex (Fujiyama et al. 2001; Kaneko and Fujiyama 2002), are densely distributed in middle cortical layers as well as in the superficial part of L1. Thus, these areas may receive afferents at their middle layers from thalamic nuclei other than the VM (Fig. 9). The mediodorsal thalamic nucleus (MD) is generally known to project to the “prefrontal areas,” which consists of 1) infralimbic and prelimbic areas, 2) orbital areas, 3) anterior cin- gulate areas, and 4) agranular insular areas in rat brain (Krettek and Price 1977; Divac et al. 1978; Vogt et al. 1981; Groenewegen 1988; Zeng and Stuesse 1991; Ray and Price 1992; Reep et al. 1996). The MD sends axon fibers to the middle layers as well as to L1 of cingulate and orbital areas (Krettek and Price 1977; Vogt et al. 1981; Groenewegen 1988), and thus their thalamo- cortical fibers may collaborate with the afferents from VM neurons that mainly project to orbital and/or cingulate areas. This relation of MD axons to VM axons in the cingulate and orbital areas might be similar to that of EZ axons to IZ/VM axons in the motor-associated areas. Thus, a further single- neuron-labeling study on the axon fibers of MD neurons is necessary to understand the structural organization of VM afferents to these higher-order prefrontal areas.

[rcoreilly]

AokiSmithLiEtAl19

Aoki, S., Smith, J. B., Li, H., Yan, X., Igarashi, M., Coulon, P., Wickens, J. R., Ruigrok, T. J., & Jin, X. (2019). An open cortico-basal ganglia loop allows limbic control over motor output via the nigrothalamic pathway. eLife, 8, e49995. https://doi.org/10.7554/eLife.49995

• Key finding of VS -> DS SNr pathway for "open loop" control pathway.

Though some studies have implicated mechanisms for limbic-motor interactions in the dopaminergic system (Beier et al., 2015; Belin and Everitt, 2008; Haber et al., 2000; Lerner et al., 2015; Watabe-Uchida et al., 2012; Yang et al., 2018), how limbic information ulti- mately reaches motor circuitry remains largely unknown, as do the characteristics of its influence.

A ‘partially-open’ loop architecture in the cortico-basal ganglia circuitry has been suggested from primate studies (Joel and Weiner, 1994; Kelly and Strick, 2004; Miyachi et al., 2006), but this previous evidence is incomplete and the precise anatomical basis underlying connections between functionally distinct loops has not been identified. This lack of clarity in the cortico-basal ganglia connections is due to technical limita- tions, which include the lack of sophisticated viral tools and the complicated geometry of basal gan- glia nuclei.

In addition to the closed loop within each functional domain, we identified a novel one-way interaction across these cortico-basal ganglia loops, in which the limbic loop exerts a unidirectional influence over the motor loop that is mediated by ventral striatum-medial SNr-motor thalamus circuitry. Using slice physiology and optogenetics, we show that VS functionally inhibits medial SNr neurons that project to motor thalamus. We then characterized the influence of limbic and motor striato-nigral outputs onto downstream cortical targets by optogenetically activating striato-nigral terminals from VS or DLS and by recording neuronal activity in mPFC and M1. Consistent with our anatomical findings, the in vivo recording experiments showed, in addition to the within-loop activation in which VS acti- vated mPFC and DLS activated M1, significant activation of M1 when VS terminals in the SNr were stimulated. Conversely, DLS output did not effectively modulate activity in mPFC.

The open-loop struc- ture that we revealed here is consistent with earlier studies in primates that have identified multi-syn- aptic connectivity from ventral putamen to M1 (Kelly and Strick, 2004; Miyachi et al., 2006), with conceptual work on the convergence of basal ganglia outputs to motor circuits (Allen and Tsu- kahara, 1974; Haber, 2003; Joel and Weiner, 1994; Kemp and Powell, 1971), and with behavioral findings that suggest the involvement of VS in modifying motor output (Belin and Everitt, 2008; Floresco, 2015; Sawada et al., 2015).

TS = Tail of striatum (?)

In addition to VS, the other striatal territories, including DMS and TS, were also found to connect to the motor cortex, suggesting a wider theme of all modalities converging onto motor circuits, in addition to their within-domain interactions.

[rcoreilly]

Recent computational models of BG sequence learning

CalderonVergutsFrank21 ACDC model

Calderon, C. B., Verguts, T., & Frank, M. J. (2021). Thunderstruck: The ACDC model of flexible sequences and rhythms in recurrent neural circuits (p. 2021.04.07.438842). https://doi.org/10.1101/2021.04.07.438842

ACDC = Associative Cluster-Dependent Chain

First, we propose a bio- logically-plausible model of cortico-basal ganglia-thalamic loops that decomposes the func- tions of cortex and basal ganglia and learns sequences based on simple local and (biologically motivated) supervised learning rules. Second, this decomposition affords greater flexibility in generating desired action sequences, supporting temporal asynchrony, shifting, rescaling, and compositionality in a single model. Crucially, our model factorizes action sequence features within the circuit, with cortical RNN representing latent states within a sequence, and BG con- trolling both the timing of the transitions from one state to the next and which actions are linked to sequence positions. Factorizing order and timing information by storing them sepa- rately in a premotor cortical RNN, which is dynamically gated by a basal ganglia-thalamus module, affords independent (and flexible) manipulation of sequence order and action timing, and thus increases computational flexibility.

Notes:

• Overall, has an implausibly localist character with discretized clusters in PMC and localist projections from thalamus: VM thalamus in particular has very broad connectivity (e.g., KuramotoOhnoFurutaEtAl15), and at the level of something like finger presses or trunk actions (forward, left, right), it is likely that the motor cortical reps are highly overlapping.

• Transitions happen via Go -> Thal -> NoGo which is inconsistent with simultaneous activity of Go & NoGo at start of actions (CuiJunJinEtAl13, KlausAlvesdaSilvaCosta19), and with idea that Thal -> Striatum is specific to NoGo.

• Despite these issues, the general idea that cortex has a more distributed, continuously-evolving state representation, and BG drives a more discrete transition sequence according to a specific timing, and that these transitions involve (broad) thalamic signals that coordinate between cortex and BG, makes a lot of sense.

LogiacoAbbottEscola21

Logiaco, L., Abbott, L. F., & Escola, S. (2021). Thalamic control of cortical dynamics in a model of flexible motor sequencing. Cell Reports, 35(9), 109090. https://doi.org/10.1016/j.celrep.2021.109090

To generate an arbitrary sequence, the basal ganglia switches between inhibiting different subsets of thalamic units to prepare and execute different motifs. This framework ascribes to the basal ganglia the roles of sequence selection and transition timing and to the thalamocortical loop, in conjunction with motor cortex, the roles of motor preparation and motif execution. Our framework suggests that the cortex-thalamus architecture is well suited to flexibly control prolonged and complex sequential motor behavior.

Second, activity patterns in GPi/SNr have been shown to be sustained during motif execution, with switches pre- dominantly at the transitions between motifs (Jin et al., 2014). In our model, we assume that the synapses from basal ganglia to thalamus are strong enough such that, when active, they completely inhibit their thalamic targets (Edgerton and Jaeger, 2014)—which we assume to be a specific subset of thalamic neurons involved in learning new motifs. This assumption is further supported by the presence of recurrent inhibition within the thalamus (e.g., via the thalamic reticular nucleus; Arcelli et al., 1997), which—though ignored by our model—would only facilitate the ability of the basal ganglia to select among different thalamocortical loops via winner-take-all dynamics (Murray and Escola, 2017). This is compatible with the classic view that motor thalamus is in an inhibited state by default and is disinhibited by targeted removal of inhibitory input from the basal ganglia, which gates movement (Deniau and Chevalier, 1985; Edgerton and Jaeger, 2014; Kim et al., 2017; Aoki et al., 2019; but see Schwab et al., 2020 and STAR Methods, section 4.2.5). Importantly, this introduces a critical nonlinearity in our model by allowing thalamic units to be switched ‘‘on’’ and ‘‘off’’ across motifs. Spe- cifically, inhibited thalamic units are silent while disinhibited ones are free to respond to their cortical inputs—in which case they participate to the motif-specific linear dynamics.

• Note: this seems like an over-strong set of assumptions about VM thalamic activity..

First, we restrict learning to the synapses within the thalamocortical loops and assume that the intracortical and output synapses are fixed. Second, we restrict the subsets of thalamic neurons that are active during different motifs to be non-overlapping. These constraints guar- antee that motifs do not interfere and offer a procedure for add- ing new motifs: identify a new subset of thalamic units and set their synapses to and from cortex such that, when they are released from basal ganglia inhibition, the cortex-thalamus sys- tem drives the output to generate some new target movement (Figure 1B).

• Note: this is similar to the ACDC model and highly implausible given the nature of VM. Both of these models are implausibly using the thalamus as a kind of player piano roll to drive discrete sequencing.

Recently, motor cortex recordings have shown that motor preparation occurs before starting each sequence element during rapidly executed sequences (Zimnik and Churchland, 2021), supporting our hypothesis that—at least in certain settings—chunking does not occur at the level of motor cortex.

KaoSadabadiHennequin21

Kao, T.-C., Sadabadi, M. S., & Hennequin, G. (2021). Optimal anticipatory control as a theory of motor preparation: A thalamo-cortical circuit model. Neuron, 109(9), 1567-1581.e12. https://doi.org/10.1016/j.neuron.2021.03.009

Fast ballistic movements (e.g., throwing) require spatially and temporally precise commands to the musculature. Many of these signals are thought to arise from internal dynamics in the primary motor cortex (M1; Figure 1A; Evarts, 1968; To- dorov, 2000; Scott, 2012; Shenoy et al., 2013; Omrani et al., 2017). In turn, consistent with state trajectories produced by a dynamical system, M1 activity during movement depends strongly on the ‘‘initial condition’’ reached just before move- ment onset, and variability in initial condition predicts behav- ioral variability (Churchland et al., 2006; Afshar et al., 2011; Pandarinath et al., 2018). An immediate consequence of this dynamical systems view is the so-called optimal subspace hy- pothesis (Churchland et al., 2010; Shenoy et al., 2013): the network dynamics that generate movement must be seeded with an appropriate initial condition prior to each movement. In other words, accurate movement production likely requires fine adjustment of M1 activity during a phase of movement preparation (Figure 1B, green).

Moreover, optimal feedback inputs, but not feedforward inputs, robustly orthogonalize preparatory- and movement-epoch activ- ity, thus accounting for one of the most prominent features of perimovement activity in reaching monkeys (Kaufman et al., 2014; Elsayed et al., 2016).

Finally, we propose a way in which neural circuits may imple- ment optimal anticipatory control. In particular, we propose that cortex is actively controlled by thalamic feedback during motor preparation, with thalamic afferents providing the desired optimal control inputs. This is consistent with the causal role of thalamus in the preparation of directed licking in mice (Guo et al., 2017). Moreover, we posit that the basal ganglia operate an ON/OFF switch on the thalamo-cortical loop (Jin and Costa, 2010; Cui et al., 2013; Halassa and Acsa ́dy, 2016; Logiaco et al., 2019; Logiaco and Escola, 2020), thereby flexibly control- ling the timing of both movement planning and initiation.

The loop is gated ON/OFF at preparation onset/ offset by the (dis)-inhibitory action of basal ganglia outputs (Jin and Costa, 2010; Cui et al., 2013; Dudman and Krakauer, 2016; Halassa and Acsa ́ dy, 2016; Logiaco et al., 2019; Logiaco and Escola, 2020). S pecifically, cortical excitatory neurons project to 160 thalamic neurons, which make excitatory backprojections to a pool of 100 excitatory (E) and 100 inhibitory (I) neurons in cortex layer 4. In turn, these layer 4 neurons provide both excitation and inhibition to the main cortical network, thereby closing the control loop. Here, inhibition is necessary to capture the negative nature of optimal feedback.

[rcoreilly]

Churchland et al: Monkey Motor Cortex

ZimnikChurchland21

Zimnik, A. J., & Churchland, M. M. (2021). Independent generation of sequence elements by motor cortex. Nature Neuroscience, 24(3), Article 3. https://doi.org/10.1038/s41593-021-00798-5

We recorded from motor cortex (dorsal premotor and primary motor cortex) in monkeys trained to execute two-reach sequences. Monkeys were required to modulate the timing between reaches, performing them either in rapid succession (compound reach) or separated by an imposed pause (delayed double-reach). As expected, the imposed pause produced clearly independent prep- aration and execution: the prepare–trigger–execute motif was observed twice. Unexpectedly, independent preparation and execu- tion persisted during compound reaches. This was true even though the full sequence was known in advance, had been practiced tens of thousands of times and unfolded so swiftly that muscle activity for the for the second reach began as the first ended. Neural responses revealed how this rapid pace was possible: preparation for the sec- ond reach occurred during execution of the first, yet did not disrupt ongoing execution. Thus, at the level of motor cortex, skilled perfor- mance depends not on fusing elements, but on the ability to swiftly prepare one element while executing another.

The dominant aspect of the motor cortex response is a condition-invariant signal (CIS) that begins roughly 150 ms before reach onset and is highly predictive of the moment of reach onset (KaufmanSeelySussilloEtAl16 20,21). A similar CIS occurs in recurrent networks trained to produce reach-related muscle activity (SussilloChurchlandKaufmanEtAl15 31), and relates to the triggering of output-generating dynamics. In agreement, the dominant response component in mouse motor cortex during reaching depends on thalamic inputs, whose silencing interrupts execution (Dacre et al; SauerbreiGuoCohenEtAl19 32,33). Under the assumption that the CIS relates to movement triggering, there should be a monophasic CIS if compound reaches are generated holistically and a biphasic CIS if they are generated sequentially.

Our results also help extend the prepare–trigger–execute model. Preparation has historically been considered a time-consuming process17,36,40 and thus to contribute to the longer time between sequence elements at chunk boundaries5. This made it reasonable to assume that minimizing temporal separation required a holistic strategy. Yet preparation can be remarkably swift for well-practiced movements25,41. The present results confirm this swiftness, and illus- trate that preparation can overlap execution without interference. Given this, preparation need not prevent one movement from fol- lowing another with zero latency, or even negative latency, as has been proposed to occur during corrective movements42.

KaufmanSeelySussilloEtAl16

Kaufman, M. T., Seely, J. S., Sussillo, D., Ryu, S. I., Shenoy, K. V., & Churchland, M. M. (2016). The Largest Response Component in the Motor Cortex Reflects Movement Timing but Not Movement Type. eNeuro, 3(4). https://doi.org/10.1523/ENEURO.0085-16.2016

The largest condition-invariant component was much larger than any of the tuned components; i.e., it explained more of the structure in individual-neuron responses. This condition-invariant response component underwent a rapid change before movement onset. The timing of that change predicted most of the trial-by-trial variance in reaction time. Thus, although individual M1 and PMd neurons essentially always reflected which movement was made, the largest component of the population response reflected movement timing rather than movement type.

• The simplest explanation here is that this is thalamic gating of layer 5 PT neurons, consistent with Dacre et al, Sauerbrei, Guo et al. etc.

While a few neurons (Fig. 2C,D) had an unusually large contribution from the condition-invariant components, we found no evidence for separate populations of condition- invariant and condition-specific neurons. Weights wn,1 were continuously distributed, and could be positive (Fig. 2C) or negative (Fig. 2D). We also note that the average �wn, 1� was similar for neurons recorded in PMd and M1, indicating that the CIS is of similar size in the two areas.

If motor cortex undergoes a large condition-invariant change prior to movement, what drives that change? What other area(s) might supply the relevant input? A number of candidate regions exist, including the basal ganglia (Romo et al., 1992; Hauber, 1998), superior col- liculus (Werner, 1993; Philipp and Hoffmann, 2014), pari- etal cortex (Scherberger and Andersen, 2007; Pesaran et al., 2008), supplementary motor area (Orgogozo and Larsen, 1979; Eccles, 1982; Romo and Schultz, 1992), the dentate nucleus of the cerebellum (Meyer-Lohmann et al., 1977), and, in rodents, secondary motor cortex (Murakami et al., 2014).

yeah its the BG -> VM..

SussilloChurchlandKaufmanEtAl15

Sussillo, D., Churchland, M. M., Kaufman, M. T., & Shenoy, K. V. (2015). A neural network that finds a naturalistic solution for the production of muscle activity. Nature Neuroscience, 18(7), Article 7. https://doi.org/10.1038/nn.4042

PSTH for 5 different neurons (a) and 5 neurons from strongly regularized model (b). unconstrained model did not look like the real data.

Analysis of regularized model dynamics revealed a sequence of four events. First, during the preparatory period, condition-specific inputs produce a set of states (one per condition) that act as initial states for the upcoming movement-period dynamical system. Second, the movement-period dynamical system is produced by the simultaneous removal of the hold cue and the condition-specific inputs. Third, movement-period dynamics are dominated by a single, condition-independent fixed point with approximately linear and strongly oscillatory dynamics. Fourth, those dynamics yield neural trajectories whose projections onto the output dimensions produce the patterns of EMG. The similarity of this sequence in model and data lend support to the view that motor cortex concerns itself with low-level features of movement generation1,2,4,11,17,18,33,34.

DacreColliganAmmerEtAl19

Dacre, J., Colligan, M., Ammer, J., Schiemann, J., Clarke, T., Chamosa-Pino, V., Claudi, F., Harston, J. A., Eleftheriou, C., Pakan, J. M. P., Huang, C.-C., Hantman, A., Rochefort, N. L., & Duguid, I. (2019). Cerebellar-recipient motor thalamus drives behavioral context-specific movement initiation (p. 802124). bioRxiv. https://doi.org/10.1101/802124

Blocking MThDN/IPN output suppressed cued movement initiation. Stimulating the MThDN/IPN thalamocortical pathway in the absence of the cue recapitulated cue-evoked M1 membrane potential dynamics and forelimb behavior in the learned behavioral context, but generated semi-random movements in an altered behavioral context. Thus, cerebellar-recipient motor thalamocortical input to M1 is indispensable for the generation of motor commands that initiate goal-directed movement, refining our understanding of how the cerebellar-thalamocortical pathway contributes to movement timing.

Focal inactivation of MThDN/IPN suppressed layer 5 membrane potential dynamics in forelimb motor cortex (M1FL) and blocked cued movement initiation. Direct stimulation of MThDN/IPN neurons, or their axon terminals in M1FL, in the absence of the cue recapitulated motor cortical activity dynamics and forelimb behavior in the learned behavioral context, but generated semi-random movements in an altered behavioral context where the lever and reward were absent.

• Consistent with basic thal -> PT gating dynamic.

SauerbreiGuoCohenEtAl19

Sauerbrei, B. A., Guo, J.-Z., Cohen, J. D., Mischiati, M., Guo, W., Kabra, M., Verma, N., Mensh, B., Branson, K., & Hantman, A. W. (2019). Cortical pattern generation during dexterous movement is input-driven. Nature, 1–6. https://doi.org/10.1038/s41586-019-1869-9

Activation of thalamocortical axon terminals at different frequencies disrupted cortical activity and arm movement in a graded manner. Simultaneous recordings revealed that both thalamic activity and the current state of cortex predicted changes in cortical activity. Thus, the pattern generator for dexterous arm movement is distributed across multiple, strongly interacting brain regions.

• They don't exactly say where, but likely VM thalamus based on bregma coordinates.

• optogentic perturbation of thalamus in lots of different ways affected reach initiation and interrupted ongoing reaches.

[rcoreilly]

Dorsal Striosomes and Local Credit Assignment

Summary:

• Striosomes / patch directly inhibit SNc DA, and indirectly modulate via GPh / border GPi, consistent with PVLV model. This is strongly confirmed by modern methods, especially Watabe-UchidaZhuOgawaEtAI12

• All of the available evidence consistently shows that vmPFC areas (32 pgACC, 13 mOFC) project to striosomes, not dlPFC / ALM / M1 motor areas. Most is older: BerendseGalis-DeGraafGroenewegen92, RagsdaleGraybiel90, ElbenGraybiel95. ElbenGraybiel95 has the best overall data in monkey, and seems quite definitive: dlPFC areas 9/46 and principal sulcus 8/46 strongly and uniformly project to matrix and specifically avoid striosomes. Specific studies focusing on M1 projections consistently find it avoiding striosomes ParentParent06 FlahertyGraybiel93a.

Thus, the DLS striosomes are under BOA control, driving local SNc DA according to goal/outcome-related information, not according to specific motor plans.

GerfenBolam16

Gerfen, C. R., & Bolam, J. P. (2016). Chapter 1—The Neuroanatomical Organization of the Basal Ganglia. In H. Steiner & K. Y. Tseng (Eds.), Handbook of Behavioral Neuroscience (Vol. 24, pp. 3–32). Elsevier. https://doi.org/10.1016/B978-0-12-802206-1.00001-5

• deep layer 5 projects to striosomes; more superficial 5 and also 2/3 project to matrix: striosomes only get "gated" inputs!

A similar organization is also found following injec- tions into the infralimbic, anterior cingulate, medial agranular, and lateral agranular cortices (Berendse et al., 1992a; Gerfen, 1989). Differences between these cortical areas are mainly related to the relative density of inputs to the patch compartment. Infralimbic and prelimbic cortical injections provide denser inputs to the patch compartment, while cingulate and medial agranular cortical areas show a slightly less dense input to the patch compartment. The lateral agranular cortex provides only very sparse inputs to the patch compartment.

Thalamic afferents to the striatum from the parafas- cicular/intralaminar nuclei are also organized relative to the patch/matrix compartments (Beckstead, 1985; Berendse et al., 1988; Herkenham and Pert, 1981; Xu et al., 1991). Inputs from the parafascicular/centrome- dian thalamic nuclei provide inputs directed to the matrix compartment, whereas inputs to the striatal patch compartment arise from more restricted parts of the intralaminar thalamic nuclei.

This distribution pattern thus confirms axonal tracing studies, which suggest that patch compartment neurons target dopamine neu- rons in the SN and that matrix compartment neurons provide inputs to the GABA neurons in the SNr.

In one study, Cre-depen- dent retrograde transsynaptic viral vectors were injected into the SN of animals expressing Cre in dopa- mine neurons, which resulted in selectively labeling of neurons in the striatal patch compartment (Watabe- Uchida et al., 2012). In another study, injection of viral expression vectors identified the axonal projections of individual neurons in the patch compartment (Fujiyama et al., 2011).

Studies have shown that the striatal matrix compart- ment provides inputs directed to the thalamic project- ing part of the GPi, whereas the patch compartment provides inputs to the habenular projecting part of the GPi (Rajakumar et al., 1993). As the lateral habenula projects to the SNc, this suggests that the organization of the patch and matrix projections to the GPi provides differential pathways targeting GABAergic output neurons and dopamine neurons in the SNc. This is similar, though different, from the organization of striatal patch�matrix projections respectively to the SNc and SNr. Of note is the fact that the direct striatal patch projection to the SNc targets the ventral tier dopamine cell group, whereas the patch projection sys- tem through the GPi�habenular connections appears to provide inputs directed to the dorsal dopamine cell group. As discussed earlier, the dorsal dopamine cell group provides input directed to the matrix compart- ment, whereas the ventral dopamine cell group pro- vides input directed to the patch compartment.

When patch�matrix compartments are defined on the basis of input�output connectional systems, the organization in the dorsal and ventral striatum is similar.

BerendseGalis-DeGraafGroenewegen92

Berendse, H. W., Galis-De Graaf, Y., & Groenewegen, H. J. (1992). Topographical organization and relationship with ventral striatal compartments of prefrontal corticostriatal projections in the rat. The Journal of Comparative Neurology, 316(3), 314–347. http://dx.doi.org/10.1002/cne.903160305

Initially, the compartmental distribution of the corticostriatal inputs was related to the cytoarchitectonic area of origin. The prelimbic cortex was considered to project to the patch compartment throughout the striatum, whereas the projec- tions of the other cortical areas examined were thought to be confined to the matrix (Gerfen, '84; Donoghue and Herkenham, '86).

The prefrontal cortexcan be partitioned into medial, lateral or sulcal, and orbital or ventral subdivisions. The medial subdivision comprises the infralimbic area, the prelimbic area, the anterior cingular area, and the medial precentral area. The lateral subdivision includes the dorsal, ventral, and posterior agranular insular areas, whereas the orbital subdivision encompasses the medial
orbital area, the ventral orbital area, the ventrolateral orbital area, and the lateral orbital area. The medial precentral area of Krettek and Price ('77a) is identical to the medial agranular field of the motor cortex, as delineated by Donoghue and Parham ('83) and Reepetal.('87). In the present account, we will use the term medial agranular area. The medial orbital area will be included in the medial prefrontal cortex, as it is not possible, on the basis of cytoarchitectonics,to distinguish this area from the infra- limbic area with which it is caudally continuous (Van Eden and Uylings, '85).

• IL = 25, PL = pgACC (32), medial precentral / medial agranular is key question: this is secondary motor cortex (alm?)

Taking together the similarities in cytoarchitec- ture (Van Eden and Uylings, '851, thalamic afferents (Ber- endseandGroenewegen,'911,andstriatalefferents(present results), it can be argued that the medial orbital and infralimbic areas should be considered as a single cortical field. If the presence of projections from the mediodorsal nucleus is a prerequisite for the inclusion of a particular corticalareaintheprefrontalcortex,themedialorbitaland infralimbic areas should not be denoted as prefrontal
(Krettek and Price, '77a; Groenewegen, '88). However, since the projections from the parataenial thalamic nucleus to both areas (Krettek and Price, '77a; Berendse and Groenewegen, '91) seem to be comparable to those of the mediodorsal nucleus to the prefrontal cortex, the medial orbital and infralimbic areas are here included in the prefrontal cortex.

Further- more, rostrocaudal differences were observed within the projections of both the dorsal anterior cingular and the medial agranular areas. Rostra1 (pregenual) parts of these areas project ventrally and rostrally within the striatum, whereas the fibres from caudal (supragenual)parts termi- nate in more dorsal and caudal striatal regions. A similar pattern is apparent when the projections of the ventral part of the anterior cingular area are compared with those of the rostrally adjacent, dorsal part of the prelimbic area.

It does not look like AGm = medial agranular motor cortex projects much to patches.

The medial agranular area, particularly its rostra1 part, is generally considered as the rodent homo- logue of the frontal eye field,with additional supplementary motor and premotor characteristics (Hall and Lindholm, '74; Neafsey et al., '86a,b; Reep et al., '87).

RagsdaleGraybiel90

Ragsdale, C. W., & Graybiel, A. M. (1990). A simple ordering of neocortical areas established by the compartmental organization of their striatal projections. Proceedings of the National Academy of Sciences, 87(16), 6196–6199. https://doi.org/10.1073/pnas.87.16.6196

Note: this is in cats -- not much dlPFC to look at here:

• Parietal never fills patch, but does fill dorsal matrix (DLS). This makes sense for spatial sensory inputs guiding movements.
• Dorsomedial prefrontal = ACC 24 (most likely) -> DLS patch, DMS matrix. Motor effort activates DLS patch, closed loop with DMS. Yes.
• Ventrolateral prefrontal = OFC (most likely) -> DLS (barely), DMS patch; VS matrix (fairly dorsal)
• Insular (maybe some vmPFC, 25 in there?) -> DMS, VS patch; VS, VVS (basal) matrix
• Rostral temporal (perirhinal??) -> DMS, VS, VVS striosomes; VS, VVS matrix

Bottom line is that DLS patch gets ACC 24, some of OFC -- this is not really the motor program actually.. more BOA level inputs. Consistent with a general story that patch gets preferentially "limbic" inputs.

FlahertyGraybiel93a

Flaherty, A. W., & Graybiel, A. M. (1993). Two input systems for body representations in the primate striatal matrix: Experimental evidence in the squirrel monkey. Journal of Neuroscience, 13(3), 1120–1137. http://www.ncbi.nlm.nih.gov/pubmed/7680067

These experiments demonstrate the existence of two sets of motor cortical input zones within the matrix of the squirrel monkey striatum (Fig. 10). The first receives somatotopically matched, largely convergent inputs from ipsilateral MI and SI. The second receives inputs from contralateral MI. Both sets of zones are distinctly modular, and-with the apparent exception of those in the face region of the putamen-the two sets are most often nonoverlapping. Our results suggest that (1) the pu- tamen does not segregate motor and somatosensory information as much as it segregates information about the ipsilateral and contralateral body, (2) the putamen receives more information from contralateral MI about axial body parts than distal ones, and (3) matrisomes receiving motor and somatosensory infor- mation about a given body part are positioned such that they can interact along their borders with matrisomes receiving in- formation from nearby ipsilateral body parts, with matrisomes receiving information from the contralateral body part, with other matrisomal systems with unknown inputs, and with strio-somes. The sensorimotor input map in the putamen is thus considerably patchier than maps in other CNS representations of the body.

At all levels of the striatum at which both striosomes and MI projections were visible, both ipsilateral and contralateral motor cortex projections predominantly avoided striosomes.

• Confirmed: motor does not go to striosomes.

EblenGraybiel95

Eblen, F., & Graybiel, A. M. (1995). Highly restricted origin of prefrontal cortical inputs to striosomes in the macaque monkey. Journal of Neuroscience, 15, 5999. http://www.ncbi.nlm.nih.gov/pubmed/7666184

Selective innervation of striosomes was found for only 5 of the 20 cortical injections made.Three of the injections sites involved the posterior orbital cortex, and the other two were centeredi n the medial prelimbic cortex at its border with the anterior cingulate cortex.

In monkeys in which the dorsolateral prefrontal cortex or cortex in and near the principal sulcus was injected, there was dense and
apparently nearly uniform labeling of the matrix within the striatal projection fields (Fig. 2). There were pockets of very low, near background levels of labeling within these fields, and every such zone that we could score corresponded to a striosome. This pro-
jection pattern is shown in Figure 2 for case IB, in which a large tracer deposit was placed in the dorsal and ventral banks of the principal sulcus. A similar projection pattern was seen in case 13A, in which a large injection of tracer was made in the dorsolateral prefrontal cortex (Fig. 4B). It was notable that within the main projection fields labeled in these cases, the labeled inner- vation appeared nearly to fill the matrix; the borders of the held were themselves irregular with some outlying patches of label, but the main field was strongly labeled throughout.

• This seems definitive. no dlPFC -> striosomes.

These characteristics suggest that the major functional specializations of the striosome-projecting prefrontal cortex are tightly linked to those of the subcortical limbic system; these range from control of visceral responses and of emotions, such as the fear response for the amygdaloid complex, to the memory-re- lated functions posited for the hippocampal formation (see, e.g., Davis, 1992; Zola-Morgan and Squire, 1993).

Watabe-UchidaZhuOgawaEtAI12

Watabe-Uchida, M., Zhu, L., Ogawa, S. K., Vamanrao, A., & Uchida, N. (2012). Whole-brain mapping of direct inputs to midbrain dopamine neurons. Neuron, 74, 858–873. http://www.ncbi.nlm.nih.gov/pubmed/22681690

Previous studies have indicated that striatal neurons in the patches project to SNc, whereas those in the matrix project to SNr (Fujiyama et al., 2011; Gerfen, 1984), although a recent study indicated that these projections are not as specific as previously thought, at least in primates (Le ́ vesque and Parent, 2005), and the cell-type specificity of postsynaptic neurons has not been established.

ParentParent06

Parent, M., & Parent, A. (2006). Single-axon tracing study of corticostriatal projections arising from primary motor cortex in primates. The Journal of Comparative Neurology, 496, 202–213. http://www.ncbi.nlm.nih.gov/pubmed/16538675

The corticostriatal termi- nals are heterogeneously distributed within these puta- menal bands, being particularly abundant in small areas of the striatal matrix (matrisomes) but absent from striosomes (Flaherty and Graybiel, 1993). A precise topo- graphic distribution of the cortical motor input in the sensorimotor striatal territory is important because this territory harbors neurons directly involved in the execu- tion of movements (Parent, 1990; Flaherty and Graybiel, 1993).

This pattern of arborization suggests that projections aris- ing from very small cortical areas may extend through a large region of the striatum, implying that any given striatal cell receives convergent inputs from multiple cor- ticostriatal axons (Parent and Hazrati, 1995a).

The thalamostriatal projec- tion also appears more focused than the corticostriatal projection in primates. Thalamostriatal axons form dense terminal clusters concentrated in a relatively small por- tion of the striatum, indicating that relatively small num- bers of medium spiny striatal neurons are contacted by numerous terminals of thalamic origin. In contrast, the scarce and widely distributed corticostriatal innervation suggests that each M1 neuron provides only a few axonal varicosities to medium spiny striatal neurons scattered over a vast region of the putamen.

GuoLiHuberEtAl14 ALM paper

Location of ALM: bregma + 2.5mm -- fairly anterior and lateral.

source areas from BerendseGalis-DeGraafGroenewegen92

closest to ALM: a bit medial and thus in dorsal ACC / 24, not ALM, but closest case shown:

[rcoreilly]

Cerebellum, and <-> BG more generally

CaligiorePezzuloBaldassarreEtAl17

Caligiore, D., Pezzulo, G., Baldassarre, G., Bostan, A. C., Strick, P. L., Doya, K., Helmich, R. C., Dirkx, M., Houk, J., Jörntell, H., Lago-Rodriguez, A., Galea, J. M., Miall, R. C., Popa, T., Kishore, A., Verschure, P. F. M. J., Zucca, R., & Herreros, I. (2017). Consensus paper: Towards a systems-level view of cerebellar function: the interplay between cerebellum, basal ganglia, and cortex. Cerebellum (London, England), 16(1), 203–229. https://doi.org/10.1007/s12311-016-0763-3

Bostan & Strick

These experiments demonstrate that cerebellar output influences not only M1 but also premotor, prefrontal, and parietal areas (Fig. 1a, b). Output channels to M1 and the premotor areas are clustered in a dorsal region of the cerebellar dentate nucleus, identifying a motor domain within this nucleus. Output channels to prefron- tal cortex are clustered together in a ventral (non-motor) re- gion of the nucleus, entirely outside the motor domain [45] (Fig. 1b).

To date, all of the areas in the cerebral cortex that are targets of cerebellar output also have prominent projections to the cerebellar cortex via the pontine nuclei (Fig. 1a). Thin axon collaterals from approximately 20 % of ponto-cerebellar neu- rons reach the dentate nucleus [46] and may allow excitatory inputs from cortical areas, such as M1, to reach the dentate

a) is projections to pontine nuclei from cortex. It is pretty interesting how sparse it is from 9 / 46 and v/m PFC. But parietal is pretty strongly interconnected. Also S1 seems strangely sparse!

Furthermore, there is evidence that cerebellar output can alter striatal plasticity [52]. = ChenFremontArteaga-BrachoEtAl14

However, a closer look at the resultant behavior reveals that, rather than totally replacing reactive feedback commands, feed-forward cerebellar commands are merged with them [59, 65, 68].

In the instrumental contin- gency that uses air puffs as US, the anticipatory response es- tablishes a behavioral feedback that, as learning progresses, diminishes the sensed intensity of the aversive and learning- inducing US [65–67]. One may expect that the subject will completely avoid any aversive effect of the unconditioned stimulus by blinking preemptively, namely, to expect a com- plete substitution of reactive feedback by predictive feed- forward control after learning. However, behavior indicates that at learning asymptote subjects only display a partial antic- ipatory closure of the eyelid to the CS and will only completely close their eyes after perception of the US [59, 68–70]. This implies that a significant part of the protective action of the putative CR still takes place reactively and is thus a UR.

Thus, at the level of the cerebellum, this nucleo-olivary projection achieves, internally, what the behavioral negative feedback achieves externally: it reduces the intensity of the learning- inducing error signal, specifically, the driving of the activity of the IO cells above their baseline firing rate, which is in the range of 1 Hz.

plasticity in the cerebellar cortex because the excitation it pro- vides to the IO has been canceled by the NOI recruited by the cerebellar CR [73, 74]. In consequence, under the assump- tions of bidirectional plasticity in the cerebellar cortex and spontaneous activity in the IO, the NOI prevents the complete substitution of a feedback by a feed-forward mode of control [65] and thus implements a mixed feedback/feed-forward controller.

In the broader context of this consensus paper, we propose that the NOI realizes a mechanism that allows to contextually ad- just the contribution of distinct cortico-cerebellar loops to cor- tical computations and that such regulation might depend on neuromodulatory output from the basal ganglia.

If the basal ganglia initiate the release of motor programs or motor command signals via the neocor- tex, corticospinal axons will inevitably activate large parts of the circuitry in the spinal cord. As many neuronal components in the spinal circuitry project to the cerebellum, in addition to alpha-motor neurons or spinal motor pools [87], the spinocerebellar pathways are an important source of informa- tion about the neocortical activity [80, 87].

The spinocerebellar and spino-reticulo-cerebellar path- ways represent information from spinal motor circuits that are composed of various spinal interneurons. Almost all corticospinal tract axons terminate within the pool of spinal interneurons rather than the alpha-motor neurons directly [83, 84], which puts the interneurons in a key position for the majority of the motor control functions exerted by the brain. This pool of interneurons integrates the motor command sig- nal from the neocortex with local sensory feedback, such as cutaneous information, tendon organ afferents, and muscle spindle afferents both of type I and type II. Spinal interneurons can either have direct recurrent connections that ascend all the way to the cerebellum [85] or the lateral reticular nucleus [86, 87], or they can utilize specialized ascending neurons such as the neurons of Clarke’s column [80].

• perhaps this is why no connections from S1! it gets it all directly from the source!

Hence, the cerebellum can use the spinocerebellar systems to be informed about the relative excitatory drive on specific synergical components and on the low level motor functions resident in the spinal cord [89, 90].

In the case of spinocerebellar systems, such functions can already be repre- sented in the information that reaches the cerebellum, and the task of the cerebellum becomes that of finding which of these many functions should be associated with each other, and under which context they should be associated.

An interesting observation [52] is that optogenetic stimulation of the dentate-thalamo-striatal path- way reverts cortico-striatal long-term depression (LTD) to LTP, which had been reported to happen with dopamine stim- ulation [114]

ChenFremontArteaga-BrachoEtAl14

Chen, C. H., Fremont, R., Arteaga-Bracho, E. E., & Khodakhah, K. (2014). Short latency cerebellar modulation of the basal ganglia. Nature Neuroscience, 17(12), Article 12. https://doi.org/10.1038/nn.3868

To date there is no functional evidence for the presence of rapid communication between the cerebellum and the basal ganglia. Thus, whether cerebellum and basal ganglia can communicate rapidly, and indeed the utility of the disynaptic pathway as a short-latency conduit between the two structures, remains to be established.

In freely moving mice, we found that cerebellar stimulation effectively altered the activity of striatal neurons in about half of the neurons examined. Cerebellar modulation of striatal neurons had a short latency of ~10 ms and was mediated via the disynaptic cerebello-thalamo-striatal pathway. We explored whether this pathway is capable of altering corticostriatal plasticity.

We found that high-frequency stimulation of the cortex alone predominantly depressed corticostriatal responses. However,
concurrent activation of the cerebellum during high-frequency stimulation of the cortex altered the direction of the plasticity, resulting in the long-term potentiation of corticostriatal responses.

Selective activation of cerebellar axons in the intralaminar thalamic nuclei was as effective as direct stimulation of the dentate nuclei in altering the activity of striatal neurons and produced comparable response types and latencies in the majority of the cells examined.

Corroborating this hypothesis, we found that cerebellar-induced striatal responses were critically dependent on the state of the animals and were reversibly abated when the mice were anesthetized

PisanoDhanerawalaKislinEtAl21

Pisano, T. J., Dhanerawala, Z. M., Kislin, M., Bakshinskaya, D., Engel, E. A., Hansen, E. J., Hoag, A. T., Lee, J., de Oude, N. L., Venkataraju, K. U., Verpeut, J. L., Hoebeek, F. E., Richardson, B. D., Boele, H.-J., & Wang, S. S.-H. (2021). Homologous organization of cerebellar pathways to sensory, motor, and associative forebrain. Cell Reports, 36(12), 109721. https://doi.org/10.1016/j.celrep.2021.109721

Wealth of connectivity data!

Cerebellum -> cortex:

Cortex -> Cerebellum:

Nonmotor functions have been suggested for lobule VI in the posterior vermis and crus I in the posterior parts of the hemi- spheres (Badura et al., 2018; Stoodley et al., 2017). We found that lobule VI sends strong projections to MD and PV thalamic nuclei and to frontal neocortical regions, which serve a range of cognitive and affective functions (Marton et al., 2018; Yama- muro et al., 2020). Because postnatal refinement of neural circuitry is activity dependent (Hubel and Wiesel, 1965), this pro- jection may explain why lobule VI perturbations affect cognitive and social development in rodents (Badura et al., 2018) and hu- mans (Wang et al., 2014) and the association of posterior vermal abnormalities with a high risk of ASD (Courchesne et al., 1988; Limperopoulos et al., 2007).

Our observed connectivity patterns suggest the cerebellum in- corporates information from the somatomotor, somatosensory, and other cortexes to exert returning influence on a wide distri- bution of thalamic and neocortical targets. Our results provide an anatomical framework to be used by future studies to probe function and circuit mechanisms.

KangJunBaekEtAl21

Kang, S., Jun, S., Baek, S. J., Park, H., Yamamoto, Y., & Tanaka-Yamamoto, K. (2021). Recent Advances in the Understanding of Specific Efferent Pathways Emerging From the Cerebellum. Frontiers in Neuroanatomy, 15. https://www.frontiersin.org/articles/10.3389/fnana.2021.759948

• mRN = medial Red Nucleus

The mRN receives inputs from the IPN, but the pRN receives inputs from the DN (Basile et al., 2021). Consistent with the evolutionary difference, the IPN-mRN appears to be involved in motor functions. Specifically, studies have demonstrated that the pathway from the simplex lobule of the cerebellar cortex to the anterior IPN, and then to the mRN, is responsible for the delay eyeblink conditioning, which has often been used as a model system to investigate cerebellum-dependent associative motor learning (Freeman and Steinmetz, 2011)

The reticular formation includes many interconnected nuclei throughout the brainstem, and acts as a relay center for many fundamental brain functions, including somatic motor control (Mangold and Das, 2021). The DCN sends prominent projections to the reticular formation (Teune et al., 2000; Ruigrok and Teune, 2014). Three nuclei in the DCN, except for the posterior IPN, appear to contribute substantially to regulation of the pontomedullary reticular formation, which is the main source of the reticulospinal tract controlling posture and locomotion (Prentice and Drew, 2001; Stapley and Drew, 2009)

All three DCNs project to wide areas of the thalamus (Teune et al., 2000; Gornati et al., 2018; Fujita et al., 2020), although regions projected by the IPN shifted dorsolaterally and regions projected by the FN shifted ventromedially relative to the DN (Kebschull et al., 2020).

The efferent pathway from the cerebellum to the superior colliculus (SC) has also been long known (Angaut, 1969; Batton et al., 1977; Roldán and Reinoso-Suárez, 1981; Kawamura et al., 1982). All three nuclei of the DCN appear to project to the SC in a topographically distinct manner, although the DN and IPN send stronger projections than the FN (Roldán and Reinoso-Suárez, 1981; Doykos et al., 2020). The SC is thought to be an area for sensorimotor integration to initiate motor behaviors, including eye and head movements (Ito and Feldheim, 2018),

Alternatively, considering the functions of the SC in visually guided limb movement (Courjon et al., 2004; Steinmetz et al., 2019) and the importance of the cerebellum in controlling the precision of limb positions (Becker and Person, 2019; Wagner et al., 2021), the DCN-SC pathway may provide predictive information of limb positions to successfully achieve a target, as proposed (Doykos et al., 2020).

In addition to the abovementioned cerebellar efferent pathways to other brain regions associated with motor control, it has been long known that the cerebellum also sends direct projections to the spinal cord [e.g. (Thomas et al., 1956; Fukushima et al., 1977; Matsushita and Hosoya, 1978; Asanuma et al., 1980; Liang et al., 2011; Wang et al., 2018)]. A study using retrograde labeling further characterized the cerebellospinal pathways, demonstrating that distinct populations of DCN neurons in the anterior IPN send ipsilateral projections to the cervical, thoracic, and lumbar cords, whereas contralateral connections from the posterior IPN and FN are limited to the cervical cord (Sathyamurthy et al., 2020).

Interestingly, as nucleo-olivary, olivo-cortical, and corticonuclear pathways follow zonal network arrangements, a closed loop appears to be formed between the cerebellar cortex, DCN, and IO (Bengtsson and Hesslow, 2013; Chaumont et al., 2013). Thus, this nucleo-olivary pathway is thought to send feedback information to the IO, and was suggested to provide negative feedback mechanisms to block associative motor learning assessed in the eyeblink conditioning (Kim et al., 1998).

A recent study demonstrated that the nucleo-olivary pathway generated negative prediction error signals that proactively trigger the extinction of associative motor memory (Kim et al., 2020). In addition, the nucleo-olivary projections have been proposed to control synchronous activity of IO neurons (Lefler et al., 2014), which are electrically coupled through gap junctions. Indeed, another recent study showed that the inhibition of nucleo-olivary projections triggered synchronous activity of IO neurons, and this was the mechanism of state changes underlying skilled movement (Wagner et al., 2021).

KebschullRichmanRingachEtAl20

Kebschull, J. M., Richman, E. B., Ringach, N., Friedmann, D., Albarran, E., Kolluru, S. S., Jones, R. C., Allen, W. E., Wang, Y., Cho, S. W., Zhou, H., Ding, J. B., Chang, H. Y., Deisseroth, K., Quake, S. R., & Luo, L. (2020). Cerebellar nuclei evolved by repeatedly duplicating a conserved cell-type set. Science, 370(6523), eabd5059. https://doi.org/10.1126/science.abd5059

SengulWatson15

Sengul, G., & Watson, C. (2015). Chapter 8—Ascending and Descending Pathways in the Spinal Cord. In G. Paxinos (Ed.), The Rat Nervous System (Fourth Edition) (pp. 115–130). Academic Press. https://doi.org/10.1016/B978-0-12-374245-2.00008-5

Plenty of efferent copy signals going to the DCN.

The ventral spinocerebellar tract neurons are activated by group Ia and Ib afferents (Bras et al., 1988) and transmit information for coordinated movement and posture of the entire lower limb. They differ from those in the dorsal spinocerebellar tract in that they lack sub-divisions for different modalities, and transmit information from large receptive fields covering a range of different spinal cord segments (Kim et al., 1986).

The rostral spinocerebellar tract is the upper extremity homolog of the ventral spinocerebellar tract.

[rcoreilly]

Cerebellum: DCN Circuitry

UusisaariKnopfel11

Uusisaari, M., & Knöpfel, T. (2011). Functional Classification of Neurons in the Mouse Lateral Cerebellar Nuclei. The Cerebellum, 10(4), 637–646. https://doi.org/10.1007/s12311-010-0240-3

However, evidence clearly suggests that the DCN are not merely a relay station, but rather constitute a site of signal integration if not signal generation. Indeed, a role in adaptive information processing and perhaps pattern generation may be deduced from the presence of specific forms of synaptic plasticity within the DCN [2–4].

Fig 1

The mouse DCN comprises three distinct nuclei (medial, interposed, and lateral), each of which features a specific composition of neurons of various sizes [13]. There are significant differences in gross anatomy and neuronal morphology between the three DCN—especially the lateral DCN, also known as the “dentate nucleus” in primates [5, 14].

However, all of the DCN engage in two pathways of information flow (Fig. 1): the olivo-cortico- nucleo-olivary (OCNO) loop, and the pathway formed by mossy fibers, granule cell axons, Purkinje cells, and DCN projections to premotor nuclei [1, 15, 16].

These pathways are organized in compartments that follow the zonal topography of the cerebellar cortex, each of which is likely to be related to different modes of cerebellar function [17].

In vitro and in vivo recordings from unidentified DCN neurons have revealed that the vast majority of these neurons are spontaneously active at action potential (AP) firing rates of 35–55 Hz driven by tonic cation currents [20–24].

The GAD+IO neurons have not yet been carefully examined by intracellular recordings, but our few recordings suggest that they do exhibit spontaneous firing behavior that is similar to the local GABAergic neurons (mean firing frequency 7–10 Hz at 32°C; M. Uusisaari, unpublished observations).

It would thus seem that the DCN neurons cover a range of signal transduction modes, from tonic firing and rate-coding (GADnL) to burst-firing and onset or time-interval coding (GAD+, Gly-I).

This is unfortunate, as our targeted staining of these silent, GlyT2+ cells revealed that they project directly to the cerebellar cortex, unlike any other cell type that we have encountered in the DCN so far [12], making further elucidation of their behavior mandatory.

The actual synaptic targets within the cerebellar cortex remain unknown for these putatively glycinergic—and thus, inhibitory—neurons, but are likely to be cerebellar Golgi cells as these are the only neurons in most areas of the cerebellar granule cell layer that express glycine receptors [36].

Purkinje neuron (PN) axons from the cerebellar cortex provide the main efferent input to the DCN, and this is their main target. Based on counts of Purkinje and DCN neurons (estimates for mouse are ~200,000 PNs [37] and ~30,000 neurons in all three DCN [38]), signal transmission from the cortex to the nuclei shows overall convergence, but the true extent of this convergence at the cellular level is not known. Based on the count of Purkinje axon terminals on large DCN cell bodies, and taking into account various estimates of how many terminals a single axon will form on a given neuronal body and the spread of PN axons within the nuclei, estimates of the number of PNs contacting a single large DCN neuron range between several tens and several hundreds [39–41].

• DCN is lower dimensional relative to PN, but not as dramatically so as in BG

This led to the proposition that GABAergic synapses on GAD+ cells are predominantly formed on distal dendrites rather than cell bodies; on the other hand, PNs form large numbers of synapses on the bodies of large glutamatergic DCN neurons, to the extent that GAD− cell bodies seem to be entirely covered by GABAergic presynaptic terminals [5, 11].

Taken together, it appears that the slower IPSC time course seen in GAD+ cells arises from a combina- tion of distal dendritic localization of synapses and a specific receptor subtype composition, as well as possible differences in the spatiotemporal dynamics of GABA release [3, 45, 49].

• Potential for temporal derivative computation here with these strong temporal differences.

Conversely, it is known that climbing fiber (CF) collaterals contact both GAD+IO and GADnL DCN neurons and that mossy fiber (MF) collaterals synapse on GADnL cells, but the relative strength of these connections and whether MF or CF synapses occur on other cell types is unknown [5, 19]. Unpublished observations from our laboratory suggest, however, that both GAD+ and GADnS cells receive glutamatergic extracerebel- lar synaptic afferents (M. Uusisaari, unpublished observa- tions).

DAngeloAntoniettiCasaliEtAl16

D’Angelo, E., Antonietti, A., Casali, S., Casellato, C., Garrido, J. A., Luque, N. R., Mapelli, L., Masoli, S., Pedrocchi, A., Prestori, F., Rizza, M. F., & Ros, E. (2016). Modeling the Cerebellar Microcircuit: New Strategies for a Long-Standing Issue. Frontiers in Cellular Neuroscience, 10. https://www.frontiersin.org/articles/10.3389/fncel.2016.00176

The cerebellar inputs are elaborated in the GCL before being further processed in the ML and distributed to PCs, from which signals are sent to DCN. While signals flow along the GrC → PC → DCN neuronal chain, they are thought to undergo an initial “expansion recoding” in the GCL followed by a “perceptron-like” sampling in PCs before converging onto the DCN (the validity of these assumptions is further considered below).

Local computations in the cerebellar cortex are regulated by two extended inhibitory interneuron networks, one in the GCL and one in the ML. Since the DCN is also activated by mf collaterals, the cerebellar cortex de facto operates as a modulator of DCN activity.

Finally, the IO → PC → DCN neuronal chain forms another pathway probably implied in controlling network learning and timing capabilities. Recently, relevance has been given to recurrent DCN → GrC and DCN → IO connections, which can directly send output information back to the input.

The main input to the cerebellum comes through the mfs. The mfs originate from neurons located in the brain-stem nuclei (including the cuneate nucleus, vestibular nucleus, reticular nucleus, red nucleus and APN) and spinal cord (dorsal columns). Moreover, relevant to external connectivity, GrCs have recently been shown to receive a blend of modalities from brain-stem and cortical afferences (Huang et al., 2013; Ishikawa et al., 2015). In the GCL, mfs, GrC dendrites, GoC dendrites and axons interact into specialized structures called glomeruli. The mfs emit collaterals forming synapses in the DCN. The other important input originates from a brain-stem nucleus, the IO, giving rise to the cfs contacting PCs and DCNs.

Therefore, perhaps the most striking aspect in the cerebellar microcircuit is that, while mfs, cfs, GoC axons and PC dendrites are oriented longitudinally, they are orthogonal to the pfs that cross the PC dendritic trees.

The cerebellum is characterized by two extended inhibitory interneuron networks. The GCL layer inhibitory network is made of feedforward and feedback loops driven by mfs: (i) the mfs excite GrC and GoC dendrites and these latter inhibit GrCs in a feedforward loop, and (ii) the mfs excite GrCs and then pfs excite GoCs and these latter inhibit GrCs in a feedback loop (Simões de Souza and De Schutter, 2011; Mapelli et al., 2014). The GoCs are interconnected through gap-junctions and reciprocal inhibitory synapses. The ML inhibitory network is formed by a series of MLIs (SCs and BCs) activated by pfs and inhibiting PCs in feed-forward (Santamaria et al., 2002, 2007). The MLIs are interconnected through gap-junctions and reciprocal inhibitory synapses (Astori et al., 2009; Alcami and Marty, 2013).

The most renowned recurrent loop passes through the IO. The small DCN GABAergic neurons inhibit the IO cells regulating their coupling and oscillations (Najac and Raman, 2015).

The DCNs are involved in the cerebellar circuitry with a one way connection between the glycinergic DCN, projecting to the GCL, inhibiting GABAergic GoCs and the glutamatergic DCN that excite the GRCs and GOCs (Ankri et al., 2015; Houck and Person, 2015; Gao et al., 2016). A similar connectivity characterizes the medial vestibular nucleus in the vestibulo-cerebellum.

The state of the art for the cerebellar GCL is currently set by the 2010 model (Solinas et al., 2010), which was intended to generate a core computational element of the GCL microcircuit (about 10,000 neurons).

Perhaps the aspect most relevant to cerebellar modeling on the mesoscale is the organization of subcircuits, in which the cfs and the mfs contacting a certain group of PCs and DCN neurons are connected to the same area of origin to form fully connected cerebellar modules. Furthermore, the cerebellar modules can be organized according to the longitudinal stripes, in which some neuronal and synaptic mechanisms are differentiated depending on the type (Z+ or Z−) of the stripe (Wadiche and Jahr, 2005; Wang et al., 2011; Zhou et al., 2014).

BroersenAlbergariaCarulliEtAl23

Broersen, R., Albergaria, C., Carulli, D., Carey, M. R., Canto, C. B., & De Zeeuw, C. I. (2023). Synaptic mechanisms for associative learning in the cerebellar nuclei. Nature Communications, 14(1), Article 1. https://doi.org/10.1038/s41467-023-43227-w

The cerebellum has a remarkable ability to learn sensorimotor asso- ciations. One well-known example of such associative learning is delay eyeblink conditioning (EBC), a Pavlovian task where well-timed con- ditioned eyelid closures (conditioned response; CR) arise after pairing a neutral conditioned stimulus (CS) with an aversive unconditioned stimulus (US) at a fixed interval, with both stimuli co-terminating in time1,2.

The main inputs of the cerebellar system that convey the CS and US information have been characterized. The current working hypothesis is that the CS is at least partially conveyed to granule cells in the cerebellar cortex by mossy fiber (MF) axons originating in the pontine nuclei (PN)3,4. Granule cells give rise to parallel fibers and innervate Purkinje cells (PCs). Climbing fibers, derived from the inferior olive and powerfully innervating PCs, respond strongly to the US 5.

As incoming MFs and climbing fibers innervate both the cerebellar cortex and the cerebellar nuclei (CN)6–10, both sites are strong candi- dates that may contribute to the plasticity mechanisms that underlie associative learning.

In the cerebellar cortex, synaptic weight changes at parallel fiber synapses onto PCs11–13 and molecular layer interneurons14,15, potentially in combination with intrinsic mechanisms16 and structural changes of axonal fibers17,18, lead predominantly to reductions in PC simple spike activity during the CS-US interval14,19,20. This in turn leads to disinhibi- tion of CN neurons21,22 and ultimately increased spiking activity of premotor and motor neurons downstream in the red nucleus and facial nucleus, respectively23,24, thereby generating a conditioned eye- lid movement.

Within the complex of the CN, the anterior interposed nucleus (AIP) contains most EBC encoding neurons21,22,25. Manipulating their activity disrupts both the acquisition and expression of CRs26–28. Fur- thermore, CRs remain partly intact after removal of the cerebellar cortex in fully conditioned mice29, suggesting that part of the memory is stored at the level of the nuclei30. However, in contrast to the synaptic plasticity mechanisms in the cerebellar cortex that contribute to eyeblink conditioning, those in the CN remain largely obscure.

Together, our findings raise the intriguing possibility that there is the capacity for sensorimotor learning and temporal coding within the CN themselves, carried by a synergy of structural and physiological changes at the level of individual nuclei neurons.

Following conditioning, the predominant response of AIP neurons is spike facilitation induced by the CS, with the spike rate increasing during learning22.

We show that stimulating MF afferents to AIP as a CS can effectively drive learning, with conditioned mice exhibiting well-timed CRs that are adaptive to changing CS-US delays. Chemo- genetic inhibition of cerebellar granule cells selectively impairs learn- ing to a cortical optogenetic CS, but leaves learning to a MF-AIP CS intact. Moreover, learning with MF-AIP stimulation as a CS is not modulated by locomotor activity, in contrast to learning with cortical MF stimulation31. Taken together, these results strongly suggest that there is a capacity for well-timed learning within the AIP, without the need for MFs to convey CS information to the cerebellar cortex55–57.

We further show that learning leads to structural changes in both excitatory MF and inhibitory inputs to the AIP, while MF input to the granule cell layer remains unchanged. At the same time we find that optogenetic stimulation of the PN neurons, the main source of MF input to the cerebellum, elicits larger eyelid movements after con- ditioning. Last but not least, by performing in vivo whole-cell record- ings in awake behaving mice during EBC, we show how these learning- induced changes shape the Vm dynamics of AIP neurons. We find that conditioned neurons sustain a depolarized Vm state until US onset, while naive mice show transient Vm depolarizations. After conditioning the Vm response becomes temporally precise in that it starts earlier than eyelid closures, away from the initial delay between Vm response and eyelid openings in naive mice. This indicates that conditioned AIP neurons could play a role in initiating the CR, whereas naive neurons likely process an efference copy of the eyelid opening with movements initiated by other brain areas53,54.

• This is direct evidence that naive animals show late activation at time of US, but conditioned animals use CS signal from MF to predict that US -- basic temporally advanced predictive learning. In this context, the function of the cerebellar cortex (CC) is to further modulate that basic CS -> US association formed in the DCN, according to broader state variables. This further modulation is disinhibitory on the basic CS -> US DCN association -- so it can modulate that.

There is broad consensus that CRs emerge as the result of well- timed pauses in PC activity in response to the CS14,19, where plasticity in the AIP would be under the control of PC inputs58. However, our results indicate that at least part of the memory trace may be generated and stored directly at the level of the nuclei30, at least in part through structural modifications of VGLUT1 + MF and inhibitory inputs. This would be in agreement with the observation that many AIP neurons gradually increase their spike rates over the course of learning, coin- ciding with larger conditioned eyelid movements22.

Lesion and reversible inactivation studies have suggested that these two areas may encode different features of learning, with plasticity at MF-AIP synapses mediating the expression of motor responses, while cortical cells regulate their gain and timing21,28,63–72. One of the most influential models in the field posits that temporally specific plasticity in the cerebellar cortex leads, over learning, to well-timed pausing of PCs, and, consequently, a downstream disinhibition of AIP neurons63. Indeed, AIP neurons, which are like PCs highly intrinsically active22,50, are sensitive to the differential impact of synchronized activity of PCs during behavior73.

• 63 = PerrettRuizMauk93, 73 = ZhengRaman10

[rcoreilly]

Contreras-LopezAlatriste-LeonDiaz-HernandezEtAl23

Contreras-López, R., Alatriste-León, H., Díaz-Hernández, E., Ramírez-Jarquín, J. O., & Tecuapetla, F. (2023). The deep cerebellar nuclei to striatum disynaptic connection contributes to skilled forelimb movement. Cell Reports, 42(1), 112000. https://doi.org/10.1016/j.celrep.2023.112000

Highlights
• The DCN are connected with the striatum mainly through the CL, PF, and the VL
• Inhibition of the DCN→thalamo→striatal or the DCN→PF projections impairs reaching
• Inhibition of the DCN→VL and DCN→CL projections impairs unidirectional displacements
• Inhibition of the DCN→Th→Str pathways impairs the execution of sequential displacements

HabasMantoCabaraux19

Habas, C., Manto, M., & Cabaraux, P. (2019). The Cerebellar Thalamus. The Cerebellum, 18(3), 635–648. https://doi.org/10.1007/s12311-019-01019-3

KunimatsuSuzukiOhmaeEtAl18

Kunimatsu, J., Suzuki, T. W., Ohmae, S., & Tanaka, M. (2018). Different contributions of preparatory activity in the basal ganglia and cerebellum for self-timing. eLife, 7, e35676. https://doi.org/10.7554/eLife.35676

Here, we show that in monkeys trained to generate a self-initiated saccade at instructed timing following a visual cue, neurons in the caudate nucleus kept track of passage of time throughout the delay period, while those in the cerebellar dentate nucleus were recruited only during the last part of the delay period. Conversely, neuronal correlates of trial-by-trial variation of self-timing emerged earlier in the cerebellum than the striatum. Local inactivation of respective recording sites confirmed the difference in their relative contributions to supra- and sub-second intervals. These results suggest that the basal ganglia may measure elapsed time relative to the intended interval, while the cerebellum might be responsible for the fine adjustment of self-timing.

• Nice demonstration that BG = outer loop motor timing, Cerebellum = fine-grained inner loop timing

WolffKoOlveczky22

Wolff, S. B. E., Ko, R., & Ölveczky, B. P. (2022). Distinct roles for motor cortical and thalamic inputs to striatum during motor skill learning and execution. Science Advances, 8(8), eabk0231. https://doi.org/10.1126/sciadv.abk0231

To address this, we silenced the inputs in rats trained on a task that results in highly stereotyped and idiosyncratic movement patterns. While striatal-projecting motor cortex neurons were critical for learning these skills, silencing this pathway after learning had no effect on performance. In contrast, silencing striatal-projecting thalamus neurons disrupted the execution of the learned skills, causing rats to revert to species-typical pressing behaviors and preventing them from relearning the task.

We targeted the thalamic nuclei that provide the main source of thalamic input to the DLS, the Pf, rILN, and midline nuclei (Fig. 4A and fig. S6B) (29, 32, 33, 64, 65) but not motor or somatosensory thalamic nuclei.

While these results demonstrate that DLS-projecting thalamus neurons in the Pf and rILN are necessary for executing the learned motor skills we train, their specific function remains unclear.

Furthermore, our experiments did not address whether the nonhomogeneous thalamus inputs have distinct functions in this process (33). Inputs originating from Pf and the rILN differ in striatal projection patterns (broad versus focused arborization and targeting of dendritic shafts versus spines) (71) and physiological characteristics such as short-term plasticity (72). These differences and the different information the individual inputs may transmit could indicate distinct contributions to both learning and execution of new motor skills and to memory formation and storage.

• PF, CL -> Str is a chaining auto-associative pathway, key for keeping the dynamics moving along, in response to efferent copy inputs from CL too.

DeZeeuwI21

De Zeeuw, & I, C. (2021). Bidirectional learning in upbound and downbound microzones of the cerebellum. Nature Reviews Neuroscience, 22(2), Article 2. https://doi.org/10.1038/s41583-020-00392-x

However, there still seem to be contradictions between the various hypotheses with regard to the mechanisms underlying cerebellar learning. The challenge is therefore to reconcile these different views and unite them into a single overall concept. Here I review our current understanding of the changes in the activity of cerebellar Purkinje cells in different ‘microzones’ during various forms of learning. I describe an emerging model that indicates that the activity of each microzone is bound to either increase or decrease during the initial stages of learning, depending on the directional and temporal demands of its downstream circuitry and the behaviour involved.

The specificity of the actions of microzones is supported by their targeted projections: each microzone projects to a particular set of cerebel- lar and/or vestibular nuclei, together establishing a microcomplex21–23. Even though the cerebellar and ves- tibular nuclei comprise a bewildering palette of differ- ent projection neurons and local interneurons24–27, the dendrites of which can extend even into neighbouring regions28, the baseline activity patterns of the main projection neurons in these nuclei appear to adhere to the organization of the microcomplexes29. Moreover, the cerebellar and vestibular nuclei of individual micro- complexes in turn inhibit the specific subnucleus of the inferior olive from which the corresponding Purkinje cells receive their climbing fibre input30–32. As a result, the olivocerebellar system as a whole comprises many parallel micromodules that are formed by topographi- cally organized three-element loops, comprising specific sets of Purkinje cells, neurons in the cerebellar and/or vestibular nuclei and inferior olive neurons31,33–35 (Fig. 1b).

What is the challenge? Over the past half century the prevailing view has been that cerebellar motor learning in vertebrates depends largely on long-term depression (LTD) at the parallel fibre to Purkinje cell synapse52,54. This synaptic change can be induced when parallel fibres and climbing fibres are activated simultaneously or within a relatively narrow time frame55–58. However, evi- dence has been accumulating that this synapse can also be subject to long-term potentiation (LTP) and that many other forms of synaptic and non-synaptic plasticity also occur in the cerebellar cortex as well as downstream in the cerebellar and vestibular nuclei53,59–68. It has also been shown that the Purkinje cells of different micro- zones often show opposite changes in their simple spike activity during the expression of learned behaviours69.

The second key to understanding how the learning cerebellum may operate is the notion that the upbound or downbound nature of the microzone involved in a particular behaviour can probably be predicted by the net polarity of neurotransmissions in the downstream cir- cuitry that mediates it. More specifically, given the inhib- itory nature of Purkinje cells90, they can be expected to participate in an upbound or downbound microzone if the net polarity of the downstream circuitry is inhibitory or excitatory, respectively. In this way, the target neuron at the end of the downstream circuit that mediates the initial primary effect will, in general, be automatically excited.

TeunevanderBurgvanderMoerEtAl00

Teune, T. M., van der Burg, J., van der Moer, J., Voogd, J., & Ruigrok, T. J. H. (2000). Topography of cerebellar nuclear projections to the brain stem in the rat. In Progress in Brain Research (Vol. 124, pp. 141–172). Elsevier. https://doi.org/10.1016/S0079-6123(00)24014-4

The results presented here clearly show that small injections placed into selected parts of the CN invariably result in terminal projection patterns that are distributed over wide areas of the brain stem. It is likely that these output patterns relate to single cerebellar modules or submodules as based on intrinsic and olivocerebellar connections (Buisseret-Delmas and Angaut, 1993; Voogd et al., 1996).

Moreover, the material shows that the projections to the IO are usually restricted to a single patch in one of its subdivisions indicating that in these cases the injection relates to a homogeneous olivocerebellar projection zone (Ruigrok, 1997; Voogd, 1995). Hence, we conclude that the modular output is distributed to and processed in parallel in many brain stem areas.

In addition, as discussed above, several brain stem regions that receive a conspicuous cerebellar input, such as the medial mesodiencephalic region, pretectum, and DpMe, may serve as a source of IO input (Bull and Berkley, 1991; De Zeeuw and Ruigrok, 1994). (4) Several areas may be classified as 'effector' areas, e.g. such as the RN and supraoculomotor nuclei. Some effector regions, such as the vestibular nuclei, the SC and the reticular formation and the trigeminal nuclei, also serve as sources of mossy fiber input and as pre-olivary relay nuclei.

• Efferent copy signals come back from brainstem motor areas for sure, into mossy fiber inputs and IO training signals -- dual role as predictors and trainers.

• In fact, in many cases, most diencephalic terminals were observed in the ZI and/or the parafascicular thalamic nucleus.

cerebellum also contributes to PF?

YoshidaOnateKhatamiEtAl22

Yoshida, J., Oñate, M., Khatami, L., Vera, J., Nadim, F., & Khodakhah, K. (2022). Cerebellar Contributions to the Basal Ganglia Influence Motor Coordination, Reward Processing, and Movement Vigor. Journal of Neuroscience, 42(45), 8406–8415. https://doi.org/10.1523/JNEUROSCI.1535-22.2022

[rcoreilly]

SNr recordings

The single most informative type of data for BG influence over motor control would be direct recordings from SNr. Surprisingly, this is relatively rare data -- so many recordings of MSNs, so little from SNr / GPi!

BarterLiSukharnikovaEtAl15

Barter, J. W., Li, S., Sukharnikova, T., Rossi, M. A., Bartholomew, R. A., & Yin, H. H. (2015). Basal Ganglia Outputs Map Instantaneous Position Coordinates during Behavior. Journal of Neuroscience, 35(6), 2703–2716. https://doi.org/10.1523/JNEUROSCI.3245-14.2015

Shows modulation of eye movement parameters in individual SNr outputs

LarryZurJoshua24, ZurLarryJoshua23

Larry, N., Zur, G., & Joshua, M. (2024). Organization of reward and movement signals in the basal ganglia and cerebellum. Nature Communications, 15(1), 2119. https://doi.org/10.1038/s41467-024-45921-9

Zur, G., Larry, N., & Joshua, M. (2023). High-Dimensional Encoding of Movement by Single Neurons in Basal Ganglia Output. bioRxiv, 2023–05. https://www.biorxiv.org/content/10.1101/2023.05.17.541090.abstract

SNr has higher dimensional, sharper firing for smooth eye pursuit actions, compared to other areas.

Falasconi, Kanodia, Arber, in prep (at GRC meeting)

• Ferreira-PintoKanodiaFalasconiEtAl21, YangKanodiaArber23 related papers

Very strong evidence of specific disinhibition of brainstem motor (lateral rostral medulla = latRM) by individual SNr neurons.