To assess the correspondence between distinct basal ganglia pathways and hypothesized cognitive processes, we applied the high spatio-temporal resolution of single-unit electrophysiology to a stop-signal task based around our prior studies of rodent decision-making20,21
(, Table S1
). Each rat was trained to place its nose in a central port until the onset of a Go cue (1kHz or 4kHz tone) that directed a brief lateral head movement (to left or right; see Supplemental Movie
). On 30% of trials the Go tone was followed by a Stop cue (white noise), instructing that the rat should stay in the central port (). For both Go trials and Stop trials correct performance was rewarded by delivery of a sugar pellet. As typically observed for stop-signal tasks, reaction times for Failed Stop trials corresponded to the faster portion of the Go trial reaction time distribution (). This is consistent with race models: when the Go process happens more quickly, a Stop process is less likely to successfully suppress behavior.
For our first set of recordings (Experiment 1), well-trained subjects (n=4) received tetrode implants that simultaneously targeted striatum, STN, globus pallidus (GP), and SNr (for further details see ref. 21
). We isolated spikes from individual neurons during task performance, from each brain region (for anatomical locations see Fig. S1
). A challenge when studying behavioral inhibition is to disentangle neural activity specifically linked to stopping, rather than going. To do this, we followed a latency matching procedure22,23
(see Methods), which exploits the similarity in reaction times -and thus presumably the Go process - between Failed Stop trials and Fast Go trials. We compared the firing rate of each neuron between these trial types, and between Correct Stop trials and Slow Go trials. We then assessed the fraction of each neuronal population that showed significant differences at each moment in time ().
Figure 2 Distinct processing of the Stop cue across basal ganglia components. (a) For each brain area, bars indicate the fraction of neurons whose firing rate significantly differs between the trial types under comparison. To screen for Stop-related activity, (more ...)
Striatal neurons showed little or no fast population-level response to Stop cues. In contrast, both STN and SNr contained a significant proportion of neurons with rapid responses to the Stop signal (, red and purple filled bars). For STN, this population response was the same for Correct and Failed Stop trials (for the two STN bins just after Stop cue onset, p = 0.17 and 0.21, shuffle test) and thus resembled a fast “sensory”-like response to the Stop cue (see for a representative single unit example). Strikingly however, for SNr this fast change in activity was only
observed for Correct, rather than Failed Stop trials (for the two filled red SNr bins, p = 0.008 and 0.005, shuffle test; see also for a single unit example and Fig. S3
for a direct comparison between Correct and Failed Stop trials). Thus, while the activity in STN was consistent with a sensory response, activity in SNr instead reflected the behavioral outcome on each trial.
Our GP recordings did not yield such unambiguous results. Although the initial screen indicated that some neurons may selectively respond for Correct Stops (), the direct comparison did not confirm a selective GP response in Correct, rather than Failed Stop trials (Fig. S3
). We therefore focus on STN and SNr below.
We next examined the time course of activity in these Stop-related STN and SNr neurons. STN neurons responded to the Stop cue onset with transiently increased firing (), that in some cases took the form of just a single, precisely timed extra spike (). These STN increases had consistently very low latencies (peak response ~15 ms; ; see ref. 24
for similarly low STN latencies) that were not different between Correct and Failed Stop trials (p = 0.41, paired t-test; ). The magnitude of the peak STN response also showed no systematic population preference for Correct versus Failed Stop trials (). SNr neurons also increased firing to the Stop cue (), but with a longer latency (peak response ~36ms; ; p = 0.004 comparing STN to SNr latencies, one-sided Kolmogorov-Smirnov test) and preferentially on Correct Stop trials (). This latency difference was seen even when the analysis was restricted to units recorded in the same session (n = 15 pairs; STN cells preceded SNr cells by an average of 13.6 ms, p = 0.041, shuffling statistic). All SNr neurons that responded to the Stop cue on Correct trials did so before the Stop-Signal Reaction Time (SSRT; grey lines in ), a standard, inferred behavioral measure for how quick a process must be to influence stopping performance3,4,7
. Thus, SNr activity not only distinguishes between Correct and Failed Stop trials, it does so quickly enough to affect the trial outcome.
Figure 3 Stop cues increase firing in STN before SNr. (a) Firing rate time courses for the neuronal subpopulations that distinguish Stop from Go trials (in contralateral trials; see and Methods). In each case colored lines show the mean (± s.e.m.) (more ...)
Most of the SNr units with fast Stop cue responses (10/18) also markedly decreased their activity beginning just before movement initiation (Fig. S3
). This suggests that the Stop cue may not alter SNr activity globally, but rather have a selective influence over cells and subregions involved in controlling the specific movement that needs to be inhibited. We therefore performed a second set of recordings (Experiment 2) using both high-density silicon probes (n = 3 rats, ) and more tetrodes (n = 2 rats) to target a wide range of SNr locations. Combining all SNr results together revealed a clear “hotspot” of SNr cells that fired significantly more on Correct than on Failed Stop trials, briefly after the Stop cue (, S4a–d
). This hotspot corresponds remarkably well to to the SNr sensorimotor “core” subregion described in anatomical studies25
, located dorsolaterally and extended along the rostral-caudal axis. This subregion projects to specific parts of the superior colliculus involved in orienting movements25,26
, so the Stop signal influences activity in an SNr subregion that is likely critical for exerting fast behavioral control27
Figure 4 An SNr hotspot for Stop cue responses. (a) Example of a silicon probe recording from SNr. Tips of the 8 probe shanks were coated in DiO (green) for histological visualization. One tip is visible here (the others were more anterior and posterior). SNr (more ...)
The distinct latencies of STN and SNr cue responses are consistent with Stop information being conveyed along the STN - SNr pathway. Yet, the selectivity of this transmission to Correct Stop trials suggests some form of gating mechanism. In other words: given that the glutamatergic STN cells spike on Failed Stop trials, why are SNr neurons not responsive to this input? The answer may lie in the movement-related firing rate decreases of SNr neurons (Figs. S3 and S4e,f
). Such SNr firing pauses are well-known from studies of eye and limb movements27,28
, and are thought to facilitate action through disinhibition of superior colliculus and other structures more directly linked to motor output16,29
. SNr pauses are driven by increased firing of the GABAergic striatal direct pathway neurons17,30
, plausible participants in a Go process.
To assess how striatal neurons may contribute to movement preparation and initiation, we looked for units that distinguish movement direction before movement onset. We found an abrupt increase in contralateral coding starting ~140ms before movements (arrow in , right). When we compared the activity of these direction-selective striatal cells (n = 74) between different trial types, we observed a rapid acceleration of firing rate just before movement onset31
(, right), that followed the same trajectory for Fast Go, Slow Go and Failed Stop trials. Aligned on the earlier Go cue, this striatal activity remained very similar between Fast Go and Failed Stop trials, but distinct to Slow Go and Correct Stop trials (, left).
Figure 5 Variable timing of a striatal Go process critically determines whether stopping is successful. (a) Fractions of striatal units distinguishing between contra- and ipsilateral movements, at each time point during Go trials. Layout is as . On right, (more ...)
These results fit well with a simple race model, in which variability in the timing of a striatal-based Go process determines the outcome on Stop trials. On Failed Stop trials, movement-related striatal activity has already begun to increase by the time of Stop cue onset (, S6
). This effect was particularly pronounced when examining individual presumed striatal projection neurons (Fig. S6b
). Therefore, the lack of SNr responses to the Stop cue on Failed Stop trials may be due to the early arrival of striatal GABAergic input, shunting away the effects of glutamatergic inputs from STN.
To confirm the viability of this idea we studied gating of the Stop cue responses in a simple integrate-and-fire model of an SNr neuron. This neuron received excitatory pulses, mimicking STN sensory responses to Stop cues, and (as in prior basal ganglia models32,33
) this excitatory input was influenced by GABAergic inhibition34,35
(see Methods). For GABAergic input we used the average striatal population activity during movement initiation () to approximate real input patterns. We adjusted synaptic strengths of inhibitory and excitatory inputs to provide a good qualitative match with the cue-evoked increases and movement-related decreases in SNr firing.
A critical parameter in the model is the relative timing of excitation and inhibition. We define Δ as the interval between Stop cue onset and the point in the striatal output at which movements begin on Go trials. If the Stop cue begins long before movement onset (; Δ = 200ms), striatal inhibition is low at that time and the Stop cue evokes a full response in the SNr cell. By contrast, if the Stop cue occurs only briefly before movement onset (; Δ = 50ms), a high level of striatal inhibition suppresses the SNr cue response. A systematic variation of Δ in the behaviorally relevant range yielded a gating curve that quantified the model response to the Stop cue (). The gating phenomenon required strong divisive inhibition, e.g. through shunting inhibition, rather than simple summation of inhibitory and excitatory conductances (, right). We then used the behavioral data of each rat to estimate the actual distribution of Δ for both Correct and Failed Stop trials (Fig. S7
). These Δ distributions could then be used to calculate model firing rates for these trial types (). Importantly, just as in real rat SNr cells, the model SNr cell selectively responded to the Stop cue in Correct but not Failed Stop trials. We conclude that the integration of distinct excitatory and inhibitory synaptic inputs by individual SNr neurons provides a straightforward, mechanistic account of how Go and Stop processes can “race” within the brain.
Figure 6 Modeling sensorimotor gating in SNr neurons. (a) Model responses for two illustrative values of Δ, the interval between Stop cue and Move onset. Red and green lines indicate STN and striatal (STR) inputs to the SNr model, and blue line shows the (more ...)