Removal of inhibition uncovers latent movement potential during preparation

The motor system prepares for movements well in advance of their execution. In the gaze control system, the dynamics of preparatory neural activity have been well described by stochastic accumulation-to-threshold models. However, it is unclear whether this activity has features indicative of a hidden movement command. We explicitly tested whether preparatory neural activity in premotor neurons of the primate superior colliculus has ‘motor potential’. We removed downstream inhibition on the saccadic system using the trigeminal blink reflex, triggering saccades at earlier-than-normal latencies. Accumulating low-frequency activity was predictive of eye movement dynamics tens of milliseconds in advance of the actual saccade, indicating the presence of a latent movement command. We also show that reaching a fixed threshold level is not a necessary condition for movement initiation. The results bring into question extant models of saccade generation and support the possibility of a concurrent representation for movement preparation and generation.


Significance Statement 48
How the brain plans for upcoming actions before deciding to initiate them is a central question in 49 neuroscience. Popular theories suggest that movement planning and execution occur in serial stages, 50 separated by a decision boundary in neural activity space (e.g., "threshold"), which needs to be crossed 51 before the movement is executed. By removing inhibitory gating on the motor system, we show here that 52 the activity required to initiate a saccade can be flexibly modulated. We also show that evolving activity 53 during movement planning is a hidden motor command. The results have important implications for our 54 understanding of how movements are generated, in addition to providing useful information for decoding 55 movement intention based on planning-related activity. 56

Introduction 71
The ability to interact with the world through movements is a hallmark of the animal kingdom. 72 Movements are usually preceded by a period of planning, when the nervous system makes decisions 73 about the optimal response to a stimulus and programs its execution. Such planning behavior is seen in a 74 wide variety of species, including, insects (Fotowat and Gabbiani, 2007;Card and Dickinson, 2008), fish 75 (Preuss et al., 2006), frogs (Nakagawa and Nishida, 2012), and mammals 76 Churchland et al., 2006a). A fundamental question in sensorimotor neuroscience is how planning activity 77 is appropriately parsed in order to prepare and execute movements. 78 In the primate gaze control system, premotor neurons that produce a volley of spikes to generate a 79 movement are typically also active leading up to the movement. Build-up of low frequency activity prior 80 to the saccade command, or its signatures, have been observed in a wide variety of brain regions involved 81 in gaze control, including the frontal eye fields ; Gold and Shadlen, 2000), lateral 82 intraparietal area (Platt and Glimcher, 1999), and superior colliculus (SC) (Dorris et al., 1997). Since 83 variability in the onset and rate of accumulation of low-frequency activity is correlated with eventual 84 saccade reaction times Ratcliff and Rouder, 1998;Usher and McClelland, 85 2001), it is thought that this activity primarily dictates when the movement is supposed to be initiated. 86 However, it is unclear how (or if) downstream motor networks distinguish activity related to movement 87 preparation from the command to execute one. 88 One possibility is that the activity of premotor neurons undergoes a transformation from representing 89 preparation into movement-related commands at a discrete point in time (Thompson et al., 1996;Juan et 90 al., 2004;Schall et al., 2011), when the activity reaches a movement initiation criterion, thereby acquiring 91 the potential to generate a movement. Indeed, this is the basis for stochastic accumulator models of 92 saccade initiation, in which premotor activity must reach a threshold level in order to generate the 93 movement Ratcliff and Rouder, 1998;Zandbelt et al., 2014). Recent work in the 94 skeletomotor system has also suggested that neuronal population activity undergoes a state space 95 transformation just prior to the movement, thus permitting movement preparation without execution 96 (Churchland et al., 2012;Kaufman et al., 2014;Elsayed et al., 2016). These related views dictate that 97 movement planning and execution are implemented as serial processes in the motor system. Alternatively, 98 it is possible that neurons in sensorimotor structures represent these signals concurrently, gearing up to 99 execute a movement in proportion to the strength of the planning activity. In other words, preparatory 100 build-up of neural activity can multiplex higher order signals while simultaneously relaying those signals 101 to effectors; we call this latter property the "motor potential" of preparatory activity. This idea is in fact 102 the premise of the premotor theory of attention (Rizzolatti et al., 1987;Hoffman and Subramaniam, 103 1995), and represents a latent behavioral manifestation of movement preparation. 104 How might we test for the presence of a motor potential in low frequency preparatory activity? The 105 following thought experiment helps illustrate one approach. Consider the activity of a premotor neuron 106 accumulating over time. Under normal circumstances, inhibitory gating on the saccadic system is released 107 at an internally specified time, possibly when activity crosses a purported threshold level or when the 108 population dynamics reach the optimal subspace, thus resulting in movement generation (top row in 109 Figure 1a). We can then infer that the high frequency burst during movement execution has motor 110 potential if neural activity is correlated with dynamics of the ensuing movement, i.e., saccade velocity is 111 faster when burst activity is higher, and vice versa (match gray traces between top and bottom rows in 112 Figure 1a). Now, if inhibition was somehow removed at a prior time through an experimental 113 manipulation instead of allowing the system its natural time course (thick red line in Figure 1b), the 114 occurrence of an early movement would indicate that ongoing low frequency preparatory activity also 115 possesses motor potential. Importantly, this potential can be quantified by correlating neural activity with 116 kinematics of the eye before the onset of the saccade proper (pre-saccade velocity traces in Figure 1b), 117 and comparing it to the previously estimated potential after saccade onset. Furthermore, the dynamics of 118 activity following the manipulation would indicate whether the activity must cross a decision boundary 119 (i.e., threshold or optimal subspace) in order to produce the movement (dashed traces in Figure 1b). This 120 hypothetical manipulation would therefore simultaneously shed light on both concurrent processing of 121 preparatory signals and the criterion for movement initiation. 122 In this study, we used the trigeminal blink reflex to remove inhibition on the gaze control network during 123 ongoing low-frequency activity. The omnipause neurons (OPNs) in the brainstem, which discharge at a 124 tonic rate during fixation and are suppressed during saccades (Figure 1c, Cohen and Henn, 1972;Keller, 125 1974), also become quiescent during eye movements associated with blinks (Schultz et al., 2010). 126 Previous work in our lab has shown that removal of this potent source of inhibition on the saccade burst 127 generator with reflex blinks triggers saccades at lower-than-normal latencies (Gandhi and Bonadonna, 128 2005), an observation that has been used to study latent sensorimotor processing in SC (Jagadisan and 129 Gandhi, 2016), the motor potential of a target selection signal during visual search (Katnani and Gandhi,130 2013), and the dynamics of movement cancellation during saccade countermanding (Walton and Gandhi, 131 2006). Here, we first established that the saccade-related burst in SC has motor potential under normal 132 conditions, by correlating the activity during the burst to saccade kinematics on individual trials. 133 Critically, when performing the same analysis in the perturbation condition, we found that the level of 134 preparatory activity at the time of the blink was also strongly correlated with initial dynamics of the 135 evoked movement, prior to the saccade proper, suggesting that ongoing sub-threshold activity in SC also 136 possesses motor potential. Finally, we show that although these movements were preceded by an 137 acceleration of ongoing activity following the perturbation, it is not necessary for preparatory activity in 138 SC to reach a threshold before a saccade is producedneural activity just prior to saccades triggered by 139 reflex blinks was lower at both individual neuron and population levels. In order to explicitly test whether saccade preparatory activity contains a latent movement command, we 143 transiently disinhibited the motor system during the preparatory period in monkeys performing the 144 delayed saccade task. Briefly, each subject fixated on a central fixation point while a saccade target 145 appeared in the periphery. After a random delay interval, the fixation spot disappeared, which was the cue 146 (go cue) for the animal to make a saccade to the target. We induced reflex blinks during this preparatory 147 epoch -after the go cue and before saccade typical reaction times. A reflex blink is a suitable perturbation 148 because it removes inhibition on the saccadic system by turning off the OPNs and triggers gaze shifts at 149 lower-than-normal latencies (Gandhi and Bonadonna, 2005). We first describe important aspects of the 150 blink technique here. When induced during fixation in the absence of any other target, a reflex blink is 151 accompanied by a blink-related eye movement (BREM)the eyes turn nasally and downward before 152 returning to the original fixation position in a loop-like trajectory (e.g., Rottach et al., 1998). Gaze shifts 153 triggered by the blink in the presence of a peripheral target thus have the BREM component in addition to 154 the saccade directed towards the target; we refer to them as blink-triggered movements or blink-triggered 155 saccades. 156 Figure 2a shows example velocity profiles of the BREM (thin gray traces in the left column) and the 157 velocity profiles and spatial trajectories of three blink-triggered movements (colored traces in left and 158 right columns, respectively) from one session. We computed onset times of saccades embedded in blink-159 triggered movements using a previously used model-free approach that is agnostic to any possible 160 interaction between the BREM and saccade components (Katnani and Gandhi, 2013). Saccade onset was 161 determined as the time at which the movement velocity on a given trial deviated (colored circles) from the 162 expected BREM profile distribution (thick black traces in Figure 2a; only the mean BREM trace is shown 163 in the right panel for clarity) for that session. As seen from the example blink-triggered movements, there 164 was considerable variation in the time at which the eye movement deviated from the BREM profile 165 towards the saccade goal, marking saccade onset. Figure 2b shows the distribution of saccade onset times 166 relative to blink time obtained using this approach (for more details, see Methodsmovement detection). 167 It is worth noting here that the bimodality apparent in the distribution of saccade onset times in Figure 2b  168 likely reflects the divide between trials in which the process behind saccade initiation was already 169 underway by the time of the blink (delays <20 ms, to the left of the vertical black line), and trials in which 170 the saccade was triggered due to disinhibition by the blink (delays >20 ms, to the right of the vertical 171 black line). Since most analyses in this study were focused on processes occurring before saccade onset, 172 we split the data from perturbation trials into two sets, along the aforementioned divide: blink-triggered 173 movements with saccade onset greater than 20 ms after blink onset were used for the motor potential and 174 accumulation rate analyses, and all blink-triggered movements were used for the threshold analysis (for 175 more details, see Methodsinclusion criteria). 176

177
The blink perturbation triggers reduced latency, but accurate, saccades 178 First, as one way to assay the behavioral manifestation of motor preparation, we verified that reflex blinks 179 during the preparatory period produced low-latency saccades. Figure 3a shows saccade reaction time 180 (from GO cue) as a function of the time of blink across all perturbation trials (red circles). To visually 181 compare reaction times on blink trials with those in control trials, it was necessary to include the 182 distribution of control reaction times in this figure. To do this, we created a surrogate dataset by randomly 183 assigning blink times to control trials, and plotted them on the same axes as blink trials in Figure 3a (blue 184 circles). Reaction times in perturbation trials were correlated with time of blink, and were significantly 185 lower than control reaction times (mean control reaction time = 278 ms, mean blink-triggered reaction 186 time = 227 ms, p = 2.2 x 10 -197 , one-tailed t-test), consistent with previous observations (Gandhi and 187 Bonadonna, 2005). 188 We then verified whether saccades triggered by the blink were as accurate as normal saccades, in order to 189 eliminate any potential confounds due to differences in accuracy. We calculated saccade accuracy as the 190 Euclidean endpoint error normalized with respect to target location (inset in Figure 3b). The distributions 191 of relative errors for all control and blink trials are shown in Figure 3b. Blink-triggered saccade accuracy 192 was not significantly different from control saccades (mean control accuracy = 0.136, mean blink-193 triggered accuracy = 0.133, p = 0.3, two-tailed t-test), as reported previously (Goossens and Van Opstal,194 2000b; Gandhi and Bonadonna, 2005). We also tested for and found no relationship between blink time 195 and saccade accuracy (Spearman's correlation = 0.04, p = 0.22). Note that the eyes are closed due to the 196 blink, so visual feedback does not contribute to endpoint accuracy. Nonetheless, we have demonstrated 197 previously that blanking the visual target during the blink-triggered movement does not compromise the 198 accuracy on perturbation trials (Gandhi and Bonadonna, 2005). The blink perturbation thus provides an 199 assay to study the question of motor potential without introducing confounding factors related to saccade Next, as a crucial prerequisite for our motor potential analysis, we examined how the motor burst in SC is 204 correlated with saccade kinematics. Specifically, we computed the correlation between the trial-by-trial 205 firing rates of a neuron and the corresponding velocities. This approach to estimating motor potential is 206 illustrated in Figure 4a. Since our eventual goal was to study pre-saccade motor potential in blink-207 triggered saccades, we used the component of velocity in the direction of the saccade goal as our 208 kinematic variable on a given trial (inset in Figure 4a, see Methodskinematic variables, for details). The 209 choice to use projected kinematics is to maintain uniformity with analyses on blink-triggered movements 210 (see next section), but performing the analysis on the raw, unprojected kinematics for control saccades 211 yielded very similar results (Supplementary Figure 1). Further, to avoid assumptions about the efferent 212 delay between SC activity and ocular kinematics, we computed the activity-velocity correlation at various 213 potential delays between the two signals (the blue bars in Figure 4a show an example delay of 12 ms -214 compare the similarly shaded bars in the velocity and activity panels). Figure 4b shows, for an example 215 neuron, the trial-by-trial scatter of velocities at three time points (15 ms before, at, and 15 ms after 216 saccade onsetlight, medium, and dark blue circles, respectively) plotted against neural activity 217 preceding those respective velocities by 12 ms. 218 It is not surprising to see a lack of correlation with SC activity for pre-saccade velocities, since they are 219 largely zero or constant, by definition, for normal saccades, because the inhibitory gating by OPNs has 220 not been removed yet. In contrast, the strong correlation between kinematics and activity following 221 movement onset indicates the presence of a motor potential in the saccade-related burst. We 222 systematically explored the time course of this motor potential by computing the correlation at different 223 time points before and during the saccade, for a range of delays between activity and kinematics. We did 224 this for each neuron, and the population average correlation coefficients at each time point and delay are 225 shown in the heat map in Figure 4c. To aid interpretation of this figure, the light blue asterisk in the heat 226 map refers to the correlation computed at the time points with the corresponding asterisk in Figure 4a. 227 Motor potential of SC activity emerged only after the onset of the saccade, and lasted throughout the 228 movement (streak of correlation below the unity line in Figure 4c). 229 For each time point in the velocity signal, we also computed the activity time points at which correlation 230 was highest, shown as the running mean (black trace) in the heat map. This provides a measure of the 231 efferent delay at which neural activity is most effective in driving movement kinematics. Note that the 232 black trace is roughly parallel to the unity line, suggesting that the efferent delay was consistent for the 233 duration of the movement. To characterize this property better, we plotted optimal efferent delay, 234 calculated as the difference between the black trace and the unity line, as a function time with respect to 235 movement onset (Figure 4d). The mean delay during the movement was -12 ms, meaning that 236 instantaneous saccade kinematics were best predicted by SC activity 12 ms before. This value is centered 237 within the range of previous functional estimates of the conduction delay between SC and extra-ocular 238 muscles (8-17 ms, Miyashita and Hikosaka, 1996). The values for the delay before saccade onset are 239 highly variable, likely due to noise in the pre-saccade velocities, and therefore should be ignored (also 240 note that these delays are positive and therefore non-causal). Finally, Figure 4e shows the correlation 241 values at the -12 ms delay as a function of time. The gray region is the ± 95% confidence intervals of the 242 population average bootstrapped (trial-shuffled) distribution of correlations (see Methodssurrogate data 243 and statistics). Thus, for normal trials, motor potential, in the form of correlation between neural activity 244 and eye velocity, only manifests after the onset of the saccade proper (starting 3 ms after saccade onset, p 245 < 0.05, bootstrap test). 246 247 SC preparatory activity preceding the saccade-related burst possesses motor potential 248 Having found a correlation between SC activity and ocular kinematics during saccades, we wanted to 249 know whether the time course of such motor potential expands when the saccadic system is disinhibited at 250 earlier time points. Specifically, we wanted to know whether ongoing preparatory activity contained any 251 motor potential before its maturation into a motor command. We hypothesized that, if the low-frequency 252 preparatory activity that precedes the high-frequency burst had motor potential, removing inhibitory 253 gating on the system would result in a slow eye movement proportional to the level of activity before 254 accelerating into a saccade. The rationale was that, because the downstream OPNs become quiescent 255 during the blink, activity in SC neurons that possessed motor potential would be allowed to drive the 256 burst generators, and subsequently, the eye muscles (see circuit diagram in Figure 1c). To test this, we 257 computed the correlation between SC activity and eye velocity before and during blink-triggered saccades 258 in a manner like that for control saccades (Figure 5a). It is important to note that we subtracted the BREM 259 component from the blink-triggered saccade velocities before projecting these residuals in the direction of 260 the saccade goal (inset in Figure 5a). This was done to prevent independent variations in BREM 261 kinematics (unrelated to SC activity) from masking any underlying motor potential-related correlation, 262 which we found might be the case when we performed this analysis on the raw velocities for blink-263 triggered saccades ( Supplementary Figures 2a-b). Figure 5b shows an example scatter plot of the trial-by-264 trial activities versus projected residual velocities for three time points with respect to saccade onset. 265 We then computed the population average correlation at different time points before and after the onset of 266 the high velocity saccade, at different efferent delays. Neural activity was highly correlated with 267 movement kinematics after the onset of the high velocity saccade component (after time 0 in the heat map 268 in Figure 5c), in agreement with data from control saccades. Importantly, activity was also correlated with 269 eye kinematics before saccade onset (time points before 0 in Figure 5c), suggesting that upstream SC 270 activity leaked through to the eye muscles as soon as the OPNs were turned off by the blink, causing 271 activity-related deviations in the kinematics around the BREM. The black trace in Figure 5c also shows 272 that the estimated efferent delay was similar to that observed in control trials and consistent before and 273 after saccade onset. This is better observed in Figure 5d   If the observed pre-saccade correlation resulted solely from estimating saccade onset to be later than the 291 ground truth, i.e., it is actually a peri-saccade correlation in disguise, then it should persist even with the 292 raw, unprojected blink-triggered velocity residuals, and at the same efferent delay as for the peri-saccade 293 correlation. This is because we have already seen that motor potentials exist once the saccade has started. 294 Second, it adds support to the notion of motor potential itself: if spikes from the preparatory activity of 295 these neurons leaked through to the muscles, you would expect them to only drive kinematics in the 296 neurons' preferred direction (as opposed to a non-selective impact on all movements). Note that the time 297

Blink-triggered saccades are evoked at lower thresholds compared to normal saccades 304
Next, we used the fact that blink-triggered saccades are evoked at lower latencies to test an influential 305 model of saccade initiationthe threshold hypothesis . We asked whether it is 306 necessary for activity in SC intermediate layer neurons to reach a fixed activity criterion in order to 307 generate a movement. Previous studies have estimated the threshold for individual neurons in SC and 308 FEF by assuming a specific time at which the threshold could be reached before saccade onset or by 309 computing the time, backwards from saccade onset, at which premotor activity starts becoming correlated 310 with reaction time Paré and Hanes, 2003). Given the heterogeneity of the 311 activity profiles of premotor neurons, we think this approach is too restrictive to obtain an unbiased 312 estimate of the threshold, if any. Instead, we took a non-parametric approach and scanned through 313 possible times at which a putative threshold might be reached prior to saccade onset (Jantz et al., 2013). 314 Figure 6a shows a snippet of the average population activity, normalized by the peak trial-averaged 315 activity during control trials for each neuron, and aligned on saccade onset for control (blue traces) and 316 blink (red traces) trials. For each neuron in this population (n = 50), we computed the average activity in 317 10 ms bins slid in 10 ms increments from 50 ms before to 20 ms after saccade onset (colored windows at 318 the bottom of Figure 6a). If activity on control trials reaches the purported threshold at any one of these 319 times before saccade onset, a comparison with activity in blink trials at that time should reveal the 320 existence, or lack thereof, of a fixed threshold. Figure 6b shows the activity in these bins for control trials 321 plotted against blink trials, colored according to the bins in Figure 6a. Note that the majority of points for 322 early time bins lie below the unity line. Activity on blink trials was significantly lower compared to 323 control activity from 50 ms before to 10 ms after saccade onset (square points, Wilcoxon signed-rank test, 324 comparisons in each of these windows were significant at p = 10 -6 , at least). The systematic trend in the 325 linear fits (solid lines) to these points suggests that the activity on blink trials gradually approaches that on 326 control trials; however, the earliest time at which control activity was not different from activity in blink 327 trials was 20 ms after saccade onset (circles, Wilcoxon signed-rank test, p = 0.8)too late to be 328 considered activity pertaining to a movement initiation threshold. Thus, activity at the population level 329 need not reach a threshold level in order to produce a movement. 330 Nevertheless, we wanted to know if there exist individual neurons in the population that might obey the 331 threshold hypothesis. For each neuron, we calculated whether activity on blink trials was higher, lower, or 332 not significantly different from activity in control trials, at each time point from -50 before to 20 ms after 333 saccade onset (Wilcoxon rank-sum test, comparisons in each of these windows were significant at p = 334 0.001, at least). The three traces in Figure 6c represent the proportion of neurons that showed each of 335 those three characteristics as a function of time. As late as 10 ms before saccade onset, more than 60% of 336 the neurons had lower activity on blink trials compared to control trials (blue trace), inconsistent with the 337 idea of a fixed threshold. Roughly 30% of the neurons did not exhibit significant differences in activity on 338 blink and control trials at that time point (black trace); however, this observation is insufficient to 339 conclude that the activities in the two conditions were identical, or that it must reach a threshold. Of 340 course, it is possible that some of these neurons belong to a class for which fixed thresholds have been 341 observed in previous studies. Together, these results suggest that it is not necessary for premotor activity 342 in SC intermediate layers to reach a threshold level at the individual neuron or population level in order to 343 produce a movement. 344 345

Rate of accumulation of SC activity accelerates following disinhibition by the blink 346
Since SC neurons do not necessarily cross a fixed threshold to produce movements, as we saw above, it is 347 possible that blink-triggered saccades are initiated directly off the ongoing level of preparatory activity. 348 Alternatively, low frequency SC activity may be altered by the blink, even if the saccade is triggered at a 349 lower level compared to control trials. Therefore, we studied whether the dynamics of SC activity are 350 modulated by the blink prior to saccade initiation. Since we wanted to test for a change in dynamics 351 before the actual saccade started, we restricted our analysis to the subset of trials in which saccade onset 352 occurred at least 20 ms after blink onset. This restriction reduced our population to 38 neurons. For each 353 neuron, we estimated the rate of accumulation of activity in 20 ms windows before and after blink onset 354 with piecewise linear fits (Figure 7a, dashed red trace and solid lines). It is important to note that while 355 the evolution of premotor activity is commonly modelled as a linear process, the actual dynamics of 356 accumulation may be non-linear, causing spurious changes in linear estimates of accumulation rate over 357 time. To account for this, we created a surrogate dataset of control trials for each neuron, with blink times 358 randomly assigned from the actual distribution of blink times for that session. We then performed linear 359 accumulation fits for the control dataset as well (dashed blue trace and solid lines in Figure 7a). Changes 360 in accumulation rate on control trials following the pseudo-blink should reflect the natural evolution of 361 activity at typical blink times and provide a baseline for comparing any changes observed in blink trials. 362 Figure 7b shows a scatter plot of pre-and post-blink accumulation rates on control and blink trials. Pre-363 blink rates were not different between the two conditions (light circles, Wilcoxon signed-rank test, p = 364 0.4), but post-blink rates were significantly higher on blink trials (dark circles, Wilcoxon signed-rank test, 365 p = 2.5 x 10 -6 ). Next, we tested for a change in accumulation rates following the blink by calculating a 366 rate change index, defined as the difference of post-versus pre-blink rates divided by their sum, for each 367 condition (Figure 7c). This index was positive for most neurons, even for control trials, highlighting the 368 natural non-linear dynamics mentioned above. The change in accumulation rate was significantly higher 369 following the actual blink on blink trials (Wilcoxon signed-rank test, p = 4.4 x 10 -6 ) compared to after the 370 pseudo-blink on control trials. Thus, removal of inhibition seems to cause an acceleration in the dynamics 371 of ongoing activity in the lead up to a saccade. 372 373

Discussion 374
In this study, we sought to uncover the dynamics of movement preparation and, specifically, test whether 375 the low-frequency preparatory activity of SC neurons encodes movement-related signals. We first 376 established a baseline for this question by showing that, under normal conditions, saccade-related activity 377 in the intermediate layers of SC has "motor potential", defined as correlated variability between firing rate 378 and saccade kinematics. Then, by disinhibiting the saccadic system much earlier than its natural time 379 course with a reflex blink, we showed that low-frequency preparatory activity in these neurons also has a 380 latent motor potential, indicating the presence of a hidden movement command. These results suggest that 381 the output of higher order gaze control regions like the SC possesses motor potential during both low-and 382 high-frequency activity, but a correlation with movement kinematics may only present itself when 383 intermediary gating (in this case presented downstream of SC by the OPNs) between the two observables 384 is turned off. We also found that SC activity does not necessarily have to reach a threshold at the single 385 neuron or population level in order to initiate the saccade, contrary to the postulates of an influential 386 model of saccade initiationthe threshold hypothesis . 387 388

Relationship between movement preparation and execution 389
Studies on the neural correlates of movement generation have largely focused on the divide between 390 preparatory and executory activity, mainly due to the substantial natural latencies between the cue to 391 perform a movement and its actual execution. Since variability in the properties of post-cue neural 392 activity is correlated with eventual movement reaction times Churchland et al., 393 2006b), it is thought that this activity is purely preparatory in nature, influencing only when the 394 movement is supposed to be initiated , and is devoid of the potential to generate a 395 movement, until just before its execution. Indeed, there is some evidence that movement preparation and How do we reconcile such previous observations with the finding in this study of a latent motor potential 407 in preparatory activity? One possibility is that the oculomotor system operates differently from the 408 skeletomotor system. More generally, it is possible that it is necessary for population activity to be in an 409 optimal, "movement-generating" subspace in order to release inhibition and/or effectively engage 410 downstream pathways leading up to activation of the muscles, but once the motor system has been 411 disinhibited by another means (e.g., the blink perturbation in this study), preparatory activity is read out as 412 if it were a movement command. Based on the results in our study, we hypothesize that the activity is 413 likely in the movement preparation subspace when the slow eye movement is produced after the blink 414 perturbation, and its subsequent transition into the movement execution subspace results in a high 415 velocity movement. More studies that causally delink evolving population activity from physiological 416 gating are needed to clarify these mechanisms. 417 418

Implications for threshold-based accumulator models 419
The threshold hypothesis is the leading model of movement preparation and initiation in the gaze control 420 system . Inspired by stochastic accumulator models of decision-making 421 Our results show that the low-frequency preparatory activity of individual or population of SC neurons 438 does not have to transition into a high-frequency mode to trigger a movement (Figure 6a, b). If gates 439 downstream of the SC are removed, then reduced SC activity is sufficient to move the eyes. The effective 440 reduction in threshold (or equivalently, non-existence of a fixed threshold) that we observe is likely due to 441 reduced activity of the OPNs, which are a potent source of inhibition on the pathway downstream of the 442 SC. The OPNs are inhibited by the reflex blink, and thus premotor activity needs to overcome lesser 443 inhibition and is able to trigger movements off a lower level. It is also important to note that while there is 444 some evidence that premotor activity in SC is attenuated when saccades are perturbed by a reflex blink 445 (Goossens and Van Opstal, 2000a), we did not observe suppression during movements that were triggered 446 by the blink, as seen in the firing rate profile in Figure 7a we observe, either before or during the saccade, since we estimate this potential from correlated 488 variability between SC activity and eye kinematics across trials, not within a trial. Moreover, in Figure 5, 489 this correlation is computed between activity and residual velocity projected onto the direction of the 490 saccade target after subtraction of the BREM template. This can cause the kinematic variable to be 491 instantaneously negative (i.e., going away from the saccade target) on some trials, but as long as it is less 492 negative on trials when the activity is higher, we can say that the activity has motor potential. In contrast, 493 the mini-vector model will predict that the eye moves in fixed vector increments towards the saccade 494 target (which happens roughly to be the optimal vector of the recorded neuron). Moreover, in the previous 495 study, not all neurons show the fixed spike count property, and at the population level, spike counts on 496 perturbation trials are slightly higher than on control trials. This is entirely consistent with our observation 497 Cullen, 2006) -burst neuron activity is also modified to account for the observed changes in eye velocity. 504 Given these results, we predict that the lower brainstem burst generator neurons will exhibit low 505 frequency activity to produce the slow eye movement leaked by premature inhibition of OPNs, followed 506 by a high frequency burst that generates the saccade. 507 We would like to emphasize that the observed results -preparatory motor potential, reduced threshold, 508 accelerated activity dynamics -are most likely indirect effects of the trigeminal blink reflex, via inhibition 509 of the OPNs, and not directly due to the reflex itself. Prior work has shown that the activity of SC neurons 510 is not affected by the BREM produced during fixation (Goossens and Van Opstal, 2000a; Jagadisan and 511 Gandhi, 2016). For blinks produced after the saccade target is presented, some SC neurons in fact exhibit 512 attenuation (Goossens and Van Opstal, 2000a), although we did not see it in our dataset (and even if that 513 did happen, it is counter-intuitive to and does not explain the motor potential and acceleration of activity). 514 These observations collectively suggest that the acceleration of activity that leads to a reduced latency 515 saccade is not directly due to the trigeminal blink reflex but indirectly due to OPN inhibition. 516 In our study, we found the correlation between SC activity and eye kinematics during the saccade to be 517 maximal at a time shift of 12 ms between the two signals, providing us with an estimate of the optimal 518 efferent delay between SC and extraocular muscles. This is in line with previously estimated ranges for 519 the efferent delay (8-17 ms, Miyashita and Hikosaka, 1996). In that study, single pulse microstimulation 520 was delivered to the SC during an ongoing saccade, and the latency to deviation from normal saccade 521 kinematics provided an estimate of the time it takes for a spike in SC to reach extraocular muscles while 522 the gating in the pathway downstream of SC is open. The fact that we observe the same efferent delay 523 (i.e., 12 ms) prior to saccade onset, when the OPNs are already quiescent due to the reflex blink, fits 524 neatly within this picture and adds credibility to the notion of a latent motor potential in preparatory 525 spikes. 526 527

Parallel implementation of the sensory-to-motor transformation 528
An influential idea in systems neuroscience is the premotor theory of attention, which posits that spatial 529 attention is manifested by the same neural circuitry that produces movements (Rizzolatti et al., 1987).

Concluding remarks 558
It is worthwhile to end on a note of caution. The results in this study are based on experiments performed 559 in one node, SC, in a distributed network of brain regions involved in gaze control. Traditional knowledge 560 imposes a hierarchy on the sensorimotor transformations that need to occur before a gaze shift is 561 generated . It is possible that sensorimotor neurons in SC, and to some extent, FEF, 562 which project directly to the brainstem burst generator (Segraves, 1992;Rodgers et al., 2006), are more 563 likely to exhibit signatures of a motor potential in preparatory activity compared to regions higher in the 564 cascade. Neurons in other regions may still need to signal the initiation of a movement by reaching a 565 threshold, optimal subspace, or other similar decision bound. Furthermore, it is known that movement 566 initiation thresholds in SC and FEF can vary based on task context (e.g., Jantz et al., 2013). The results 567 presented here are based on a relatively simple taskthe delayed saccade task. The mechanisms of 568 movement initiation, and the presence of motor potential in preparatory activity, could in principle be 569 different in more complex tasks, e.g. those that involve competitive spatial selection of movements or 570 sequential movements. Future studies that take causal approaches to perturbing intrinsic population 571 dynamics in various premotor areas across different tasks and effector modalities are essential in order to 572 gauge whether the findings in this study point to a fundamental and generalizable property of 573 sensorimotor systems.  (Bryant and Gandhi, 2005). After initial training and acclimatization, the monkeys were trained to 600 perform a delayed saccade task. The subject was required to initiate the trial by looking at a central 601 fixation target. Next, a target appeared in the periphery but the fixation point remained illuminated for a 602 variable 500-1200 ms, and the animal was required to delay saccade onset until the fixation point was 603 extinguished (GO cue). Trials in which fixation was broken before peripheral target onset were removed 604 from further analyses. The animals performed the task correctly on >95% of the trials. During each 605 session, we presented the targets in one of two locationseither inside the response field of the recorded 606 neuron, or in the diametrically opposite location (see below). 607

Induction of reflex blinks 609
On 15-20% of trials, fixation was perturbed by delivering an air puff to the animal's eye to invoke the 610 trigeminal blink reflex. Compressed air was fed through a pressure valve and air flow was monitored with 611 a flow meter. To record blinks, we taped a small Teflon-coated stainless steel coil (similar to the ones 612 used for eye tracking, but smaller in coil diameter) to the top of the eyelid. The air pressure was titrated 613 during each session to evoke a single blink. Trials in which the animal blinked excessively or did not 614 blink were excluded from further analyses. The air-puff delivery was randomly timed to evoke a blink 615 either during fixation (400-100 ms before target onset) or 100-250 ms after the GO cue, during the early 616 phase of the typical saccade reaction time. The blink evoked during fixation produced a slow and loopy 617 blink-related eye movement (BREM, gray traces in Figure 2a). The eyes returned to near the original 618 position and fixation was re-established for 400-100 ms before a target was presented in the periphery and 619 the remainder of the delayed saccade task continued. The window constraints for gaze were relaxed for a 620 period of 200-500 ms following delivery of the air puff to ensure that the excursion of the BREM did not 621 lead to an aborted trial. The blink evoked after the GO cue typically produced a blink-triggered 622 movement that can be described as a combination of a BREM and a saccade to the desired target location 623 (colored traces in Figure 2a). We used deviations from the BREM profile to determine true saccade onset, 624 as described in more detail in the next section. 625 626

Movement detection 627
Data were analyzed using a combination of in-house software and Matlab. Eye position signals were 628 smoothed with a phase-neutral filter and differentiated to obtain velocity traces. Normal saccades, 629 BREMs, and blink-triggered gaze shifts were detected using standard onset and offset velocity criteria (50 630 deg/s and 30 deg/s, respectively). Onsets and offsets were detected separately for horizontal and vertical 631 components of the movements and the minimum (maximum) of the two values was taken to be the actual 632 onset (offset). Saccade onset times within blink-triggered movements were detected using a non-633 parametric approach (Katnani and Gandhi, 2013, also see Figure 2a). We first created an estimate of the 634 expected BREM distribution during each session by computing the instantaneous mean and standard 635 deviation of the horizontal and vertical BREM velocity profiles. Then, for each blink-triggered movement 636 in that session, we determined the time point at which the velocity exceeded the ± 2.5 s.d. bounds of the 637 BREM profile distribution, and remained outside those bounds for at least 15 successive time points. We 638 did this separately for the horizontal and vertical velocity profiles, and took the earlier time point between 639 the two components as the onset of the saccade. We further manually verified that the detected saccadic 640 deviations on individual trials were reasonable, esp., in the spatial domain. Figure 2a  During each recording session, a tungsten microelectrode was lowered into the SC chamber using a 645 hydraulic microdrive. Neural activity was amplified and band-pass filtered between 200 Hz and 5 kHz 646 and fed to a digital oscilloscope for visualization and spike discrimination. A window discriminator was 647 used to threshold and trigger spikes online, and the corresponding spike times were recorded. The 648 location of the electrode in the intermediate layers of SC was confirmed by the presence of visual and 649 movement-related activity as well as the ability to evoke fixed vector saccadic eye movements at low 650 stimulation currents (20-40 µA, 400 Hz, 100 ms). Before beginning data collection for a given neuron, its 651 response field was roughly estimated. During data collection, the saccade target was placed either in the 652 neuron's response field or at the diametrically opposite location (reflected across both axes) in a randomly 653 interleaved manner. Response field centers, and therefore, target locations (also consequently, saccade 654 amplitudes and directions) varied between 9-25 degrees in eccentricity and spanned all directions. 655

Data analysisneural pre-processing 657
Spike density waveforms were computed for each neuron and each trial by convolving the raw spike 658 trains with a Gaussian kernel. We used a 3 ms wide kernel for the motor potential and threshold analysis 659 (involving across-trial correlations or trial-averaged neural activity) and a 10 ms kernel for the 660 accumulation rate analysis (for better rate estimation on individual trials). For a given neuron and target 661 location, spike densities were averaged across trials after aligning to target and saccade onsets. Neurons 662 were classified as visual, visuomovement or movement-related, based on the presence of a significant 663 target and/or saccade-related response. We only analyzed visuomovement and movement neurons for this 664 study, the majority of which were visuomovement (47/50). Where necessary, we normalized the trial-665 averaged spike density of each neuron to enable meaningful averaging across the population. The activity 666 of each neuron was normalized by its peak trial-averaged firing rate during normal saccades. 667 668

Data analysisinclusion criteria 669
Overall, we recorded from 64 neurons for 12339 control trials and 2364 blink trials over 50 sessions. For 670 all analyses, we only considered neurons for which we had at least 10 blink perturbation trials with the 671 target in the response field. Since we only introduced the blink perturbation on a small percentage of trials 672 in a given session in order to prevent habituation, this restricted our population to 50 neurons (7891 673 control trials and 1615 blink trials over 43 sessions). We used all 50 neurons for the threshold analysis 674 ( Figure 6). For the motor potential analysis (Figure 5), since our aim was to correlate neural activity with 675 eye kinematics before saccade onset, we used only the subset of trials where the onset of the saccade was 676 delayed with respect to overall movement onset by at least 20 ms (see Figure 2b). To ensure that the 677 correlation values were reliable, we used only neurons which had at least 10 trials meeting the above 678 criterion. This restriction reduced the number of neurons available for the motor potential analyses to 38 679 (6771 control trials and 869 blink trials over 32 sessions), and we used the same neurons for control trials 680 to enable meaningful comparison (Figure 4). For the same reason, we also used this subset of neurons for 681 the accumulation rate analysis (Figure 7), where we compared the dynamics of neural activity in 20 ms 682 windows before and after blink onset, and we wanted to ensure that the post-blink window did not include 683 activity co-occurring with the saccade. 684 685

Kinematic variables 686
For the motor potential analyses in Figures 4 and 5, we computed the across-trial correlation between 687 instantaneous movement kinematics and neural activity for each neuron. We computed the kinematic 688 variable of relevance for each analysis as follows. In all cases, we used the raw or modified (see below) 689 horizontal and vertical velocity signals to compute a single vectorial velocity signal using the Pythagorean 690 theorem: ( ) = √ ℎ 2 ( ) + 2 ( ). For the analyses in Supplementary Figure 1, for control trials, we used 691 the raw, unmodified velocity signals to compute vectorial velocity as a function of time, which we then 692 used as the instantaneous kinematic variable to correlate with neural activity. For perturbation trials in 693 and spatial profiles of blink-triggered movements look largely like those of a BREM (see Figure 2a). 696 Thus, in order to extract only the saccadic component of a blink-triggered movement, we subtracted from 697 it the mean BREM template on a given session, and used only the horizontal and vertical residuals to 698 compute the vectorial residual velocity: ̃( ) = √̃ℎ 2 ( ) +̃2( ) , which was used for the correlation 699 analysis in Supplementary Figure 2. 700 A potential pitfall when using residual velocities by just subtracting out the mean BREM, given the 701 variability in BREM profiles across repetitions, is that intrinsic variability of the BREM itself may mask 702 any correlated variability that might be present between ocular kinematics and neural activity. In other 703 words, if the BREM is driven by an independent pathway compared to the saccade/SC activity, it 704 represents an orthogonal source of variability in the kinematics relative to the activity-driven variability 705 that is being examined. Therefore, for the perturbation trial analysis in Figure 5, we used the component 706 of residual velocity in the direction of the saccade goal, to isolate variability in the direction of the 707 saccade. The kinematic variable for this analysis is thus defined as: ̃( ) = √̃ℎ 2 ( ) +̃2( ) cos , 708 where is the angle between the instantaneous residual velocity vector and the direction of the saccade 709 goal (e.g., between the green and black vectors in the inset in Figure 5a). For the sake of consistency, we 710 used a similar variable: ( ) = √ ℎ 2 ( ) + 2 ( ) cos , for the equivalent control analysis in Figure 4, 711 even though the instantaneous direction of velocity is largely towards saccade goal in this condition. indicates timepoints at which the average correlation was significant (based on ± 95% CI from panel e). 851 d. Optimal efferent delay computed as the distance of the black trace in panel c from the unity line.

852
Negative values for the delay are causal, i.e., correlation was high for activity points leading the velocity 853 points. The shaded gray bar shows that the optimal delay was consistent during the movement (mean for 854 shaded region = -12 ms) e. Population average correlation as a function of time at the -12 ms estimated 855 efferent delay. The black trace is the mean and the gray region is the ± 95% confidence interval for the 856 bootstrapped (trial-shuffled) correlation distribution. causal, i.e., correlation was high for activity points leading the velocity points. The gray bar highlights the 891 fact that the optimal delay was consistent during both control and blink-triggered saccades, and both 892 before and after saccade onset for the latter (mean for shaded region = -12 ms). e. Population average 893 correlation for blink-triggered movements (red trace) as a function of time at the -12 ms estimated 894 efferent delay. The blue trace is from Figure 4e for control saccades, overlaid for comparison. The black 895 trace is the mean and the gray region is the ± 95% confidence interval for the bootstrapped (trial-896 shuffled) correlation distribution. As in Figure 4a, horizontal and vertical velocity traces (top two rows) on control trials are converted to 944 radial velocity (third row). In contrast to Figure 4a, the velocities are used as is to compute motor 945 potential, without projecting onto the direction of the saccade goal. The bottom row shows neural activity 946 traces on different trials for the neuron recorded in this example session (same as Figure 4a). b. Motor 947 potential is estimated as the correlation between neural activity and saccade kinematics in appropriate 948 time windows. The scatter plot of the projected radial velocity 15 ms after saccade onset, at saccade onset, 949 and 15 ms before saccade onset (light, medium, and dark blue points, respectively, and corresponding 950 vertical lines in panel a) against neural activity 12 ms preceding the velocity time points. Each point 951 corresponds to one trial. c. Point-by-point correlation of velocity and activity, averaged across neurons. 952 The black curve traces the contour of the highest correlation time points in the activity for each point 953 during the movement. The blue bar at the bottom of the heatmap indicates timepoints at which the 954 average correlation was significant (based on ± 95% CI from panel e). d. Optimal efferent delay (thick 955 blue trace) computed as the distance of the black trace in panel c from the unity line. Negative values for 956 the delay are causal, i.e., correlation was high for activity points leading the velocity points. The gray bar 957 shows that the optimal delay was consistent during the movement (-12 ms) e. Population average 958 correlation as a function of time at the -12 ms estimated efferent delay. The black trace is the mean and 959 the gray region is the ± 95% confidence interval for the bootstrapped (trial-shuffled) correlation converted to radial velocity (third row). In contrast to Figure 5a, the residual velocities (gray shaded 971 deviation from the BREM template) are used as is to compute motor potential, without projecting onto 972 the direction of the saccade goal. The bottom row shows neural activity traces on different trials for the 973 neuron recorded in this example session (same as Figure 5a). b. Scatter plot of the neural activity versus 974 velocity at the three time points (shaded red windows) from panel a for blink-triggered movements. c. 975 Point-by-point correlation of velocity and activity, averaged across neurons, for blink-triggered 976 movements. The velocity time points are with respect to time of saccade onset extracted from the blink-977