Extraction of saccades from eye movements triggered by reflex blinks

The trigeminal blink reflex can be evoked by delivering an air puff to the eye. If timed appropriately, e.g., during motor preparation, the small, loopy blink-related eye movement (BREM) associated with eyelid closure disinhibits the saccadic system and reduces the reaction time of planned eye movements. The BREM and intended eye movement overlap temporally, thus a mathematical formulation is required to objectively extract saccade features – onset time and velocity profile – from the combined movement. While it has been assumed that the interactions are nonlinear, we show that blink-triggered movements can be modeled as a linear combination of a typical BREM and a normal saccade, crucially, with an imposed delay between the two components. Saccades reconstructed with this approach are largely similar to control movements in their temporal and spatial profiles. Furthermore, activity profiles of saccade-related bursts in superior colliculus neurons for the recovered saccades closely match those for normal saccades. Thus, blink perturbations, if properly accounted for, offer a non-invasive tool to probe the behavioral and neural signatures of sensory-to-motor transformations. New and noteworthy The trigeminal blink reflex is a brief noninvasive perturbation that disinhibits the saccadic system and provides a behavioral readout of the latent motor preparation process. The saccade, however, is combined with a loopy blink related eye movement. Here, we provide a mathematical formulation to extract the saccade from the combined movement. Thus, blink perturbations, when properly accounted for, offer a non-invasive tool to probe the behavioral and neural signatures of sensory-to-motor transformations.


Introduction 41
As a general rule, an experimental result has a deeper impact when its correlational outcome is fortified 42 by causal evidence. In systems neuroscience, this latter objective is typically addressed by observing 43 how a normal circuit or system responds to perturbations. Apart from supplementing correlational 44 studies, perturbations can also impose new regimes on a system and reveal novel properties that are 45 hidden from observation under normal circumstances. Standard forms of perturbations in nonhuman 46 primate research are electrical microstimulation or chemical injection to excite or suppress activity in a 47 region of neural tissue. While informative, their major disadvantage (at the moment) is the inability to 48 record neural activity at the same site of experimental manipulation. Furthermore, microstimulation 49 induces extraneous activity that could confound interpretation, and neurons take a relatively long time 50 to recover from chemical inactivation which limits their reversibility and application in randomized 51 paradigms. A newer perturbation tool, optogenetics, circumvents some of these limitations, but it has 52 yet to deliver efficient outcomes in higher order mammals. 53 Our vast knowledge of the neural basis of motor control, sensorimotor integration and cognition has 54 emerged from experimental manipulations applied to the saccadic system. For example, saccades 55 interrupted by microstimulation of the brainstem omnipause neurons (OPNs) have been crucial in 56 demonstrating an internal feedback control mechanism that preserves accuracy (Keller et al. 1996), and 57 bias in target selection after inactivation of the superior colliculus (SC) has highlighted its role in 58 cognitive processing, not just motor control (Gandhi and Katnani 2011;Krauzlis et al. 2013). We have 59 been using an additional, under-appreciated perturbation tool -the trigeminal blink reflex -to obtain 60 behavioral and neural signatures of the (normally hidden) process underlying movement preparation. 61 For example, appropriately timed reflex blinks reduce saccade reaction times significantly (Gandhi and 62

Materials and Methods 102
General and surgical procedures 103 All experimental and surgical procedures were approved by the Institutional Animal Care and Use 104 Committee at the University of Pittsburgh and were in compliance with the US Public Health Service 105 policy on the humane care and use of laboratory animals. We used two adult rhesus monkeys (Macaca 106 mulatta, 1 male and 1 female, ages 8 and 10, respectively) for our experiments. Under isoflurane 107 anesthesia, a craniotomy that allowed access to the SC was performed and a recording chamber was 108 secured to the skull over the craniotomy. In addition, a post for head restraint and a scleral search coil to 109 track gaze shifts were implanted. Post-recovery, the animal was trained to perform standard eye 110 movement tasks for a liquid reward. 111

Visual stimuli and behavior 113
Visual stimuli were displayed by back-projection onto a hemispherical dome. Stimuli were white squares 114 on a dark grey background, 4x4 pixels in size and subtended approximately 0.5° of visual angle. Eye 115 position was recorded using the scleral search coil technique, sampled at 1 kHz. Stimulus presentation 116 and the animal's behavior were under real-time control with a LabVIEW-based controller interface 117 (Bryant and Gandhi 2005). After initial training and acclimatization, the monkeys were trained to 118 perform a delayed saccade task. The subject was required to initiate the trial by acquiring fixation on a 119 central fixation target. Next, a target appeared in the periphery but the fixation point remained 120 illuminated for a variable 500-1200 ms, and the animal was required to delay saccade onset until the 121 fixation point was extinguished (GO cue). Control trials in which fixation was broken before peripheral 122 target onset were removed from further analyses. The animals performed the task correctly on >95% of 123 the trials. 124 125

Induction of reflex blinks 126
On a small percentage of trials (~15-20%), we delivered an air puff to the animal's eye to invoke the 127 trigeminal blink reflex. Compressed air was fed through a pressure valve and air flow was monitored 128 with a flow meter. To record blinks, we taped a small Teflon-coated stainless steel coil (similar to the 129 ones used for eye tracking, but smaller in coil diameter) to the top of the eyelid. The air pressure was titrated during each session to evoke a single blink. Trials in which the animal blinked excessively or did 131 not blink were aborted and/or excluded from further analyses. To obtain blink-triggered movements to 132 the peripheral target, we sought to evoke blinks 100-250 ms after the GO cue, during the early phase of 133 the typical saccade reaction time. In our experimental setup, blink onset occurs approximately 150 ms 134 after the air puff reaches the eye. Thus, air puffs were administered 50 ms before to 100 ms after the 135 GO cue. Trials in which a saccade did not accompany such blinks (i.e., where the gaze did not end up at 136 the target) were removed from further analysis. To obtain BREMs without an accompanying saccade, air 137 puff delivery was timed to evoke blinks during fixation of the central target, 400-100 ms before the 138 onset of the peripheral target. The window constraints for gaze were relaxed for a period of 200-500 ms 139 following delivery of the air puff to ensure that the excursion of the BREM did not lead to an aborted 140 trial. 141

142
Electrophysiology 143 During each recording session, a tungsten microelectrode was lowered into the SC chamber using a 144 hydraulic microdrive. Neural activity was amplified and band-pass filtered between 200 Hz and 5 kHz 145 and fed to a digital oscilloscope for visualization and spike discrimination. A window discriminator was 146 used to threshold and trigger spikes online, and the corresponding spike times were recorded. The 147 location of the electrode in the SC was confirmed by the presence of visual and movement-related 148 activity as well as the ability to evoke fixed vector saccadic eye movements at low stimulation currents 149 (20-40 A, 400 Hz, 100 ms). Before beginning data collection for a given neuron, its response field was 150 roughly estimated. During data collection, the saccade target was placed either in the neuron's response 151 field or at the diametrically opposite location (reflected across both axes) in a randomly interleaved 152 manner. Additional target locations were included in some occasions, particularly when neural activity 153 was not recorded. 154 155 Data analysis and pre-processing 156 Data were analyzed using a combination of in-house software and Matlab. Eye position signals were 157 smoothed with a phase-neutral filter and differentiated to obtain velocity traces. Normal saccades, 158 BREMs, and blink-triggered eye movements were detected using standard onset and offset velocity 159 criteria (50 deg/s and 30 deg/s, respectively). Onsets and offsets were detected separately for horizontal and vertical components of the movements and the minimum (maximum) of the two values was taken 161 to be the actual onset (offset). 162 Raw spike density waveforms were computed for each neuron and each trial by convolving the spike 163 trains with a Gaussian kernel (width = 4 ms). For a given neuron and target location, spike densities were 164 averaged across trials after aligning to saccade onset. We also normalized the trial-averaged spike 165 density of each neuron to enable meaningful averaging across the population. The activity of each 166 neuron was normalized by its peak firing rate during normal saccades. For comparison with the linear combination with delay model, we also detected saccade onset in blink-197 triggered movements using a previously used backwards velocity threshold method (Gandhi and 198 Bonadonna 2005). In order to do this, we detected the peak in the velocity profile of the blink-triggered 199 movement and marched backwards in time until the standard onset velocity criterion (50 deg/s) was 200 crossed for at least 5 consecutive time points. The idea behind this approach is that the peak velocity of 201 the movement should occur during the saccade component, and going backwards in time until the 202 threshold is crossed should isolate the saccade alone. In several sessions, peak velocity was attained 203 during the initial phase of the blink-triggered movement, which was clearly contributed by the BREM. 204 Hence, we applied the backwards threshold method starting from the second significant peak for these 205 movements. This approach is depicted in Figure 2b. 206 207

Other analyses and statistical tests 208
For the analyses that involve studying the effect on model performance of an external parameter (e.g., 209 direction or optimal delay), we binned trials from all sessions according to that parameter, and 210 computed the average mean-squared error (MSE) across trials for each bin. Significant trends were 211 identified by comparing these to the null distribution (uniform distribution/no trend) generated by a 212 bootstrap approach with appropriate confidence intervals that took into account the number of trials 213 available in a given bin. Where applicable, we used the Wilcoxon-rank-sum test for a two-way 214 comparison of distribution medians. In order to extract saccades from blink-triggered movements, we needed three sets of data. In animals 227 performing the delayed saccade task, during each session, we recorded 1) a set of normal saccades to 228 two or more targets (typically more than 100 per session), 2) a set of blink-triggered gaze shifts to those 229 targets by inducing reflex blinks after the go cue (around 100 trials per session), 3) blink-related eye 230 movement (BREM) profiles by evoking blinks during initial fixation, before peripheral target onset 231 (around 50 per session). We had 100 session-target pairs in all, and analyzed data for each session-232 target pair separately. 233 First, we simulated a range of movement profiles by linear summation of randomly chosen saccade and 234 BREM velocity profiles at all possible time shifts relative to each other. Then, for each real blink-235 triggered movement, we determined the simulated profile that best fit it (see Methods), and the 236 associated triad of saccade, BREM and time shift parametrized the decomposition of the blink-triggered 237 movement. Note that the optimal time shift effectively determined the time of saccade onset within the 238 movement. Figure 3a shows the average blink-triggered movement velocity profiles for one session-239 target pair (orange traces) and the corresponding best-fitting simulated movement profile average 240 (green traces). Next, we extracted the underlying saccade by subtracting the respective BREM profile 241 from each blink-triggered movement at the appropriate time shift. The dynamics of the extracted 242 saccades closely resembled normal saccades, as judged from the shape of their velocity profiles (Figure  243 3b, compare blue vs black traces). For comparison, we also computed an alternative saccade onset time using a backwards velocity threshold criterion (see Methods) that has been used previously (Gandhi and 245 Bonadonna 2005). Note that the dynamics of the saccade estimated using this criterion (Figure 3b, red 246 traces) deviated significantly from that of normal saccades or those extracted using the aforementioned 247 method. 248 Figure 3c shows the normalized, average radial velocity profiles across all session-target pairs for 249 normal, extracted and back-thresholded saccades. Qualitatively, the extracted saccades overlap 250 substantially with normal saccades. Some of the discrepancies could be because the dynamics of the 251 motor command, and therefore the saccade, triggered in response to a blink may be slightly attenuated 252 (for more, see below). Moreover, the velocity profiles of extracted saccades were much closer to normal 253 saccades than those of back-thresholded saccades. To quantify this observation, we computed the mean 254 square error (MSE, equivalently, Euclidean distance in velocity phase space) between normal and 255 extracted saccades, as well as between normal and back-thresholded saccades, for each session-target 256 pair. Figure  It is pertinent to remark here that the method of linear combination is certain to provide an optimal 287 decomposition of a blink-triggered movement, since any set of simulated movements is going to contain 288 one that fits the actual movement best. Although the results presented above demonstrate the 289 method's utility in extracting saccades with near-normal velocity profiles, its use in studies of the neural 290 mechanisms of movement preparation requires evidence that saccade-related neural activity is not 291 unduly affected by the re-computed onset times. In order to determine whether this was true, we 292 compared motor bursts of 51 neurons in the superior colliculus (SC) for normal saccades, and both 293 extracted and back-thresholded saccades from blink-triggered movements. We only included session-294 target pairs in which the target was in the response field of the recorded SC neuron. Figure 5a shows the 295 trial-averaged spike density of one neuron, aligned on respective saccade onsets, for each of the three 296 cases. Note the remarkable similarity between the burst profiles for normal and extracted saccades 297 (blue vs black traces). Figure 5b shows the average population activity for the three cases. To quantify 298 the similarity of the bursts, we performed the correlation analysis presented earlier for the velocity 299 profiles in Figure 3e. The distribution of correlation coefficients (Spearman's rank correlation) for the 300 two pairs of comparisons plotted against each other is shown in Figure 5c. Correlations were 301 significantly higher for the extracted saccade bursts (Wilcoxon rank-sum test, n=51, p < 10 -6 ), suggesting 302 that linear decomposition with delay largely preserves the saccade-related burst compared to back-303 thresholding. 304 Previous studies have observed strong attenuation in the saccadic burst in a few SC neurons (Goossens 305 and Van Opstal 2000b), but no suppression is apparent in our data (Figure 5a, b). To resolve this discrepancy, we more closely analyzed the timing of the saccades extracted using our approach relative 307 to the occurrence of the blink. Figure 6a shows the distribution of extracted saccade onset times relative 308 to overall movement onset across all combined blink-saccade movements. In other words, this is the 309 distribution of optimal time shifts at which linearly combining a BREM and saccade would produce each 310 blink-triggered movement. Saccade onset followed BREM onset in the majority of trials. Observe that 311 saccades can begin as late as 80 ms into the movement -an eon in the time scale of sensorimotor 312 integration -highlighting the importance of precisely determining saccade onset time for studying 313 movement preparation using the blink approach. Intriguingly, the distribution of optimal delays for 314 saccade onset was bimodal, with an early peak around 10 ms and a later peak around 40 ms after 315 movement onset. We thought this might provide a clue to the issue of suppression. Note that when 316 saccade onset follows the blink by as early as 0-20 ms, the blink is occurring at a time when the 317 command to generate a saccade has already been issued (accounting for efferent delays), and is likely to 318 affect the burst when it is approaching peak value. If so, any suppression in these cases will alter the 319 dynamics of the underlying programmed saccade and affect the ability of the algorithm to extract a 320 control-like saccade. 321 We first tested whether the performance of the linear combination with delay approach was a function 322 of the optimal delay identified by the algorithm (Figure 6b). The MSE was indeed higher for optimal 323 delays between 0-20 ms (blue traces), indicated by the increase above the bootstrapped baseline (cyan 324 traces). We then computed the suppression index (see Methods) for each neuron as a function of 325 optimal delay bin. The average index across the population is shown in Figure 6c (thick trace; thin traces 326 represent +/-s.e.m.). Note that the suppression is highest in the 0-20 ms window. Indeed, neuronal 327 suppression was a reliable predictor of the MSE, as seen in the correlation of the two variables across 328 bins (Figure 6d, R 2 = 0.36, p < 0.01). However, the magnitude of suppression was still lower (15% at its 329 maximum in Figure 6c) than previously observed (Goossens and Van Opstal 2000b). It is important to 330 note that the suppression values so far were calculated based on aligning activity to extracted saccade 331 onset, which seeks to maximize the similarity of the underlying saccade to a control movement (and 332 may thus minimize differences at the neural level as well), whereas, in previous studies, activity was 333 aligned to saccade onset estimated using the backwards velocity threshold criterion. Suppression 334 magnitudes were indeed higher for back-thresholded saccades, peaking at 35% at the population level 335 (Figure 6e). Finally, we wanted to verify whether individual neurons exhibited suppression that 336 resembled the strong attenuation observed in the previous study. Figure 6f shows the average saccade-337 aligned burst profiles for five individual neurons for control (black traces) and back-thresholded saccades (red traces), in trials that fall within the 0-30 ms window of optimal delays estimated using the 339 linear combination approach. In all five examples, a strong reduction/dip in the activity is evident during 340 the peri-saccadic period. Thus, it is possible that the strong suppression observed in the previous study 341 in a handful of neurons may be a result of a combination of factors, including timing of the saccade 342 relative to the blink. 343 344 Discussion 345 Current perspective holds that rapid redirection of the visual axis during a blink cannot be accounted for 346 by a sum of a typical saccade (Goossens and Van Opstal 2000a) and a BREM and that blinks are 347 associated with potent suppression of activity in premotor circuitry of the saccadic system (Goossens 348 and Van Opstal 2000b). In contrast, we have shown that blink-triggered movements can be 349 characterized as a linear combination of BREMs and saccades with an arbitrary delay that can be 350 estimated from the movement. Furthermore, we found that the dynamics of saccades extracted using 351 this method as well as the associated motor bursts in SC are similar to those of normal saccades. 352 The discrepancies between our results and the previous studies could be due to several factors. First, in 353 the previous behavioral study, the possibility of a staggered, time-shifted linear superposition was not 354 systematically explored (Goossens and Van Opstal 2000a). As we show here, BREM onset and saccade 355 onset can be offset by several tens of milliseconds, and taking this time shift into account can increase 356 the explanatory power of the linear analysis significantly. Second, we show that the saccade-related 357 burst is largely preserved for the extracted saccade. We believe this discrepancy could be due to the 358 fact that our analyses included a large portion of blink-triggered saccades, i.e., movements in which 359 blink onset preceded saccade onset by several tens of milliseconds, whereas previous studies focused 360 largely on co-occurring blinks and saccades. Since our laboratory's focus is to use the reflex blink as a 361 tool to probe underlying motor preparation, we induced blinks early, and thus obtained many trials 362 where the blink likely triggered neural processes that resulted in a saccade much later. In the previous 363 studies, blinks were timed to occur during typical saccade reaction times, likely resulting in instances 364 where the saccadic burst had commenced and was perturbed by the blink. This notion is consistent with 365 the observation that the suppression on blink-perturbed trials, especially for saccade onset times 366 computed using previous approaches, is maximal around 20 ms after the onset of the BREM, which is 367 when the peak of the burst would have occurred if the final motor command was issued at the time of blink interferes with the motor burst around the time of its peak. These claims are further bolstered by 370 the observation that the algorithm fails to extract saccades resembling control saccades for this window 371 of optimal delays (higher MSE, arrow in Figure 6b), possibly due to altered kinematics as a result of the 372 attenuated burst peak and by the strong suppression of the peak in example individual neurons for trials 373 that fall in this window of delays (Figure 6f). For true blink-triggered saccades, i.e., for longer optimal 374 delays between blink and saccade onset, suppression of low-frequency preparatory activity, if any, may 375 have enough time to recover to produce a full-fledged saccadic burst. The dynamics of saccades 376 extracted from such blink-triggered movements do not seem to be very different from those of normal 377 saccades; thus it is not surprising that SC activity for the two types of saccades are similar. It should be 378 noted here that, even disregarding the specific focus on the time course of the saccade relative to the 379 blink, the intensity of the motor burst was attenuated when we used the velocity-based threshold 380 criterion ( Figure 5b). 381 Another possible explanation for the discrepancy in the extent of attenuation is differences in the air puff 382 stimuli used to induce reflex blinks. The observed short latency attenuation of saccade-related activity of 383 a small subset of neurons in the previous study was locked to the time of air puff delivery, and as such, 384 may have been linked to the strength of the air puff. It is entirely possible that the air puffs we used were 385 weaker, and combined with early delivery times (when preparatory activity is lower), precluded the strong 386 attenuation described previously. It should further be noted that the previous study observed strong 387 suppression in only a small percentage of the overall population, and we observed a similar strong effect 388 in several individual neurons (some of which are shown in Figure 6f). The primary purpose of this study 389 was to find a way to precisely determine saccade onset from a blink-triggered movement, and we think a 390 systematic analysis of these multiple factors and a thorough re-verification of previous results is beyond 391 its scope. 392 The fact that BREMs and saccades can be linearly combined to produce the observed movement 393 suggests that the underlying neural processes producing a BREM and saccade are independent.