We first asked if arousal-related fluctuations are present in the SC. As in previous studies that recorded from neurons in the cortex (Cowley et al., 2020), we found that the mean spiking responses of individual SC neurons during the delay period fluctuated over the course of a session while the monkeys performed the MGS task (Figure 2A, left). Note that this effect cannot be attributed to signals related to stimulus or saccade tuning as residual spike counts were computed by subtracting the mean response to each target location from trial-to-trial responses to each target location. These fluctuations are also not likely to reflect instabilities in our linear array recordings, as several steps were taken to remove neurons that exhibited changes in waveform shape over time or large sudden changes in activity level (see Methods).

To identify coordinated fluctuations in activity across the population of neurons, we applied PCA to residual spike counts of individual neurons that were binned using a 15-min sliding window stepped every 3 mins. This yielded a vector of loadings (Figure 2A, middle) that we termed the “slow drift axis” in line with previous studies (Cowley et al., 2020). The value of each loading corresponds to the correlation coefficient between the binned response of each SC neuron and the axis in population activity space that explained the most variance in the data.. Thus, neurons whose residual activity fluctuated more over the course of a recording session had a greater portion of their variance explained by the slow drift axis (e.g. Figure 2A, Channel 2). After we identified this axis from the binned data, we then projected the residual responses of each SC neuron on single trials, which resulted in a single dimension along which we examined fluctuations over time (Figure 2A, bottom right). When the data were pooled across sessions, we found that the median proportion of variance explained by the slow drift axis was 0.88 (Figure 2A, top right). By comparing the actual distribution of values to a null distribution that was generated by shuffling the trial-to-trial spike counts for each neuron prior to performing PCA, we were able to show that the variance explained by the slow drift axis was significantly greater than what would be expected by chance (p = 0.0385).

To investigate if the SC slow drift axis was related to a global signal that might be related to shifts in arousal, we computed an identical axis using simultaneously recorded PFC data. If the SC slow drift axis represented an arousal axis that reflected coordinated activity fluctuations throughout the brain, one would expect it to be correlated with simultaneously recorded data in the PFC. For each session, we computed the correlation (Pearson’s product-moment correlation coefficient) between projections onto the SC slow drift axis and projections onto the PFC slow drift axis. As shown in Figure 2B, we found that the median variance explained across sessions, as measured using the coefficient of determination (r²), was 0.2627. To determine whether or not this effect was statistically significant, we compared the actual distribution of r² values to a null distribution that was generated by shuffling across sessions and recomputing the correlation between projection values. We found that the proportion of variance explained was significantly greater than what would be expected if there was no correlation between projections onto the SC slow drift axis and the PFC slow drift axis (p = 0.0190). These findings suggest that slow fluctuations in responsivity are present in populations of SC neurons, and furthermore that they are coordinated across the brain. To further explore how they relate to changes in arousal or brain state, we next asked if the SC slow drift axis is associated with eye metrics that are known to be correlated with arousal levels.

Decades of research has shown that the size of the pupil can be used as a reliable marker of arousal (Mathôt, 2018; Joshi and Gold, 2020). Heightened levels of arousal are accompanied by concomitant increases in LC activity and pupil size (Aston-Jones and Cohen, 2005; Murphy et al., 2014; Joshi et al., 2016; Reimer et al., 2016; de Gee et al., 2017; Breton-Provencher and Sur, 2019). Thus, we next sought to explore if projections onto the SC slow drift axis were associated with pupil size. For each session, we computed the correlation (Pearson’s product-moment correlation coefficient) between projections onto the SC slow drift axis and mean pupil size during the first 200ms of the delay period when a task-related pupil response could be observed. Across sessions, we found that the median variance explained between projections onto the SC slow drift axis and pupil size was 0.0597 (Figure 2C). When the actual distribution of values was compared to an across-session shuffle distribution, we found that the proportion of variance explained was significantly greater than what would be expected if there was no correlation between SC slow drift and pupil size (p < 0.001). Interestingly, the median variance explained between projections onto the SC slow drift axis and other eye metrics including microsaccade rate was much smaller (r² = 0.0026, p < 0.001) (Figure 2 - figure supplement 1). Although microsaccade rate (i.e. the number of small eye movements made during a period of steady fixation) has been linked to arousal (Siegenthaler et al., 2014; Johnston et al., 2022a; Contadini-Wright et al., 2023), microsaccade rates are task-dependent (Winterson and Collewijn, 1976; Ko et al., 2010; Bowers et al., 2021) and may be driven by activity in cortical as opposed to subcortical regions (Hafed et al., 2021). This might explain why no significant correlation was found between SC slow drift and microsaccade rate in the present study.

Taken together, these findings point to a link between the SC slow drift axis and brain-wide arousal signals. That simultaneously recorded PFC data and pupil size were correlated with projections onto the slow drift axis suggests that the signal we identified in the SC is highly similar to that which has been observed across several cortical areas (Cowley et al., 2020; Hennig et al., 2021). The fact that signals related to arousal are present in the SC is not particularly surprising given that its functional role has evolved and is now thought to include higher order cognitive processes (Knudsen, 2011; Krauzlis et al., 2013; Clark et al., 2015; Basso et al., 2021). However, it also poses a problem. Given the well-established role of deep-layers SC neurons in saccade generation (Sparks and Hartwich-Young, 1989; Gandhi and Katnani, 2011), having signals related to cognition and arousal embedded in oculomotor circuits that send gaze commands to downstream nuclei could lead to unwanted changes in eye position. One possibility is that arousal-related fluctuations are isolated from these deep-layer SC neurons that reside closer to the motor output. If this is the case, one would expect neurons with a stronger saccadic response to minimally fluctuate over time, and have a smaller portion of their variance explained by fluctuations along the slow drift axis.

To investigate if arousal-related fluctuations are isolated from deep-layer SC neurons, we took advantage of the fact that our linear array recordings targeted neurons residing in earlier and later stages of the visuomotor transformation (Figure 1B). As described above, this was quantified by computing a motor index for each SC neuron with high values indicating that a cell elicited a strong saccadic response to the remembered target location and low values indicating that a cell elicited a weak saccadic response. Cells with a higher motor index tend to reside in the deep layers of the SC closer to the motor output (Basso and May, 2017). In several sessions, such as that shown in Figure 3A, we observed that the loadings for the slow drift axis (i.e. the correlation between the response of each neuron and the component that explained the most variance in the data) were weaker for neurons with a higher motor index. Of note in this particular session is that these neurons were recorded from channels at the bottom of the probe, which presumably targeted the deep layers of the SC.

To explore if a similar effect was found across sessions, we computed the squared component loading for all recorded SC neurons. This was done to control for the fact that the sign of the correlation between the response of each neuron and the slow drift axis is arbitrary and could be positive or negative (Jolliffe and Cadima, 2016). Next, we discretized the motor indices of all recorded SC neurons into four equally spaced bins. Permutation tests were then performed across the four bins to investigate if neurons with a stronger saccadic response explained less variance in the slow drift axis. After controlling for multiple comparisons (Bonferroni correction), we found that the mean squared component loading for neurons with the highest motor index (M = 0.0105, SD = 0.0237) (i.e. the bin centered on 0.75) was significantly weaker than the mean loading for neurons with lower motor indices (bin centered on 0.25: M = 0.1539, SD = 0.2207; bin centered on 0: M = 0.1743, SD = 0.2429) (Figure 3B). Note that similar results were obtained when motor index was entered as a continuous variable into a multiple regression analysis instead of being discretized into four equally spaced bins (Figure 1 - figure supplement 1). The proportion of variance explained by the model, which also included subject name as a categorical variable to control for intersubject differences, was 0.0350 (F(2, 368) = 6.6748, p = 0.0014). Consistent with our hypothesis, we found that motor index was significantly associated with squared component loadings (t = −3.6154, p = 0.0003) such that SC neurons with a higher motor index explained less variance in the slow drift axis. These results demonstrate that even though arousal-related fluctuations are present in the SC, they are isolated from deep-layers neurons that elicit a strong saccadic response and presumably reside closer to the motor output. As a further test of this hypothesis, we next sought to determine if these same cells explained less variance in projections onto the PFC slow drift axis.

To probe the relationship between deep-layer SC neurons and the PFC slow drift axis, we computed the correlation (Pearson’s product-moment correlation coefficient) between the trial-to-trial residual responses of each SC neuron during the delay period and projections onto the PFC slow drift axis. Consistent with our finding that SC slow drift was significantly correlated with PFC slow drift (Figure 2B), we observed that the trial-to-trial residual responses of some SC neurons were significantly correlated with projections onto the PFC slow drift axis. For example, the residual responses of the neuron shown in Figure 4A, which has a relatively low motor index and is therefore unlikely to reside close to the motor output of the SC, were strongly correlated with projections onto the PFC slow drift axis (r² = 0.5346, p < 0.001). However, other neurons showed no correlation such as that in Figure 4B. The residual responses of this neuron, which has a relatively high motor index and presumably resides closer to the motor output stage, were not significantly correlated with projections onto the PFC slow drift axis (r² = 0.0010, p = 0.2650).

To explore if a similar effect was found across sessions, we again pooled the motor indices of all recorded SC neurons and discretized them into four equally spaced bins. We performed permutation tests across the bins to investigate if the mean variance explained between residual SC neuron responses and projections onto the PFC slow drift axis was lower for neurons with a higher motor index. After controlling for multiple comparisons, we found that the mean variance explained was significantly lower for neurons with the highest motor index SC (M = 0.0209, SD = 0.0236) (i.e. the bin centered on 0.75) than it was for neurons with lower motor indices (bin centered on 0.25: M = 0.1240, SD = 0.1468; bin centered on 0: M = 0.1557, SD = 0.1591) (Figure 4C). Furthermore, the mean variance explained for neurons with an intermediate motor index (i.e. the bin centered on 0.5) (M = 0.0703, SD = 0.1095) was significantly lower than it was for neurons with the lowest motor indices (i.e. the bins centered on 0.25 and 0). Once again, similar results were obtained when the motor index was entered as a continuous variable into a multiple regression model instead of being discretized into bins (Figure 1 - figure supplement 1). The proportion of variance explained by the model, after controlling intersubject differences, was 0.1777 (F(2, 368) = 39.7739; p < 0.001). Motor index was significantly associated with the strength of the relationship between single SC neurons responses and projections onto the PFC slow drift axis (t = −2.7902; p = 0.0055) such that it was weaker for neurons with a higher motor index. These findings provide additional support for the hypothesis that arousal-related fluctuations are isolated from neurons in the deep layers of the SC. They are more characteristic of cells residing at the early to mid stages of the visuomotor transformation, which have been implicated in autonomic processes such as controlling the size of the pupil (Wang et al., 2012; Wang and Munoz, 2018). This prompted us to ask if a relationship exists between SC neurons responses and pupil size, and whether this relationship is mediated by motor index.

Previous work has shown that intermediate-layer SC neurons control the size of the pupil via: 1) direct projections to the Edinger-Westphal nucleus; and 2) indirect projections to the mesencephalic cuneiform nucleus (Joshi and Gold, 2020). Thus, we hypothesized that neurons with a lower motor index would explain more variance in trial-to-trial pupil size than neurons in deep layers. To test this, we computed the correlation (Pearson’s product-moment correlation coefficient) between the spiking responses of each SC neuron and pupil size, both of which were computed during the first 200ms of the delay period. In several sessions, we found that the strength of this relationship was inversely related to motor index. For example, the spiking response of the neuron shown in Figure 5A, which has a relatively low motor index and presumably resides at the mid stage of the visuomotor transformation, was strongly correlated with pupil size (r² = 0.2121, p < 0.001). In contrast, the spiking responses of the neuron shown in Figure 5B, which has a relatively high motor index and presumably resides closer to the motor output stage, was not significantly correlated with changes in pupil size over the course of a recording session (r² = 0.001, p = 0.7534).

To determine if a similar effect was found across sessions, we performed a multiple regression analysis. In this analysis, we sought to determine if motor index predicted the strength of the relationship between SC neuron responses and pupil size, as measured using the coefficient of determination (r²). After controlling for intersubject differences, results showed that the proportion of variance explained by the model was 0.0309 (F(2, 368) = 5.8717; p = 0.0031). Motor index was significantly associated with the strength of the relationship between SC neuron responses and pupil size (t = −3.2593; p = 0.0039) (Figure 5C). More specifically, we found a stronger relationship between neuronal response and pupil size for neurons with a lower motor index, consistent with the proposed role of intermediate-layer SC neurons in controlling the size of the pupil (Joshi and Gold, 2020).

Taken together, the findings outlined above suggest that arousal-related signals are kept separate from neurons in the deep layers of the SC. In addition to being more weakly correlated with pupil size, these neurons did not exhibit large fluctuations in activity over time (Figure 2), and when considering the neuronal population as a whole, explained less variance in the slow drift axis when it was computed using population activity in the SC (Figure 3) and PFC (Figure 4). Although these results point to a segregation of arousal- and eye movement-related signals in the SC, research has shown that all three layers receive input from the LC: a brain region that has long been associated with controlling global arousal levels (Benavidez et al., 2021). Thus, an additional mechanism may have evolved to ensure that arousal-related signals are kept separate from those used in the encoding of eye position. In primary motor cortex, signals related to movement preparation reside in an orthogonal axis in population activity space to those that are used for movement initiation so that they do not lead to premature muscle contractions (Kaufman et al., 2014; Elsayed et al., 2016). Next, we explored if a similar mechanism exists in the SC to prevent arousal-related signals interfering with the motor output.

To investigate if arousal-related signals in the SC reside in an orthogonal subspace to those that are used to encode eye position, we computed a “pupil size axis” for each session. We discretized trial-to-trial pupil size into eight equally spaced bins and, for each neuron, computed the mean spiking response per bin (Figure 6A, left). We then used PCA to find the axis in population activity space that explained the most variance in the data attributable to the pupil. The loadings (Figure 6A, right) represent the correlation between the response of each neuron across bins and the first principal component meaning that neurons exhibiting a stronger relationship between firing rate and pupil size (e.g. Figure 6A, Channel 5) explained more variance in the pupil size axis. Our goal was to compare projections onto the pupil size axis to projections onto a “saccade tuning axis”. To compute this axis, tuning curves were generated for each SC neuron by taking the mean spiking response during the saccade epoch for each target angle (Figure 6B, right). PCA was then applied to the tuning curves, which yielded a vector of loadings for the first principal component (Figure 6B, left). In this case, neurons that exhibited greater selectivity for a particular target angle explained more variance in the saccade tuning axis (e.g. Figure 6B, Channel 11).

We first investigated if the pupil size axis and the saccade tuning axis reside in orthogonal subspaces. To ensure that fair comparisons could be made, the same data (i.e. trial-to-trial residual responses during the delay period) were projected onto each of the axes. We then computed a correlation (Pearson’s product-moment correlation coefficient) between the projection values for each session, and used the coefficient of determination (r²) to assess the variance explained. If arousal-related signals in the SC reside in an orthogonal subspace to those that are used for saccade generation, one would expect projections onto the pupil size axis to be weakly correlated with projections onto the saccade tuning axis. This was found to be the case in several sessions such as that shown in Figure 6C (r² = 0.0307, p < 0.001). Furthermore, when the data were pooled across sessions, we found that the mean variance explained between the projection values was 0.5142 (Figure 6D). By comparing the actual distribution of r² values to a null distribution that was generated by correlating projections onto the pupil size axis with projections onto an axis that was randomly oriented in population activity space, we were able to show that the mean variance explained was not significantly greater than what would be expected for data projected onto a random axis (p = 0.5192).

Finally, as an additional test of our hypothesis, we investigated if the pupil size axis was more aligned to the slow drift axis than the saccade tuning axis. To do this, we projected trial-to-trial residual responses during the delay period onto the pupil size axis and the saccade tuning axis. For each session, we then correlated (Pearson’s product-moment correlation coefficient) these projection values with projections onto the slow drift axis. Across sessions, we found that the mean variance explained between projections onto the pupil size axis and the slow drift axis (M = 0.7780, SD = 0.3025) was significantly greater than the mean variance explained between projections onto the saccade tuning axis (M = 0.5319, SD = 0.3407) (t = −4.8578, p < 0.001).

Two adult rhesus macaque monkeys (Macaca mulatta) were used in this study (Monkey Do = 28 sessions, Monkey Wa = 34 sessions). Surgical procedures to chronically implant a titanium head post (to immobilize the subjects’ heads during experiments) and microelectrode arrays were conducted in aseptic conditions under isoflurane anesthesia. Recording chambers situated along the midline at an angle of 30° from vertical and aimed at a location 8mm above interaural zero were also implanted for SC recordings. Opiate analgesics were used to minimize pain and discomfort during the perioperative period. Experimental procedures were approved by the Institutional Animal Care and Use Committee of Carnegie Mellon University and were performed in accordance with the United States National Research Council’s Guide for the Care and Use of Laboratory Animals.

The spiking responses of neuronal populations in the SC were recorded using a 16-channel linear microelectrode array (V-Probe, Plexon, Dallas, TX). The total length of the linear array was 115mm. The contacts were spaced 150μm apart and spanned a distance of 2.25mm. Linear arrays were lowered into the SC through a guide tube using a custom-designed mechanical microdrive (Laboratory for Sensorimotor Research, Bethesda, MD), which was inserted through a plastic grid with 1mm spacing.

The spiking responses of neuronal populations in the PFC (Monkey Do = left hemisphere, Monkey Wa = right hemisphere) were recorded using 96-channel “Utah” arrays (Blackrock Microsystems). The arrays comprised a 10×10 grid of silicon microelectrodes (1mm in length) that were spaced 400μm apart, implanted on the pre-arcuate gyrus (8Ar).

Signals from the linear probes (SC) and Utah arrays (PFC) were amplified and band-pass filtered (0.3–7500Hz) by a Grapevine system (Ripple). Waveform segments crossing a threshold (set as a multiple of the root mean square noise on each channel) were digitized (30KHz) and stored for offline processing, which involved waveforms being automatically sorted as spikes or noise by a neural network classifier developed in our laboratory (code available at https://github.com/smithlabvision/spikesort/nasnet). Although this method cannot distinguish between multiple neurons recorded from the same channel, it provides fast and accurate classifications that are highly similar to those of a human spike-sorter (Issar et al., 2020). All waveforms recorded from a single channel that remained after the neural network classifier had been applied were considered to be from a single unit, which we refer to as a “neuron”.

Across both monkeys, we recorded from 992 sites in SC and 5952 sites in PFC. This included all channels from the 16-channel linear microelectrode array and all channels from the 96-channel utah array across all sessions (SC: Monkey Do = 448, Monkey Wa = 544; PFC: Monkey Do = 2688, Money Wa = 3264). We applied a rigorous set of exclusion criteria designed to eliminate neurons with noisy or unstable waveforms, those with low firing rates, and neurons that did not have an evoked response during either the visual or motor epoch of the MGS

To identify candidate neurons that were mostly classified as containing noise waveforms by the neural network classifier, and therefore only comprised a small number of waveforms, we removed neurons that had a mean firing rate less than 2sp/s across the entire session (Smith and Sommer, 2013). This was calculated by taking the total number of spikes in the recording session from the beginning (before the task was started) to the end (after the last trial of the day), and dividing by the total time (including periods of task performance and periods outside of the trial). In addition to removing neurons that had a low mean firing rate across an entire session, we also removed SC neurons that did not exhibit elevated firing rates during the visual and motor epochs of the MGS tasks i.e. following the presentation of the target stimulus and during the onset of a saccade. This was quantified by computing a visual and motor index for each SC neuron (see below). Neurons were removed if the value of both of these indices was less than 0.05, indicating a lack of response during both epochs. To rule out the possibility that slow drift arose due to recording instabilities (e.g. the distance between the neuron and the microelectrode(s) changing over time), we only included neurons with stable waveform shapes throughout a session. Neurons were removed if they had a percent waveform variance greater than 10% (Cowley et al., 2020). We also removed neurons that exhibited abrupt changes in activity during a recording session. This was quantified by dividing each session into 20 non-overlapping time bins and computing the mean spike count per bin (Cowley et al., 2020). Neurons were removed if the change in activity between any two consecutive time bins was greater than 25% of the maximum activity of that neuron. Adopting these exclusion criteria ensured that slow drift did not arise due to recording instabilities. We also note that if drift were to occur due to recording instabilities, it would not be expected to affect neurons from the same penetration differentially based on their motor index (e.g., Figure 3A).

In order to obtain a precise estimate of the slow drift axis, and after controlling for low firing rate and recording instabilities, we removed sessions that contained less than two SC or PFC neurons (Monkey Do = 3 sessions, Monkey Wa = 5 sessions). Furthermore, we removed sessions if they had a duration lasting less than 45 minutes (Monkey Do = 0 sessions, Monkey Wa = 3 sessions). Adopting these additional criteria ensured that there were enough variables and observations to perform PCA and accurately compute the slow drift axis. Using all of this criteria, the total number of sessions included in the analysis for Monkey Do and Monkey Wa was 25 sessions (SC = 190 neurons, PFC = 795) and 26 sessions (SC = 181, PFC = 2110), respectively.

Visual stimuli were generated using a combination of custom software written in MATLAB (The MathWorks) and Psychophysics Toolbox extensions (Brainard, 1997; Pelli, 1997; Kleiner et al., 2007). They were displayed on a CRT monitor (resolution = 1024 × 768 pixels; refresh rate = 100Hz), which was viewed at a distance of 57cm and gamma-corrected to linearize the relationship between input voltage and output luminance using a photometer and look-up-tables.

Subjects fixated a central point (diameter = 0.18°) on the monitor to initiate a trial (Figure 1A). After fixating within a circular window (diameter: Monkey Do = 2.38°, Monkey Wa = 1.10°) for 200ms, a target stimulus (diameter = 0.55°) was presented for 200ms at one of eight angles separated by 45°. The eccentricity of the target stimulus was determined on a session-by-session basis by applying low-amplitude microstimulation to contacts along the linear microelectrode array to evoke saccadic eye movements. Target eccentricity was then set such that it matched the endpoints of evoked saccades. After the target stimulus had been presented, subjects were required to maintain fixation for a delay period. The duration of the delay period chosen at random on each trial from a uniform distribution spanning 600-1100ms. If steady fixation was maintained throughout the delay period, the central point was extinguished prompting the subjects to make a saccade to the remembered target location. The subjects had 500ms to initiate a saccade. To receive a liquid reward, the subjects’ gaze had to be maintained within a circular window centered on the remembered target location (diameter: Monkey Do = 2.93°, Monkey Wa = 2.20° for the first 12 sessions and 3.01° for the remaining 22 sessions) for 100ms. In a subset of sessions, the target was briefly reilluminated, after the fixation point was extinguished and the saccade had been initiated, to aid in saccade completion.

Eye position and pupil diameter were recorded monocularly at a rate of 1000Hz using an infrared eye tracker (EyeLink 1000, SR Research).

Microsaccades were detected during the delay period and were defined as eye movements that exceeded a velocity threshold of 6 times the standard deviation of the median velocity for at least 6ms (Engbert and Kliegl, 2003) (Figure 2 - figure supplement 1). To prevent overshoots (i.e. fast, oppositely directed gaze shifts that occur shortly after a microsaccade) being treated as separate eye movements, we merged microsaccades that occurred within 50ms of each other (McCamy et al., 2015). Finally, we assessed the validity of our microsaccade detection method by computing the correlation (Pearson’s product-moment correlation coefficient) between the amplitude and peak velocity of detected microsaccades for each session. The mean correlation between these two variables across sessions was 0.7601 (SD = 0.0678) indicating that our detection algorithm was robust, as microsaccades fell on the main sequence (Zuber et al., 1965).

To characterize the neurophysiological response properties of individual SC neurons, we first visualized SC responses by computing PSTHs aligned to the onset of the target stimulus and the onset of the saccade to the remembered target location, respectively (Figure 1B). Firing rates were computed in 1ms bins, averaged across trials and smoothed using a Gaussian function (σ = 5ms).

To further quantify the strength of the saccadic response to the remembered target location, a motor index was computed for each neuron as follows: Where M represents the mean response during a peri-saccade epoch from 25ms before to 25ms after the onset of the saccade, and B represents the mean response during a baseline period defined as 100ms before the onset of the target stimulus. Higher motor index values are indicative of a stronger saccadic response relative to baseline, whereas lower values represent a weaker saccadic response relative to baseline. Note that we also computed a visual index for each neuron, which was only used for the purpose of excluding neurons that did not exhibit an evoked response during the MGS task (see above). The visual index was computed in the following manner: Where V represents the mean response during a peri-stimulus epoch spanning 25ms to 125ms after onset of the target stimulus, and B represents the mean response during a baseline period defined as 100ms before the onset of the target stimulus.