Multiple groups of neurons in the superior colliculus convert value signals into saccadic vigor

Summary

Eye movements toward high-valued objects are executed with greater vigor. To test how neurons in the superior colliculus (SC), a subcortical structure that controls these movements, are modulated by value, we recorded four SC neuron subtypes while monkeys made saccades to objects previously associated with high or low reward volumes. High-value objects elicited greater activity in three neuron subtypes (visual, visuomotor, motor). Using a bootstrapping method, we identified three distinct activity phases: early visual response (E_VIS), late visual response (L_VIS), and pre-saccadic (Pre_SAC) motor response. Value was positively correlated with activity in the L_VIS and Pre_SAC phases, but not in the E_VIS phase, suggesting that value modulates visual and motor stages of visuomotor transformation. Additionally, we discovered a class of tonically active neurons that decrease their activity upon object onset, and remain inhibited till the end of the saccade, potentially enhancing saccade execution by disinhibiting interactions among other SC neurons.

Graphical abstract

Highlights

• •Saccades to high-value objects have shorter reaction times and higher peak velocities
• •The learned object values positively modulate the majority of SC neurons
• •Value impacts both the visual and motor stages of visuomotor transformation in SC

Neuroscience; Behavioral neuroscience; Systems neuroscience; Sensory neuroscience

Introduction

Reward is a critical factor that shapes the behavior of all organisms. Positive outcomes encourage animals to approach or explore objects and lead to invigorated action, whereas negative consequences lead to avoidance and slower actions.^1,2 The influence of reward can be seen even in simple saccadic eye movements, such that reaction times are faster and peak velocities are higher for saccades to high-value objects and/or locations.^3,4,5 This behavioral modulation by reward necessitates an interaction between value processing networks and motor control networks within the brain, which enables the integration of value information with movement planning and execution.

Saccadic eye movements are planned and executed by a network of brain regions, including the frontal eye fields, parietal cortex, and superior colliculus (SC).^6,7 In this pathway, SC is the final critical node that can serve as an interface between cognitive signals and movement planning^8,9 prior to the generation of the motor commands by the brainstem circuits to initiate saccades.^10,11 The SC receives sensory input and transforms it into motor commands that drive eye movements with the help of three distinguishable neuronal cell types -visual, visuomotor, and motor neurons.^12,13 These neurons are differentially localized in the superficial and intermediate layers of SC, forming a local circuit. The activity of these neurons controls saccade initiation and execution, which can be measured as reaction time (RT) and peak velocity (PV), respectively.^{14,15,16,17,18}

The brain contains a large network of cortical and subcortical regions that process and assign value to the incoming visual inputs. The basal ganglia are a critical node in this network and are particularly well-documented for their role in influencing saccadic eye movements.¹⁹ The neural activities in multiple basal ganglia (BG) nuclei, including the caudate nucleus,^20,21 globus pallidus,^22,23 and substantia nigra^24,25 are known to differentiate between high-valued and low-valued objects/positions. This value information is then transmitted from the output nuclei of BG, the substantia nigra pars reticulata (SNr), to other brain regions dedicated to controlling saccadic behavior, such as the SC.²⁶ Previous studies have shown that BG input to SC takes the form of disinhibition^27,28,29 mediated by the SNr neurons. Additionally, SNr neurons are modulated by the value of objects present in the contralateral visual field, reducing their activity for high-valued objects and enhancing it in the case of low-valued objects.^25,30 Thus, these SNr neurons can interface with the visuomotor circuits in the SC and gate behavior by facilitating saccades to valuable objects in the environment.

In addition to BG, many other cortical regions are also involved in processing value information and provide value-related inputs to the neurons in SC. Previous studies have shown that neurons in the frontal eye fields (FEF) are modulated by the expected value of the target prior to the action.^31,32,33 Similarly, studies done in the supplementary eye fields (SEF),^34,35,36 as well as in the lateral interparietal area (LIP),^37,38 have also shown that neurons in these regions are modulated by the expected reward at the end of the trial. Previous work from the lab using fMRI identified multiple regions across the brain that are modulated by value information.³⁹ More importantly, many of these cortical regions are known to have direct or indirect projections to SC^40,41,42,43 and may influence the saccadic behavior. Value-sensitive inputs from these cortical and/or subcortical sources may interface with the local visuomotor circuits in SC formed by the multiple functional classes of neurons, affecting their responses. Understanding the dynamics of these interactions is crucial for a comprehensive mechanistic view of how reward shapes saccade behavior.

Previous studies have shown that SC neurons were modulated by reward expectation using an asymmetrically rewarded 1-DR task^44,45 in which saccades to one direction were rewarded more than others in a block of trials. Enhanced activation was seen in some neurons when a target was placed in the location where the monkey was anticipating a high reward during the pre-target baseline and the initial transient phase of the post-target response. More recently, Griggs et al. (2018) investigated the effect of value on the responses of visual neurons in the SC using visual objects previously associated with higher or lower volumes of juice through repeated, consistent exposure for more than 10 days. After the learning phase, the same objects were flashed in the receptive fields (RFs) of the SC neurons while the animal fixated on a central point. Visual neurons in the superficial layers of SC were activated during this passive peripheral viewing of the objects and responded with a higher firing rate to high-valued objects than to low-valued objects. That study⁴⁶ used a passive fixation task to record only visual neurons in SC; hence, the impact of learned value on different neuronal subtypes that form the local circuitry in SC and their effects on saccade behavior were not fully assessed.

In this study, we compared the activity of different functional subtypes of SC neurons recorded using a linear array while the animals performed visually guided saccades toward high-value vs. low-value objects. In contrast to single electrode recordings, a linear array allows the monitoring of neurons in different layers of SC simultaneously and thus provides a better picture of the local SC circuit. The recorded neurons were broadly classified into four functional subtypes. High-value objects elicited more spikes than low-value objects in three of the four subtypes of SC neurons, ∼100 ms after target onsets. We also found that the stronger activation in these neurons predicted enhanced behavioral responses - lower RT and higher PV - seen when saccades were directed to high-valued objects.

Results

The SC contains multiple functional subtypes of neurons that play distinct roles in generating saccades. Value information from other cortical (e.g., FEF, LIP) and subcortical regions (e.g., SNr in the BG) is thought to interface with these neuronal subtypes and modulate their response properties, thereby affecting saccadic responses to valuable objects. Here, we investigate how the response profiles of these different SC neurons are affected by value, thereby modulating the RT and PV of the upcoming saccade. We tested this by recording SC neurons while the animals performed a visually guided saccade task (Figure 1A) toward fractal objects⁴⁷ previously associated with large or small volumes of juice reward (Figure 1B) presented in the contralateral visual field. A linear array with 24 contact points was positioned in the SC through a posteriorly implanted recording chamber, and multiple subtypes of neurons were simultaneously recorded from different layers (Figure 1C). We used the recorded activity in different time windows (Figure 1D) to classify the SC cells into four functional subtypes (See methods). In each recording session, monkeys made saccades to fractal targets appearing at a specific eccentricity based on the identified RF of the recorded SC neurons (Figure 1E). Across sessions, targets appeared at eccentricities ranging from 3 to 25°, and monkeys, on average, made 56 (min of 20 to max of 86) saccades to the high, and 43 (min of 13 to max of 66) saccades to the low objects.

Figure 1.

Methods used in this study

(A) Progression of events in a single trial of the visually guided saccade task.

(B) The four object sets (eight fractals in each) were used as stimuli in the task. In each set, four objects were associated with a larger volume of juice (high objects), and four objects were associated with a lower volume (low objects). Animals learned the value of high and low objects through repeated exposure to the task.

(C) A sagittal section of the brain from monkey DWN, visualized by MRI, showing the location of the posterior recording chamber and grid. A recording track targeting the right hemisphere of SC is shown as a blue line. The different subtypes of SC neurons recorded simultaneously using a linear array during an example session are also shown as a schematic.

(D) The neural activity aligned on target onset (left) and saccade onset (right) is shown for a representative neuron. The firing rates in these time intervals (colored patches) were quantified and used to classify the SC cells into different subtypes. (E) A map showing the potential anatomical locations of recordings in SC derived based on the receptive field of the recorded neurons over 33 separate sessions. Sessions recorded from monkey DWN are shown in gray squares and from monkey BLY in black circles.

Behavioral modulation of saccades

First, we tested whether the saccadic eye movements in this task confirmed the previously reported value-based behavioral modulation.^4,48 We separately computed the RT distribution of saccades to high and low-valued objects for each recording session (Figure 2A). In 26 out of 33 sessions, the RT distribution for high-valued objects significantly differed (unpaired t test, p < 0.045) from those for low-valued objects. The mean high object RT across all sessions (186 ± 4 ms) was significantly lower than the low object RT (211 ± 6 ms) (Figure 2D paired t test, n = 33, p < 0.001), corroborating that saccade initiation is strongly modulated by value.

Figure 2.

Value modulation of saccadic behavior

Saccadic behavior toward high-valued (red) and low-valued (blue) objects is shown.

(A) Average cumulative reaction time distributions of high and low object trials from all the recording sessions are shown with the standard error of means as the colored envelope.

(B) Average normalized velocity profiles of saccades directed to high and low objects are shown with the standard error of means as the colored envelope.

(C) Average normalized saccadic displacement profiles for high and low trials are plotted separately. The mean RT (D), mean PV (E), and mean amplitude (F) calculated separately for high and low objects in each session (connected by the light gray line) are shown to depict the variability in behavior across different recording sessions. The data obtained from the two animals are shown separately using different symbols. Bar graphs denote the population mean, and the error bars denote the standard error of means calculated across all 33 recorded sessions. A paired t test was used to test the significance. ∗∗∗ denote p < 0.001, ∗∗ denote p < 0.01, ∗ denotes p < 0.05, and ns denotes p ≥ 0.05.

(G) A scatterplot of the value modulation (difference between high and low object trials) observed in RT and PV across all the recorded sessions. The data obtained from the two animals are shown separately using different symbols.

(H). The extent of modulation (high-low) in RT (gray squares) and PV (black circles) due to the experimental manipulation of value is plotted against saccade amplitude across all recorded sessions. N = 33 recorded sessions in all these panels.

Next, we calculated the average velocity profiles in each session separately for saccades to high and low objects to test the effect of value on saccade execution. The PV of saccades toward high objects was significantly greater than saccades to low objects in 16 out of 33 sessions (unpaired t test, p < 0.04). Within each session, we computed the mean PV of saccades to high and low objects, and then compared these session averages across sessions using a paired t test. Across sessions, the mean high object PV (582 ± 40 deg/sec) was significantly greater than the corresponding mean low object PV (555 ± 38 deg/sec) (Figure 2E paired t test, n = 33, p < 0.001). During the execution of saccades, peak velocity is affected by the intended amplitude through the main sequence relationship.^49,50 Here, monkeys made saccades of different amplitudes ranging from 3 to 25°, and there was a strong correlation between the PV and the amplitude (Pearson’s correlation = 0.93, p < 0.001). Hence, we normalized the velocity profiles as described in the methods and plotted them separately (Figure 2B) to show that saccades to high objects were executed faster compared to low objects. In addition, we tested whether the difference in PV observed during saccades to high and low-valued objects is due to changes in saccade amplitude. There was no significant difference in the average saccade displacement profiles (Figure 2C KS Test, p > 0.05) in all of the 33 recorded sessions or in the average amplitude (Figure 2F paired t test, n = 33, p = 0.16) observed between the high (12.78 ± 1.3°) and low (12.89 ± 1.3°) object saccades; hence the PV modulation observed in saccades was independent of the amplitude. This further corroborated the idea that, in this dataset, the value associated with high and low objects modulated the execution of saccadic eye movements, as previously shown.

We further tested whether the value modulations observed in RT and PV are related by plotting the difference in RT between high and low objects against the difference in PV across sessions. There was no clear evidence of a systematic relationship (Figure 2G Pearson’s correlation = −0.22; p = 0.22) between RT and PV modulations, suggesting that value may independently affect these behavioral metrics. Further, we also checked if the different target eccentricities used in this study had any systematic effect on the observed value modulation of saccadic behavior (Figure 2H). As expected, sessions with larger target eccentricities showed stronger modulation of saccadic peak velocity (Pearson’s correlation = 0.47; p = 0.006), while there was no effect of eccentricity on the observed RT modulation (Pearson’s correlation = 0.12; p = 0.53).

Value modulation in superior colliculus neurons

Next, we analyzed the responses of SC neurons as the animal directed their saccades toward the high and low-value objects (Figure S1 for representative neurons; Figure 3 for population activity). We analyzed 520 neurons recorded from two monkeys, which were then divided into four functional subtypes (See methods for classification). The population responses were calculated separately for each functional subtype of visual, visuomotor, and motor neurons aligned on target onset (Figures 3A–C) as well as saccade onset (Figures 3D–F).

Figure 3.

Population level value modulation observed in SC neurons

The average SC population response aligned on stimulus onset when saccades were directed to the high objects (red) and low objects (blue) in the RF is shown separately for visual (A), visuomotor (B), and motor (C), neurons. The population response of the SC neurons, when aligned on saccade onsets, is shown separately for visual (D), visuomotor (E), and motor (F), neurons. The firing rate calculated in the time intervals −300 to 0 (light blue) before the target onset, greater than 80 ms (light green) aligned on the target onset, and −40 to 0 ms aligned on the saccade onset (light violet) are shown as bar graphs separately for the high (red outline) and low (blue outline) trials. The bar graphs represent the average firing rates across the neurons that were separately quantified for the visual (G), visuomotor (H), and motor (I), neurons. The error bars denote the standard error of means. A paired t test was used to test the significance. ∗∗∗ denote p < 0.001, ∗∗ denote p < 0.01, ∗ denotes p < 0.05, and ns denotes p ≥ 0.05. The piecharts below show the proportion of positively modulated (gray), negatively modulated (black), and non-modulated (white) neurons within the populations of visual (J), visuomotor (K), and motor (L) neurons. N in the first and second rows of panels denotes the number of positively modulated neurons identified in the dataset, which were used to visualize the time course of value modulation. The N in the third row of panels denotes the total number of cells recorded under each functional subtype of neurons that have been included in the statistical analysis reported in the text.

Visual neurons exhibited an initial transient increase in their firing with a latency of 20–40 ms when an object appeared in their RF (Figure 3A). At about 80 ms after target onset till 170 ms, their responses, which were sustained above the baseline, became larger to high valued objects (27 ± 2 spks/s) than low valued (22 ± 2 spks/s) objects (paired t test, n = 233, p < 0.001). When aligned on saccade onset (Figure 3D), we did not observe any pre-saccadic burst (−40 to 0 ms) in these neurons; instead, the activity was maintained stably above the baseline with higher (paired t test, n = 233 p < 0.001) activation for high objects (23 ± 2 spks/s) compared to low objects (18 ± 1 spks/s). This pattern of higher activation for high-value objects (positive modulation) was observed in 55% (Figures 3A and 3D, n = 129) of the recorded neurons. Interestingly, a minority of visual neurons (25%; n = 59) showed the opposite negative modulation, i.e., higher activation for low-value objects. There were also 20% (n = 45) of visual neurons that showed no significant difference in the activity between high and low object saccades (Figure 3J).

Visuomotor neurons (Figure 3B) showed a clear excitatory phasic response to target onset, lasting until 80 ms. These neurons maintained a higher (paired t test, n = 88, p < 0.001) level of activity for high-value objects (56 ± 5 spks/s) in the time window 80–170 ms compared to low objects (46 ± 5 spks/s). A clear excitatory presaccadic burst (Figure 3E) was seen in these neurons when the activity was aligned with saccade onset, and this burst was positively modulated for high-value objects (FR_High = 73 ± 7 spks/s compared to FR_Low = 69 ± 7 spks/s, paired t test, n = 88, p = 0.010). Most visuomotor neurons were positively modulated 70% (Figures 3B and 3E, n = 61), while the negatively modulated cells constituted only 10% (n = 9). The remaining 20% (n = 18) of the visuomotor neurons showed no significant difference between the high and low valued object conditions (Figure 3K).

Motor neurons, on the other hand, had a weak visual response (Figure 3C) but a clear excitatory burst prior to the motor onset (Figure 3F). These neurons responded with higher magnitude (Figure 3I paired t test, n = 95 p < 0.001) when the ensuing saccade was directed to high valued objects (37 ± 4 spks/s) compared to low valued objects (24 ± 3 spks/s). Their value-coding was less pronounced and restricted to the pre-saccadic period (Figure 3G) when aligned on saccade onset (FR_High = 89 ± 9spks/s compared to FR_Low = 83 ± 8 spks/s, paired t test, n = 95, p = 0.031). Similar to visuomotor neurons, most (74%, Figures 3C and 3F n = 70) of the motor neurons were also positively modulated, and only a small minority 4% (n = 4) were negatively modulated. The remaining 22% of the motor neurons (n = 21) showed no difference in activity between high and low object conditions (Figure 3L).

Previous studies had clearly shown that the neural activity in SC could reliably predict the RT and, to a lesser extent, the PV of the ensuing saccade. In addition, our previous analysis also showed that eye movements directed to high-valued objects were initiated and executed with greater vigor compared to low-valued objects. This raised the possibility that the modulation in firing rate we observed in SC was related to the vigor of upcoming eye movements rather than to value information. To test this, we used two approaches. In the first approach, before calculating the firing rates of individual neurons in a session, we matched (see methods) the RT and PV distributions observed in the high and low-value conditions, thereby ensuring that the impact of saccadic vigor on firing activity was minimal. Then, we ran a repeated-measures ANOVA with value (high vs. low) and time windows (baseline, 80 to 170 ms, −40 to 0 ms) as within-neuron factors. This revealed that, even after matching the vigor of the saccades, there was a significant main effect of value (F₁,₁₈₈ = 21.73, p < 0.001) and a significant value × epoch interaction (F₂,₃₇₆ = 14.76, p < 0.001) for visual neurons, indicating that the magnitude of value modulation varied across time. The same analysis was repeated for visuomotor (Value main effect: F₁,₈₂ = 19.28, p < 0.001, Value × epoch interaction, F₂,₁₆₄ = 14.24, p < 0.001) and motor neurons (value main effect: F₁,₇₃ = 21.26, p < 0.001, Value × epoch interaction, F₂,₁₄₆ = 10.82, p < 0.001) as well. Post-hoc comparisons across all three subtypes of SC neurons showed no significant modulation during the baseline period, but strong modulation during the window 80-170 ms and during the presaccadic −40 to 0 ms epoch. The mean firing rates and corresponding p-values were tabulated in the Table S1. This showed that SC responses are strongly modulated by value independent of saccade vigor.

In the second approach, we quantified how value and saccade variables jointly influence SC activity by constructing a pooled generalized linear model (GLM, see methods) that incorporated data from all neurons across sessions. Analysis of the cross-validated GLM coefficients (Figure S5A) enabled us to identify the unique contributions of value and movement-related variables to firing-rate variability, after statistically accounting for shared variance among predictors. The cross-validated coefficient for value was significantly shifted in the positive direction for visual (β_Value = 0.198, p < 0.0001), visuomotor (β_Value = 0.152, p < 0.0001) and motor (β_Value = 0.209, p < 0.0001) neurons, indicating that high-value objects reliably evoked stronger population firing than low-value objects even after accounting for RT, PV, and amplitude. In addition, coefficients related to RT were significantly different from zero for visuomotor (β_RT = −0.203, p < 0.0001) and motor (β_RT = −0.405, p < 0.0001) neurons. But in the case of visual neurons (β_RT = −0.049, p < 0.0001), RT had weak contributions. In contrast, the coefficients for PV and amplitudes displayed weaker contributions for all three cell types. The majority of explained variance (indexed by 10-fold cross-validated R² _Visual = 0.013, R² _VisMot = 0.063, R² _Motor = 0.207) was therefore attributable to value and RT compared to the other movement parameters.

To determine whether the observed value effect could arise from spurious correlations or residual collinearity with RT or PV, we conducted two control analyses. We shuffled the value labels randomly, resulting in approximately half of the trials having their value labels reassigned, which reduced the model’s predictive power (R² _Visual = 0.005, R² _VisMot = 0.058, R² _Motor = 0.198) and centered the value coefficients around zero (Figure S5B). In a 100% label flip (high ↔ low), the sign of the value coefficient was inverted as expected, and cross-validated R-square (R² _Visual = −0.026, R² _VisMot = 0.040, R² _Motor = 0.164) reduced relative to the true data. The coefficients and R-squared values for all three neuronal subtypes are tabulated in Table S2. These results demonstrated that the value term contributed unique, cross-validated predictive power to SC firing rather than being a collinear effect of RT, PV, or amplitudes. Together, these analyses demonstrated that the SC population encoded object value robustly, even after rigorously controlling for saccade vigor.

Tonic neurons: A new class of superior colliculus neurons

Unlike the other three groups of SC neurons, tonic neurons had significant activity (45 ± 5 spks/sec) during the inter-trial interval and during fixations. (Figure 4A). In response to objects (high or low) presented in their RFs, their activity was reduced to (12 ± 2 spks/sec) at a latency of ∼40–60 ms. (Figures 4A and 4C (left), n = 88, paired t test, p < 0.001). This inhibition persisted at least until the saccade onset and activity started to increase post-saccade (Figure 4B). To the best of our knowledge, these types of tonic SC neurons had not been reported previously in non-human primates (NHPs). Earlier studies had identified a group of neurons in the rostral part of SC with foveal RFs and reduced their activity ∼40 ms prior to the initiation of saccades.^51,52 This pause in the activity of these rostral neurons was required to remove the fixation and initiate saccadic eye movements. The tonic neurons identified here had markedly different properties from those of the foveal neurons in the rostral SC.^16,53 Unlike rostral foveal neurons, the tonic neurons had eccentric receptive fields ranging from 4 to 25° (Figure 4D). They were not specifically localized in the rostral zone of the SC and were found in recording tracts in the caudal part of the SC (Figure 1E). More importantly, unlike foveal neurons, which were modulated by the onset/offset of the central fixation cue, most tonic neurons (77/88) were unaffected by the onset of the fixation cue. There was no significant difference in the activity recorded before (15 ± 2 spks/s) and after (16 ± 2 spks/s) the onset of the fixation cue (Figure 4C (right), n = 88, paired t test p = 0.06). Given this, we suggest that tonic neurons constitute a previously unreported class of SC neurons.

Figure 4.

Properties of tonic neurons

The average response of SC tonic neurons when aligned on target onset (A) and saccade onset (B) is shown.

(C) The mean firing rate from the population of tonic neurons in a 200 ms time interval before (white) and after (gray) target onset (left) and fixation onset (right) is shown as bar graphs. The error bars denote the standard error of means. A paired t test was used to test the significance. ∗∗∗ denote p < 0.001, and ns denotes p ≥ 0.05.

(D) The eccentricity of the identified RFs of all the recorded tonic neurons (n = 89) is shown. The green circles denote the tonic neurons, which had significant changes in activity in the 200 ms time interval before and after the onset of fixation. The average population responses of the negatively modulated (E) and the positively modulated (F) tonic neurons aligned on the onset of the visual target are shown. N indicates the number of neurons that contribute to visualizing the time course of positively and negatively modulated tonic neurons. Responses to high valued (red) and low valued (blue) saccade targets are shown separately. For statistical analysis, all the recorded tonic neurons are pooled together and the firing rate calculated in the time intervals −300 to 0 (light blue) before the target onset, greater than 80 ms (light green) aligned on the target onset, and −40 to 0 ms aligned on the saccade onset (light violet) are shown as bar graphs separately for the high (red outline) and low (blue outline) trials (G). The error bars denote the standard error of means. A paired t test was used to test the significance. ns denotes p ≥ 0.05. The pie charts (H) show the proportion of positively modulated (gray), negatively modulated (black), and non-modulated (white) tonic neurons.

Next, we also looked at the value responses of these tonic neurons. Unlike the other three groups of SC neurons (Figure 3), we found approximately equal proportions of negatively (Figure 4E, 42%, n = 37) and positively modulated (Figure 4F, 38%, n = 34) tonic neurons. The remaining 20% (n = 18) of tonic neurons did not show any modulation in response to high and low objects (Figure 4H). Hence, as a population, there was no significant difference in their responses to high (12.4 spks/s) and low (12.6 spks/s) valued objects when aligned on target onset (Figure 4G, n = 89, paired t test p = 0.78). Similarly, no significant difference (n = 89, paired t test p = 0.69) was observed between high (11.56 spks/s) and low (11.87spks/s) value conditions during the pre-saccadic interval when the population activity of these neurons was aligned on saccade onset (n = 89, paired t test p = 0.70).

Identifying different phases in superior colliculus neural responses

The three groups of SC neurons (visual, visuomotor, and motor neurons) played significant roles in transforming sensory visual stimuli into saccadic eye movements.^54,55,56 The above data clearly showed that value-related modulation was observed in these three types of neurons, but it is unclear which epoch of the sensorimotor transformation was affected by this modulation. Did the value modulation in these neurons affect the visual processing or the pre-saccadic motor stage? In this set of experiments, we used a visually guided saccade task, in which the visual and motor stages of sensorimotor transformations are seamlessly linked. Hence, to test this, we need to disentangle the visual and motor stages in the neural responses of these neurons, as done previously.^57,58 If neuronal activity in an epoch represents visual processing, it should exhibit minimal temporal variability when aligned to object onset. On the other hand, the motor processes will show temporal variability when aligned on object onset. If the neuronal activity in an epoch represents the pre-saccadic motor process, it should exhibit minimal temporal variability when aligned on the saccade onset; the visual process, on the other hand, will show temporal variability when aligned on the saccade onset. Using the above principle, we tested whether the value modulation observed in each neuron was temporally associated with the visual stage (onset of the object) or with the motor stages (i.e., the start of the saccade motor plan) of the sensorimotor transformation.

To discriminate these stages of sensorimotor transformation in the neural activity of SC neurons, we separated trials into three groups based on saccade latency (e.g., short, medium, and long). Next, we simulated neuronal population responses for each SC subtype using a bootstrapping method (Figure S2) that utilized simultaneous recordings of multiple neurons obtained with linear arrays (see details in methods). Although cases of biased perceptual reports driven by single-neuron activation had been reported,⁵⁹ behavioral control in natural conditions typically relied on the concerted activity of neuronal populations. The bootstrapping method we used, by pooling the responses of multiple neurons that control saccades in a single trial, mimics this natural biological process unfolding in SC. In this analysis, we divided the neural responses across trials into three quantiles (short, medium, and long) based on saccade RT. We simulated the population response of a hypothetical trial in a particular quantile by pooling single-trial responses from the same quantile across 25 random neurons. This process reduced the variability in the single neuron response related to many non-specific factors, such as object position, day-to-day variation in the animal’s motivation, and so forth. 15–35 trials were simulated for each quantile, and the mean firing rate and mean time of peak activity during different phases of the SC response were separately quantified. This bootstrap was repeated 10000 times to estimate the mean and the 95% confidence intervals. A bootstrap-based permutation test (BBP test), which compared the F-statistics from the original data with a null distribution obtained by bootstrapping and shuffling the labels (short, medium, long RTs), was conducted to determine the p-value (see methods).

The comparison of population activity simulated using the bootstrap procedure across the three RT quantiles revealed different phases with distinctive features. All three SC neuronal subtypes showed an initial visual transient response lasting ∼80 ms. More importantly, the activity in this epoch peaked approximately (Figures 5B, 5F, 5J cyan line) at 42 ms (range: 39-44 ms) irrespective of the three RT groups. Minimal temporal variability was associated with this peak when aligned on target onset for visual (Figure 5A, BBP test p = 0.366), visuomotor (Figure 5E, BBP test p = 0.690), and motor (Figure 5I, BBP test p = 0.728) neurons, indicating that this epoch represented visual processing and was dubbed as early visual response (E_VIS).

Figure 5.

Identifying different phases in SC neuronal response

Simulated population activity of SC neurons generated by bootstrapping individual neuron responses separated by short (red), medium (yellow), and long (orange) RT trials is plotted. The activity is aligned with the target onset and plotted separately for visual (A), visuomotor (E), and motor (I) neurons. The vertical dashed lines denote the three phases of SC response: (1) early visual (E_VIS) component (cyan) between 30 and 80 ms, (2) late visual (L_VIS) component (maroon) between 80 and 130 ms, (3) tertiary pre-saccadic (PRE_SAC) component (gray) beyond 130 ms. The saccade RTs of trials belonging to the short (red), medium (yellow), and long (orange) groups are shown separately at the bottom of the graphs. The average time at which firing activity peaked in each RT quantile is shown separately for the E_VIS (cyan), L_VIS (maroon), and PRE_SAC components (gray) for Visual (B), Visuomotor (F), and Motor (J) neurons. The error bars denote the 95% confidence interval for the time to peak quantified across 10000 bootstrapped repeats. P-value is determined using a bootstrap-based permutation test. ∗∗∗ denote p < 0.001, ∗∗ denote p < 0.01, ∗ denotes p < 0.05, and ns denotes p ≥ 0.05.Simulated population activities of SC neurons are aligned on saccade onset and plotted separately for visual (C), visuomotor (G), and motor (K) neurons. The vertical dashed lines denote the pre-saccadic phase of SC response: −40 to 0 ms (green). The average time at which firing activity peaked in the pre-saccadic phase for each RT quantile is shown separately for visual (D), visuomotor (H), and motor (L) neurons. The error bars denote the 95% confidence interval for the time to peak quantified across 10000 bootstrapped repeats. P-value is determined using a bootstrap-based permutation test. ns denote p > 0.05.

Following this initial epoch, the activity in visual and visuomotor neurons increased again, resulting in a second peak within the interval of 80-130 ms. Although the difference in the levels of activity observed in this epoch contributed (BBP test, p_visual = 0.56; p_vismot = 0.011; p_motor = 0.027) to the latency of the upcoming saccade (short, medium, long), the activity reached the peak on average at 106 ms; visual neurons (101 ms) visuomotor neurons (109 ms) and motor neurons (108–110 ms) (maroon line in Figures 5B, 5F, 5J). Since there was no significant variability (BBP test, p_visual = 0.407; p_vismot = 0.577; p_motor = 0.725) in the peak time when aligned on target onset, the activity in this second epoch was also related to visual processing and was henceforth called the late visual response (L_VIS).

After the L_VIS response (>130 ms), visual, visuomotor, and motor neurons continued to be active. Unlike early and late visual responses, the activity peaked in this epoch at different times, ordered by the saccade latencies of the short, medium, and long RT groups. This distinction was clear for visuomotor (Figure 5E, BBP test, p_vismot = 0.010) and motor (Figure 5I, BBP test, p_motor < 0.0001) neurons (gray line in Figures 5F and 5J) but not so well for visual neurons (Figure 5A). This indicated that the activity after the L_VIS epoch may be temporally correlated with the saccade onset, representing the pre-saccadic motor process. Indeed, when the same activities were aligned on the saccade onset, their peaks were aligned with no significant variability in the visuomotor (Figure 5G, BBP test, p_vismot = 0.439) and motor (Figure 5K, BBP test, p_motor = 0.796) neurons. The timing of the peak activity in this third epoch was found to be ∼9 ms before saccade onsets (Figures 5H and 5L), indicating that the activity represents the pre-saccadic (Pre_SAC) motor response. In the case of visual neurons, the peak activity was not aligned on saccade onset, but they varied systematically (Figure 5D) with respect to the RT groups, suggesting that these neurons encode only visual processes.

These data suggested that SC neurons (except tonic neurons) tend to have three distinct epochs/phases of activity that occur sequentially. Consistent with the previous studies, we were able to identify that the first two phases - E_VIS and L_VIS - represent the visual processing stages of sensorimotor transformation.^60,61 The third epoch, representing motor planning, was called pre-saccadic activity (Pre_SAC). Next, we compared the results from the above bootstrapping procedure with the analysis shown in Figure 3. This comparison suggested that value modulation observed in SC neurons affected the late visual and pre-saccadic phases of neuronal responses.

To test this more systematically, we applied the bootstrapping method separately for high and low objects, as shown in Figure 6. The high and low-value object responses in the short (Figures 6A, 6G, 6M), medium (Figures 6C, 6I, 6O), and long (Figures 6E, 6K, 6Q) RT trials were compared. The mean firing rate, the 95% confidence intervals and the p-values estimated from the confidence intervals for each of the three previously identified phases (E_VIS, L_VIS, Pre_SAC) are tabulated in the Table S3. In visual neurons (Figure 6B) and visuomotor neurons (Figure 6H), we found that E_VIS responses were not strongly affected by the object value (red vs. blue). In the motor neurons (Figure 6N), where the initial visual response was very weak, we found only negligible differences in the activity between high and low conditions. In contrast to the E_VIS phase, the activity during the L_VIS phase was stronger for high objects than low objects in all three subtypes of neurons: visual neurons (below the maroon line in Figures 6A, 6C, and 6E), visuomotor neurons (below the maroon line in Figures 6G, 6I,6K) and motor neurons (below the maroon line in Figures 6M, 6O, and 6Q). In the Pre_SAC phase, where the activity represented the motor process in SC, we found that high objects elicited stronger responses than low objects in the case of visuomotor (below the gray line in Figures 6G, 6I, 6K) and motor (below the gray line in Figures 6M, 6O, 6Q) neurons. The Pre_SAC activity was different based on object value (high object > low object Figures 6L and 6R) in all saccade RT groups (short, medium, and long). Even though there was no clear motor processing represented by pre-saccadic buildup of activity in the case of visual neurons (below the gray line in Figures 6A, 6C, 6E), there were clear differences between high and low conditions in all three RT quantiles (Figure 6F). This analysis clearly indicated that value modulation affects both the visual (L_VIS) and the motor (Pre_SAC) stages of SC neuronal response.

Figure 6.

Effect of value on different phases of SC response

Simulated mean responses of visual neurons to high (red) and low (blue) valued objects in the RF generated by bootstrapping individual neuron responses separately for short (A), medium (C), and long (E) RT trials are plotted. The shaded envelopes denote the standard deviation in the spike density functions measured across the 10000 repeats of the simulations. The saccade RTs of high and low object trials are shown separately at the bottom of the graphs. The vertical dashed lines denote the different phases identified in the responses of each SC subtype. The line graphs plot the mean firing rate and the 95% confidence intervals for the high (red) and low (blue) object responses during the E_VIS (B), L_VIS (D), and PRE_SAC (F) phases of the response. P-values are directly computed from the confidence intervals obtained from bootstrapped samples. ∗∗∗ denote p < 0.001, ∗∗ denote p < 0.01, ∗ denotes p < 0.05, and ns denotes p ≥ 0.05. The value effect in each RT quantile (short (red), medium (yellow), and long (orange)) is separately shown in each graph. Simulated population responses of visuomotor neurons (G-L) and motor neurons (M-R) generated by bootstrapping individual neuron responses separated by short (top), medium (middle), and long (bottom) RT trials are plotted. The rest of the conventions are the same as in panels A–F.

Role of tonic neurons in saccade generation

Previous analysis had shown that, as a population, tonic neurons were not clearly modulated by value, as they had equal proportions of positively and negatively modulated neurons. To identify the role of these neurons in sensorimotor transformation, we simulated the population activity of tonic neurons after dividing the trials into the short, medium, and long RT groups. The simulated population activity of tonic neurons (Figure 7) in each RT quantile was not significantly different from each other. It peaked ∼49 ms [47 to 50 ms] in the initial E_VIS epoch (BBP test, p_Evis = 0.841) and then fell below the baseline. In the following L_VIS epoch, the tonic neuron activity in all RT quantiles (BBP test, p_Lvis = 0.510) briefly peaked at ∼108 ms while remaining below the baseline. In the pre-saccadic phase, the activity related to the short RT trials reaches the minimum first (157 ms), followed by the medium RT trials (172 ms), and then the longer RT trials (191 ms, Figures 7A and 7B). Although this temporal ordering did not reach statistical significance (BBP test, p_preSAC = 0.501) in this dataset, the consistent temporal progression in reaching the lowest firing (trough) before the saccade onset when aligned on target onset was qualitatively compatible with motor processing. Given this weak but orderly trend, it is likely that the rate of inhibition in tonic neurons may be correlated with the onset of saccadic eye movements. When aligned on saccade onset, as expected, the variability in the motor processes was reduced (Figures 7C and 7D, BBP test, p_preSAC = 0.923) with the activity for all RT quantiles reaching the trough 33 ms prior [-29 to −38 ms] to the saccade onset. Given this, we suggest that the inhibitory activity in tonic neurons may contribute to the onset of saccadic eye movements.

Figure 7.

Role of tonic neurons in saccade generation

Simulated population activity of tonic neurons generated by bootstrapping individual neuron responses separated by short (red), medium (yellow), and long (orange) RT trials is plotted aligned on target (A) and saccade (C) onset. The time at which the activity reached its minimum firing rate is calculated for each RT quantile and plotted when the activity is aligned on the target (B) and saccade onset (D). The error bars denote the 95% confidence interval in the minimum time calculated across the 10000 repeats of the simulations. P-value is determined using a bootstrap-based permutation test. ns denote p > 0.05.

Value-modulated superior colliculus neurons control saccade behavior

According to the data so far (based on bootstrapping), we found that the activity of SC neurons (visual, visuomotor, motor) may be divided into three phases: 1) Early visual response (E_VIS), which was not modulated by object value 2) Late visual response (L_VIS), which was strongly modulated by object value (high > low) 3) Pre-saccadic activity (Pre_SAC), which was slightly modulated by object value (high > low). It was also known that SC population responses had a systematic relationship to the upcoming saccade in that higher firing rates accompany shorter RT and higher PV.^14,18 Taking these together, we hypothesized that SC neurons may act as an interface to functionally associate the object value with the behavioral enhancement of saccades to rewarding objects. This led to the prediction that the degree of value modulation found in each neuron would correlate with the strength of that same neuron’s influence on saccade behavior.

To test this prediction, we used three metrics that quantified the different influences on SC neural activity. Value modulation in each neuron was parameterized by quantifying the difference between the average vigor-matched firing rates of high and low objects during the time epoch 80-160 ms (value index). Vigor matched firing rates in this analysis ensure that the colinear influence of RT and PV on the SC firing rates was minimal. Next, we quantified the impact of individual SC neurons on saccade behavior by calculating Pearson’s correlation between firing rates and RT/PV separately, using all trials irrespective of their value condition. We defined the RT Index as the correlation between firing activity in the interval (80-160 ms after target onset) and the corresponding RT in that trial for each SC neuronal subtype (Figures S3A–S3D). Similarly, we used the correlation between firing rate in the same epoch (80-160 ms) and peak velocity (PV Index) to quantify individual neurons’ influence on the execution of the upcoming saccade separately for each SC neuronal subtype (Figures S3E–S3H). As expected, higher neural activities were associated with shorter RT (negative correlation Figures S3A–S3D) and higher PV (positive correlation Figures S3E–S3H). The correlation was calculated in three separate epochs: −100 to 0 ms, 0 to 80 ms, and 80–160 ms. Interestingly, the number of neurons with a significant correlation between firing rate and RT increased in epochs closer to saccade initiation, reflecting the evolution of the motor plan in SC (Figures S3I–S3L). As expected, a higher proportion of motor neurons (46%) have their firing activity significantly correlated with RT in the time interval 80-160 ms compared to visual (22%), visuomotor (35%), or tonic (21%) neurons. The number of neurons with firing activity correlated with PV (8% averaged across 4 subtypes) is lower than the number of neurons modulated by RT.

We found significant correlations between the value index and behavioral indices in specific subtypes of neurons (Figure 8), confirming our prediction that the degree of value modulation found in each neuron will correlate with the strength of that same neuron’s influence on saccade behavior. In visual neurons, the correlation was not significant (p = 0.853 Figure 8A) between value and RT indices, and +0.21 (p = 0.005, Figure 8B) between value and PV indices. The correlations between RT and value indices dramatically increased for visuomotor (−0.4, p < 0.001 Figure 8C) and motor neurons (−0.33, p = 0.005 Figure 8E). In contrast, the correlation between PV and value indices did not change in the case of motor neurons (+0.24, p = 0.045 Figure 8D) and was not significant in the case of visuomotor neurons (p = 0.38 Figure 8F). In the tonic neurons, the correlation between value and RT indexes was −0.43 (p < 0.001, Figure 8G), and +0.31 (p = 0.015 Figure 8H) for correlation between value and PV indices. This analysis suggested that the heightened SC response to high objects translate into shorter RT and higher PV. Thus, SC neurons act as a bridge between the value processing circuits in the brain and the motor circuits that control saccades.

Figure 8.

Relationship between value modulation and saccadic behavior

Scatterplot showing the relation between reaction time index (correlation between firing rate and reaction time) and value modulation (the difference in the vigor matched firing rate between high and low objects in the time interval 80-160 ms). Relations are assessed separately for visual (A), visuomotor (C), motor (E), and tonic (G) neurons. Each circle represents a single SC neuron. The best-fit regression line (black line) is shown with a negative slope in all cases. Filled circles denote the neurons that showed significant value and RT modulation indices. Scatterplots of peak velocity index (correlation between firing rate and peak velocity) vs. value modulation (the difference in vigor matched firing rate between high and low objects in the time interval 80-160 ms) are shown for visual (B), visuomotor (D), motor (F), and tonic (H) neurons. The best-fit regression lines have positive slopes in all cases. Filled circles denote the neurons that showed significant value and PV modulation indices. Pearson’s correlation coefficient (r) and associated p-value (p) calculated for each neuron subtype are shown above the graph.

Next, we examined the distribution of the value, RT, and PV modulated cells by plotting their cumulative probability against the indices used to quantify the extent of modulation (Figures 9A–9C). The value modulation was quantified using AUROC, which spans from 0 to 1. We found that, on average, 77% (73–81%) of the recorded neurons in each SC neuronal subtype were value modulated. More interestingly, within the value modulated units, the majority of them were positively modulated in the case of visual (68%), visuomotor (84%), and motor (97%) neurons, while only 46% of the tonic neurons showed the same positive modulation (Figure 9A). RT modulation was seen in 30% (averaged across the 4 subtypes: visual = 21%, visuomotor = 35%, motor = 44% tonic = 18%) of the recorded neurons and was lower than the proportion of value modulated cells (77%) in the same population. The majority of the RT-modulated cells in the visual (33/49), visuomotor (31/31), and motor (42/42) neurons had higher firing rates when saccades were initiated with short latencies (negative correlation Figure 9B). On the other hand, only 5 out of 16 RT-modulated tonic neurons (31%) had their firing rates negatively correlated with reaction time, while the remaining 11/16 (69%) cells were positively correlated.

Figure 9.

Distribution of modulated SC neurons

(A) Cumulative plot shows the number of neurons that are modulated by value from visual (gray), visuomotor (brown), motor (black), and tonic (yellow) neurons in SC. The value modulation is assessed using AUROC. The neurons with AUROC values from 0 to 0.5 denote positively modulated cells, and those from 0.5 to 1 denote negatively modulated neurons. The total number of neurons from each SC subtype modulated by value is also shown in their corresponding colors.

(B) Cumulative plot shows the number of neurons that are modulated by RT from visual (gray), visuomotor (brown), motor (black), and tonic (yellow) neurons in SC. The RT modulation is quantified as the Pearson’s correlation between firing rates in the time interval 80-160 ms and the corresponding trial RT. The total number of neurons from each SC subtype that are RT modulated is also shown in their corresponding colors.

(C) Cumulative plot shows the number of neurons that are modulated by PV from visual (gray), visuomotor (brown), motor (black), and tonic (yellow) neurons in SC. The PV modulation is quantified as Pearson’s correlation between firing rates in the time interval 80-160 ms and the PV of the corresponding trial. The total number of neurons from each SC subtype that are modulated by PV is also shown in their corresponding colors. Venn diagram shows the overlap between value (pink), RT (green), and PV (blue) modulated neurons calculated separately for visual (D), visuomotor (E), motor (F), and tonic (G) neurons. The intersection area where all three circles meet denotes the neurons that are modulated by all three factors. The intersection area where two circles meet denotes neurons that are modulated by value/RT, value/PV, and RT/PV. The area of the circle that does not intersect with the others denotes neurons modulated by only one factor. The number of neurons belonging to each of the above groups is also given.

In contrast, PV modulation was seen only in a minority, 8% of SC cells (averaged across 4 subtypes: visual = 6%, visuomotor = 11%, motor = 13% tonic = 3%) compared to the value (74%) and RT (31%) modulated cells. The majority of these PV-modulated neurons in the visual (12/14), visuomotor (8/10), and motor (11/12) neurons had their firing rates positively correlated with the PV of the upcoming saccade (Figure 9C). This analysis showed that there are many more value-modulated neurons in SC, of which a smaller proportion play a significant role in converting the value signal into behavioral saccadic vigor.

We also looked at the overlap between the value, RT, and PV-modulated cells separately in each functional subtype of SC neurons (Figures 9D–9G). Similar patterns of overlap were seen across all four subtypes. Interestingly, only a very small proportion of visual (2%), visuomotor (5%), motor (7%), or tonic (2%) neurons were modulated by all three factors simultaneously. Among the visual and tonic neurons, 14% of the cells were modulated by both value and RT. The proportion of such dual (value/RT)-modulated neurons increased to 28% in visuomotor and motor neurons. The number of dual (value/PV)-modulated neurons was similar across the four subtypes and was much lower (3%) in proportion to the other type of dual (value/RT) modulated neurons. This analysis suggested that PV and RT modulated cells are largely non-overlapping subgroups within the group of value-modulated SC neurons.

Value signals interface with the local circuits in the superior colliculus

The analysis so far has clearly shown that, despite their different functional roles, the different subtypes of SC neurons respond similarly to the value associated with visual objects. In the visual, visuomotor, and motor neurons, the activity differences appeared only during the L_VIS and the Pre_SAC phases of the response, whereas activity during the E_VIS phase did not distinguish between high and low-valued objects. The delayed onsets of value modulation in these three types of neurons suggest that neural activity in SC evolves over time: E_VIS in response to the appearance of visual objects in the RF, and later L_VIS in response to the object’s value. To test this, we quantified the time at which the activity profiles of high and low-valued objects diverged significantly, i.e., the onset of value modulation (See methods). To visualize the value modulation (Figure S4) more clearly, the difference between the activity of high and low objects (SDF_High – SDF_Low) was calculated and plotted. On average, the visual (Figure S4A), visuomotor (Figure S4B), and motor neuron (Figure S4C) showed a positive value modulation. Since tonic neurons did not show any clear value responses as a population, they were not included in this analysis. We observed that the value information appeared in the visual (94 ms) and visuomotor neurons (91 ms) approximately at the same time, while it appeared in the motor neurons (103 ms) about 10 ms later (Figures S4A–S4C). Next, we repeated the same procedure in individual neurons and obtained a distribution of value modulation onsets separately for each functional subtype of SC neurons. We did not observe a significant difference in the distributions of onset values across the three classes of SC neurons (one-way ANOVA, F = 1.71, p = 0.1834). On average, the effect of value modulation became significant at 112 ± 4 ms in visual neurons (Figure S4D), 100 ± 4 ms in visuomotor neurons (Figure S4E), and 108 ± 3 ms in the motor neurons (Figure S4F). This showed that value-related afferent inputs from other cortical or subcortical sources interface with the three functional subtypes of SC neurons, thereby modulating the vigor of the upcoming saccade to high-valued objects.

Discussion

We investigated how value-sensitive inputs from other brain regions interact with different subtypes of neurons in the SC that form the visuomotor circuit and modulate saccadic behavior. Using linear arrays, we assessed value modulation in four different subtypes of SC neurons while the animals were engaged in a visually guided saccade task, which is attuned to their natural behavior. SC is an important hub of sensorimotor transformation; hence, the neurons here are involved in both visual processing of the target and motor preparation of the saccades. A bootstrapping method was used to simulate population activity of different neuronal subtypes and identify the various phases of sensorimotor transformation in the response profiles of SC neurons. We identified three distinct phases in SC responses: an early visual phase agnostic to value, a late visual phase, and a tertiary presaccadic motor phase that are modulated by the value of the target present in the RF of the neurons. Three out of four SC neuronal subtypes respond with higher firing rates to high-value objects than to low-value objects during the late visual and the pre-saccadic phases of activity, which resulted in enhanced saccadic responses toward high-valued objects. Additionally, we identified a linear relationship between the extent of an individual neuron’s value modulation and its influence on saccade RT and PV. The greater the value modulation, the greater the cell’s impact on saccade behavior (Figure 8). In conclusion, this study comprehensively assessed the effect of value across different functional subtypes of SC neurons, suggesting that they act as a bridge between circuits that process value information and those that control motor actions.

Value response in superior colliculus neurons

In this study, we observed that saccades directed to high-valued objects are initiated and executed faster than those directed to low-valued objects. This observation is consistent with other studies in humans^3,62,63 as well as NHPs^4,64 that reported similar value-driven saccadic invigoration.

We observed higher firing rates in three functional subtypes of SC neurons when the NHPs directed saccades to high-valued objects in the RFs of these neurons in the time interval 80-160 ms after target onset. Interestingly, previous studies that attempted to address value modulation in SC neurons reported different responses.^44,45 In those studies, visual neurons showed two types of significant modulations in response to reward expectation. (1) bias modulation, in that the activity of neurons before the target onset (baseline) showed a ramping of activity, (2) gain modulation, in that the initial transient response (E_VIS here) of SC neurons showed differential activation, both well correlated with the animal’s expectation of reward from a particular target location.⁴⁵ In addition, the value modulation observed in the motor neurons had two distinct temporal profiles: (1) positive coding neurons exhibited higher activation throughout the trial, starting prior to the presentation of the cue till the execution of the saccade (2) negative coding neurons exhibited higher activation to low value, largely restricted to the pre-saccadic buildup phase.⁴⁴ In contrast, this study found that the value responses observed in three subtypes of visual, visuomotor, and motor SC neurons shared common features. The baseline activity and the E_VIS phases of the response did not differentiate between objects' values, while the L_VIS and Pre_SAC phases are significantly modulated in all three neuronal subtypes (Figures 3 and 6). In addition, both the positively and negatively modulated neurons reported here showed opposite firing patterns in response to value, but they had similar temporal profiles of activation.

We hypothesize that the variation in SC neuron responses may arise from differences in the tasks used in these studies. Here, the value was associated with the identity of an object through repeated and consistent exposure for more than ten days, giving rise to long-term object value. In the earlier two studies, an asymmetrically rewarded (1-DR) task was used, in which one target position on the screen was associated with a higher reward, while the remaining positions yielded lower rewards. The reward assigned to the location flipped after the end of a block (∼30 trials), giving rise to a short-term position-value association. Given this difference in task design, there are two plausible explanations.

• 1.It is likely that SC neurons encode the positional and object values differently. The positional value response appears in SC visual neurons even before target onset in “bias-type” neurons, whereas in “gain-type” neurons, value modulation appears concomitantly with the early visual transient.⁴⁵ In contrast, the modulation due to object value appeared in SC neurons in the interval 80-160 ms after target onset. We intend to test this in a follow-up study by recording the same pool of neurons while the animal performs the two different tasks.
• 2.Previous studies have shown two parallel circuits in the basal ganglia^65,66 that separately mediate memory for long-term stable values (such as the stable object task used in the current study) and memory for short-term flexible values (such as the 1-DR task). The flexible circuit originates from the head of the caudate nucleus and projects to the rostral ventral medial (rvm) part of SNr. On the other hand, the stable circuit originates from the tail of the caudate nucleus and projects to the caudal dorsal lateral (cdl) part of the SNr.^21,25,67 Both the circuits from rvmSNr and cdlSNr converge onto SC and modulate behavior. The differences in the value responses of SC neurons between the two studies may arise from the differential activation of these parallel circuits in BG triggered by the long-term stable vs. short-term flexible values involved in the tasks used.

The higher activation seen in these neurons might be related to the value of the expected reward or to the degree of motivation that the animal had to acquire the reward. Prior experiments^35,68 were conducted in which the monkeys were trained to saccade toward objects that provided different volumes of juice reward (positive outcome) and objects that necessitated a long wait time before the next trial (negative outcome). In this task, the degree of motivation was controlled independently by the magnitude of the reward promised for success and the magnitude of the penalty threatened for failure. The large reward and long wait time cues have high behavioral relevance, and neurons sensitive to motivation will have similar activation profiles. In contrast, neurons sensitive to the expected value will have distinct activation profiles since the animals clearly prefer the positive outcome. Unfortunately, the task design used here does not allow us to distinguish between these alternatives. These studies have further shown that neuronal representations of expected value, independent of the actions, are seen in prefrontal regions, while more posterior regions in the saccade network are strongly influenced by the motivation to acquire the reward and the ensuing motor preparation.⁶⁹ The neural activity in SC is also known to be related to motor output; hence, the modulation we observed in SC neurons is likely related to the motivation the animal had to acquire the high-valued object. The strong relationship observed between value modulation and saccade parameters strongly favors this possibility.

Anticipation of a high-valued reward leads to stronger motivation, as well as greater arousal and attention. Interestingly, previous studies have shown that the activity of SC neurons is strongly influenced by such factors.^70,71,72,73 For example, neurons have higher firing rates when an object appears in a location the subjects had covertly attended. The activity is lower when the object appears in the same location while the subject’s attention is directed elsewhere. High-valued objects, due to their higher behavioral relevance, tend to draw more attention compared to low-valued objects.⁷⁴ Hence, it may have contributed to the activity modulation observed here. In other words, the value modulation we observed here may also subserve the attentional mechanisms⁷⁵ previously reported in the SC, or vice versa; however, given the task we employed, it is not possible to distinguish the specific contribution of attention from that of value information.

Similarly, previous studies have also shown that SC neurons discriminate targets of the upcoming saccade from a distractor in their RF with higher firing rates during an oddball visual search task.^76,77 Visuomotor neurons showed a biphasic activation profile where the initial visual response did not discriminate the target from the distractor, while the second visual phase was strongly modulated to signal the targets.⁶⁰ Here, we find a similar biphasic response not only in the visuomotor but also in the visual and motor neurons. The same underlying mechanism of target discrimination may be utilized in the current task to discriminate between high and low targets and manifest as the value modulation during the late visual phase. Regardless of the underlying mechanism, the observed modulation in SC neurons with extrafoveal RFs would help the monkeys identify and saccade toward the high-valued object among a multitude of distracting objects presented in their peripheral vision.⁷⁸

In addition, the SC neurons in the superficial layers, which receive inputs from the early visual processing areas, are known to generate a saliency map of the external world.^79,80 Hence, objects that have higher luminance, contrast, or other such physical attributes will affect the firing rates of visual neurons. In this study, we used multiple sets of 8 fractal objects as targets, generated randomly with varying colors, contrast, shapes, and so forth, and half were randomly assigned as high-valued objects. The animals were repeatedly exposed to these objects and their associated reward for more than 10 days, allowing them to learn the ecological salience of these fractal objects, which is verified using their gaze position under free viewing.^65,74 This procedure will nullify any systematic effects that low-level visual features of the objects have on the firing rate of SC neurons. Hence, the contribution of physical saliency to the observed value modulation of visual neurons might be minimal.

Source of value signals

The SC receives inputs from multiple regions of the brain, which are organized into different layers. The superficial input layers receive major inputs from the early visual areas in the cortex^81,82,83 and the retina,^84,85,86 while the intermediate layers receive information from a variety of sources such as the basal ganglia, frontal eye fields, supplementary eye fields, and lateral inter-parietal area.^87,88,89 An early low-latency visual epoch is seen in all four subtypes of SC neurons, which is agnostic to the value of objects in the RF and is likely driven by the sensory inputs from multiple early visual regions. Conversely, the activity during the later visual epoch, i.e., 80-160 ms after target onset, may be driven by inputs from other brain regions such as the cdlSNr of the BG, FEF, LIP, and so forth. We have seen that activity in this later visual epoch, as well as the presaccadic motor epoch of SC response, is modulated by value and may originate from these sources. More importantly, this shows that reward value does not indiscriminately alter sensory representations but selectively modulates the later integrative stages of sensorimotor transformation.

One of the most prominent inputs to SC that actively modulates saccadic responses is the output nucleus of the BG, SNr, which tonically fires to inhibit SC neurons, keeping the animal’s eye fixated. Prior to a saccade, SNr neurons reduce their firing rates, disinhibit the SC, and facilitate movement initiation.^19,90,91 A previous study has shown that the antidromic stimulation of visual, visuomotor, and saccadic SC neurons localized in different layers of SC can activate SNr neurons.²⁷ In addition, orthodromic stimulation of SNr neurons influenced the motor neurons in the SC.²⁸ More importantly, previous work from the lab showed that SNr neurons are phasically inhibited by a high-valued object, facilitating a saccade toward it. In contrast, the same neurons are phasically excited by a low value, which delays the saccade toward the low object.^24,47 This long-term object-based value modulation of SNr activity is seen around 100 ms and is thought to be driven by visual input from the tail of the caudate nucleus.²⁵ Here, using a similar long-term stable value saccade task, we showed that the value modulation appeared in SC between 95 and 115 ms after target onset in three subtypes of neurons (Figure S4). This timing suggests that cdlSNr neurons may interface with visual, visuomotor, and motor neurons in SC, thereby inducing value responses quickly. This is supported by a recent study in which the optogenetic activation of the CDt-cdlSNr pathway by laser applied in the SNr modulated the activity of SC visual neurons.³⁰ Given this evidence, we speculate that value-sensitive output from the cdlSNr interfaces with all three subtypes of SC neurons, which are localized in different layers of the SC, and facilitates the observed behavioral modulation of saccades.

In addition to SNr, neurons in the frontal eye field also send saccade-related information to the SC^88,92 and are known to have neurons modulated by reward expectation.^32,35 In a gambling task, they also showed persistently higher firing for high-valued objects even after the saccade, which was considered to be helpful in learning the reward association.³³ The FEF neurons are known to have persistently higher firing for high-valued objects after the cue onset, and sometimes during their baseline periods.³¹ Similar to the results shown here, all the 3 functional subtypes of neurons seen in the FEF were also affected by the value task, with visuomotor neurons having the highest number of modulated cells. In contrast, we have shown that the firing activity of ∼80% of visuomotor and motor neurons in SC is affected by the value of the objects. Similar neuronal responses to value have also been seen in other cortical areas such as SEF and LIP,^34,93,94,95 which play a major role in generating saccades and send direct projections to the colliculus.^40,89,96 Given the similarity of value responses in these different regions, it is challenging to speculate on the source of value inputs to the SC neurons. In addition, the tasks used in these studies were not comparable to the fractal-based visually guided saccade task used here. Moreover, the timing of value onsets was not quantified in these studies, making comparisons difficult. We intend to pursue this using more rigorous methodologies in future studies.

Local circuits in superior colliculus for saccade invigoration

The superior colliculus has a well-defined laminar structure, with the different functional subtypes of neurons organized into specific layers. Previous research with SC slice recordings has shown local circuitry comprising connections between the different layers that facilitate the flow of activity in a structured manner from visual neurons in the superficial layers to the visuomotor neurons and motor neurons in the intermediate and deep layers.^97,98,99 There are also feedback connections, both excitatory and inhibitory, from the motor neurons in the deep layers back to the superficial neurons.^100,101 This interacting local circuit in SC can have profound impacts on the incoming inputs. For example, a brief optogenetic stimulation of the cdlSNr-SC pathway (<20 ms) had a prolonged effect (∼200 ms) on SC responses.³⁰ The brief disinhibition from cdlSNr may trigger the local excitatory interactions in the intermediate and deep layers of SC,^102,103 thereby enhancing the burst activity of motor neurons until a saccade is generated. Value inputs from other cortical and subcortical regions impinge on the intermediate layers of SC and may propagate through the local circuit to the other classes of neurons. Consistent with this idea, albeit statistically not significant, we found a trend that the value modulation first appeared in the visuomotor neurons at a latency of 95 ms and then propagated to the visual (98 ms) and motor neurons (113 ms). Laminar probes allow us to track these local interactions between various cell types in the different layers of SC, and we intend to pursue this in the future.

We have shown that the L_VIS phase of SC neurons significantly differentiates between high and low objects. The same activity also has a systematic relationship with the onset of saccades as well as the peak saccade velocities. Consistent with this result, previous studies have shown that motor neuron activity directly influences behavior by modulating saccade RT and peak velocity.^14,16,18,104 In another report, the activity of visual neurons was modulated when stimuli of higher/lower intensities appeared in their RFs, resulting in the modulation of saccadic RT.^105,106 In this study, we observed that all four subtypes of SC neurons have a systematic relationship to the initiation and execution of the saccade. Moreover, we have shown that the modulation observed in the SC neurons is strongly influenced by value independent of the ensuing saccade preparation. The extent of an individual neuron’s modulation with respect to high and low-valued objects is linearly related to the same neuron’s influence on saccade RT and PV. The greater the value modulation, the greater the cell’s impact on saccade behavior. As expected, visuomotor and motor neurons exert relatively stronger influences on behavior than visual and tonic neurons. Hence, we were able to assess the relative contributions of different neuronal subtypes in controlling the value-based enhancement of saccadic behavior.

We also found that the neuronal activity of the majority of SC neurons (>70%) in each subtype is modulated by the expected value of reward that the animals received after the saccade. However, we found that only a small subset of value modulated neurons showed significant correlations between their firing rates and the RT(∼30%) and PV(∼8%) of the upcoming saccade (Figure 9). Interestingly, similar patterns were seen in other cortical regions, such as FEF, which is modulated by value. It is possible that the non-saccade value-modulated SC neurons send their outputs to modulate other effectors, such as the neck and hand,^107,108,109 and may facilitate the invigoration of the coordinated movements to high-valued objects, which is critical for animals foraging in natural environments. SC is an important hub that controls the escape behavior – those innate, fast, reflexive responses seen in animals in aversive situations. The value-modulated non-saccade neurons in SC may also project to the downstream regions such as peri-aqueductal gray or amygdala via pulvinar and control this escape behavior.^9,110,111 In addition to output structures, these value-modulated non-saccadic neurons may send ascending projections to other brain regions. There is a well-known direct projection from SC to the substantia nigra compacta, thought to generate a phasic response in dopamine neurons.^112,113 There is also evidence for the existence of disynaptic connections to the caudate nucleus through the centromedian and parafascicular nucleus of the thalamus from SC.^114,115,116 Through these pathways, the higher activity in SC neurons in response to rewarding stimuli may play a critical role in learning novel value associations.

Tonic neurons

In this study, we observed a class of SC cells that were named tonic neurons and had not been extensively studied previously. As the name suggests, these neurons are tonically active and are inhibited when a target appears in the cell’s RF in the contralateral visual field and remain inhibited till after the end of the saccade. Such neurons were reported in the superficial layers of the mouse SC.^117,118 These tonic neurons have activity profiles distinct from those of the foveal neurons in rostral SC, which briefly pause their tonic activity during the saccade execution.^51,53,119 Unlike these foveal neurons, tonic neurons have RFs ranging from 4 to 25° and are found mingled with other subtypes along the dorsoventral axis of SC. These neurons are also distinct from the visual tonic neurons found in SC, which show a brief pause in the delay period activity before the saccade onset, driven by the appearance of the central cue that indicates the nature of the upcoming movement.¹²⁰ Interestingly, we found that activity in the majority of the tonic neurons is unaffected by the onset of the central fixation cue. So, tonic neurons might be a novel class of neurons that have not been reported earlier in NHPs.

We speculate that the tonic neurons may be part of the local inhibitory network, presumably, the GABAergic neurons reported previously in SC slice recordings, which help coordinate the flow of information between the different layers of SC. Fast-spiking GABAergic interneurons with different types of axonal morphologies are reported in the intermediate SC to provide intralaminar and interlaminar inhibition. These neurons play an important role in shaping the activations in the motor maps of SC and facilitate accurate saccades.^103,121 It has been shown that reducing the level of GABA-mediated inhibition in the local circuit can induce bursts of spikes in premotor intermediate-layer neurons with minimal excitation in the superficial visual layers.¹²² Hence, these local inhibitory neurons may modulate the firing of premotor neurons in the intermediate layers to facilitate faster saccades or further inhibit the non-target motor neurons, thereby sharpening the activity in the SC motor maps to facilitate accurate saccades. Consistent with this idea, we found that when the activity of tonic neurons reached their lowest activity state earlier, it resulted in faster saccades RTs, suggesting a potential disinhibitory influence. We suggest that when there are no objects in the RF, higher baseline firing in the tonic neurons may inhibit the interaction between the other classes of SC neurons. Hence, they may play a significant role in enabling stable fixation, which is very important for perception, attention, and social interaction. When an object (high or low value) appears, tonic neurons lower their firing activity (Figures 4A and 4B), which would disinhibit visual, visuomotor, and motor neurons, thereby facilitating saccades. We speculate that this mechanism of disinhibition by tonic neurons is particularly relevant in the case of low objects, where a saccade has to be generated toward the object despite lower activation in the visual, visuomotor, and motor neurons. We have also seen equal proportions of positive and negatively modulated tonic neurons, suggesting that they may have a complex interaction with other subtypes of SC neurons to modulate saccade behavior.

Taken together, this study provides a comprehensive analysis of the various functional cell types of SC neurons and their role in invigorating saccadic responses to high-valued visual objects in the environment. These results illustrate a circuit-level mechanism by which a subcortical sensorimotor hub integrates value information into the visuomotor planning to facilitate value-guided behavior.

Limitations of the study

Despite the strengths of the dataset and the multimodal analysis framework, this study has a few limitations that should be acknowledged. This study design is purely correlational, and the proposed mechanism of enhancing saccade vigor by converting value signals into motor commands via local circuitry in SC, even if supported by previous literature, must be causally tested using techniques such as optogenetics/chemogenetics. The use of linear arrays provided an unbiased sampling of SC neurons across layers, but classifying them into four broad functional categories, as done here, limited the scope for understanding the functional diversity of SC cell types. A better approach would have been using unsupervised clustering algorithms to classify neurons, which would have strengthened classification and may have revealed additional subtypes. This approach, even though technically challenging, would have provided an opportunity to compare SC functional subtypes from NHPs to other models.¹¹⁸ Thirdly, the recorded activity in the superior colliculus (SC) is influenced by several factors. These include the perceived value of visual objects, the level of attention directed toward high versus low-value objects, the motivation to acquire these objects, and the kinematics of saccades. These factors are closely intertwined, especially in the context of the visually guided saccade task employed in this study. Although the GLM models and kinematic matching procedures helped to dissociate the impact of saccade kinematics, the simple task design used here made it difficult to fully the dissociate value-related modulation of SC activity from other cognitive factors, such as attention and motivation. More complex task designs are needed to disambiguate these effects.

Resource availability

Lead contact

Further information and requests for resources and reagents should be directed to and will be fulfilled by the lead contact, Dr. Atul Gopal, Ph.D. (pagopal@ucdavis.edu).

Materials availability

This study did not generate new, unique reagents.

Data and code availability

• •All data reported in this article are available from the lead contact upon request.
• •This article does not report original code.
• •Any additional information required to reanalyze the data reported in this article is available from the lead contact upon request.

Acknowledgments

We thank Hikosaka lab members for their critical, in-depth discussion of these results. We also thank A.M. Nichols, D. Yochelson, G. Tansey, D. Parker, I. Bruna, A. Lopez, and H. Warnock for their technical assistance. We acknowledge Dr. Leor Katz’s help in setting up the spike-sorting pipeline used in this study. We also thank Dr. Richard Krauzlis for the support provided while performing experiments and drafting this article. This research was supported by the Intramural Research Program of the National Institutes of Health (NIH), United States (1ZIAEY000415).

Author contributions

A.G. and O.H. designed these experiments and performed surgeries. A.G. performed animal training and neural recordings. A.G. also performed data analysis and prepared the figures. A.G. and O.H. drafted the article. A.G. edited and revised the article. All authors have read and approved the final version of the article.

Declaration of interests

No conflicts of interest, financial or otherwise, are declared by the authors.

Published: December 29, 2025

STAR★Methods

Key resources table

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Experimental models: Organisms/strains
Macaque Mulatta	NEI primate facility	–
Software and algorithms
MATLAB	Mathworks	https://www.mathworks.com/products/matlab.html; RRID:SCR_001622
OmniPlex	Plexon, Inc	https://plexon.com/software-downloads/; RRID:SCR_014803
Kilosort2.5	Cortex Lab	https://github.com/cortex-lab/KiloSort; RRID:SCR_016422
Other
BLIP	Okihide Hikosaka Lab	http://www.robilis.com/blip/
Eyelink	SR Research	RRID:SCR_009602
Plexon V-probe	Plexon, Inc	RRID:SCR_018784
Behavioral Data	This Paper	Available from the lead contact upon request
Neurophysiology Data	This Paper	Available from the lead contact upon request
Analysis Codes	This Paper	Available from the lead contact upon request

Experimental model and study participant details

Participants

Two non-human primates (Macaca Mulatta) males with weights of 9kgs (DWN) and 13kgs (BLY) were used in this experiment. All animal care and experimental procedures were approved by the National Eye Institute Animal Care and Use Committee (ASP:NEI-622) and followed the Public Health Service Policy on the Humane Care and Use of Laboratory Animals.

Method details

Surgery

Survival surgeries were performed under sterile conditions and isoflurane gas anesthesia. A head holder was implanted at the stereotaxic zero to restrict the head movements of the animal. A rectangular recording chamber was implanted posterior to the head post at an angle of 40 degrees off vertical (Figure 1C). This orientation of the recording chamber ensured that the receptive fields (RFs) of the different neurons recorded simultaneously from the different layers of the SC mostly overlapped.

Recording setup

The monkeys performed a visually guided saccade task (Figure 1A) that was controlled by a custom real-time software, BLIP (www.simonhong.com). The visual stimuli (Figure 1B) used in the task were fractal objects⁶⁷ (average size ∼7° X 7°, range 5-10°) with several randomly determined features (amplitude, colors, and shapes) created uniquely for these monkeys. We used four distinct fractal sets, each consisting of 8 objects, for each monkey. The objects were randomly assigned to high and low-valued groups. This was done to avoid differences in the physical properties of the stimuli systematically affecting our results. Eye positions were tracked at 1 kHz using an infrared eye-tracking system (EyeLink1000, SR Research).

Behavioral task and training

The monkeys were trained to perform visually guided saccades (Figure 1A) to single objects appearing in the RF of the neurons or in a position 180-degree opposite (AF) to the RF. These single-object trials were interleaved with two object choice trials, that occurred in pseudo-random order within a block. A single block consisted of 128 trials with equal valued choices, unequal valued choices, and single-object trials in the RF and AF. The monkey performed 2-3 blocks of trials in a recording session. This paper focuses exclusively on the behavior and neuronal data obtained during single-object saccade trials presented in the RF of the neurons.

A single stimulus set is composed of eight fractal objects, with four fractals arbitrarily designated as high objects (associated with large reward volumes) and four as low objects (associated with small reward volumes). Juice volumes were controlled by keeping the solenoid valves open for 250ms (high objects) or for 50ms (low objects). The task began with a fixation spot (green-colored square) appearing at the center of the screen. The monkeys maintained their gaze at the central fixation spot for 300ms, after which visual targets appeared at the RF or AF locations. The animals were trained to direct their gaze to the target and hold it there for 500ms, at which point an audio cue and juice reward delivery occurred. Diluted juice was used as positive reinforcement. The trial was aborted if the animals broke fixation during the pre-target (300ms) or post-saccade (500ms) interval. The same conditions were repeated immediately following an aborted trial to ensure a comparable number of correct trials in each block.

The monkeys were trained on each fractal set for more than ten days prior to the neural recording. This consistent long-term exposure resulted in the formation of long-term, stable value memory that boosts the animals' performance. The monkeys correctly chose the high valued objects more than 98% of the time when high and low objects were presented simultaneously, suggesting that they could readily associate value with a large number (8 objects ∗4 sets = 32) of objects.

Electrophysiology

We used a 24-channel, linear electrode array (V probe – Plexon Inc), with inter-electrode distances (50 microns/100 microns), to record the neural activity from different layers of the SC simultaneously. The signals from the 24 channels of the V-probe were filtered, amplified, and digitized using a 64-channel OMNIPLEX-D data acquisition system (Plexon Inc) at a sampling frequency of 40 kHz. Electrodes were inserted into the brain through a stainless-steel guide tube and advanced using an oil-driven micromanipulator (MO-97A, Narishige Int’l. USA). Recording sites were targeted using a grid system that enabled electrode placement at 1-mm intervals in x and y directions, orthogonal to the guide tube.

General experimental procedures

In a typical experimental session, we chose a grid location that targets the SC based on MRI images (Figure 1C), and a 24-channel v-probe was then lowered into the brain using the manual hydraulic manipulator. We encountered SC neurons, typically at a depth of 40mm from the bottom of the grid, characterized by visual neurons with a low-latency phasic response when a target was presented at a specific location in the visual field. After identifying the dorsal-most layer of SC, the probe was lowered further until all 24 channels on it were in contact with different layers of SC. A v-probe with a 100-micron inter-electrode distance spans 2400 microns of SC and can typically record multiple functional subtypes of neurons across the different layers simultaneously (Figure 1C).

We used a passive viewing task to map the neurons' RF. The monkeys were trained to fixate at the center of the screen while objects were flashed at different visual field locations. The stimuli were presented along eight radial angles, starting at zero (horizontal) at 45-degree increments and with amplitudes ranging from 5 to 25 degrees at 5-degree increments. The screen location that elicited high firing in the most number of channels along the dorsoventral axis of SC was identified as the RF and used in the subsequent behavioral tasks. This approach ensured that most of the neurons we recorded had some task-specific responses, even if the targets were not placed in their ideal fields due to the inability to finely tune and identify RFs for each cell recorded on the probe. The broad RFs identified in all 33 recording sessions from the two monkeys are depicted on the SC map (Figure 1E), which shows that we have recorded neurons situated in the middle and caudal portions of the colliculus.

Quantification and statistical analysis

Dataset

This dataset contains 33 experimental sessions, collected from two monkeys: DWN (22 sessions) and BLY (10 sessions).

Pre-processing behavioral data

We considered an object fixated when the eye position was stationary and within 8° of the center of the object based on horizontal and vertical eye position traces. Saccade onsets were defined as the time point when the instantaneous velocity crossed 30 degrees/s, and the end of the saccade was marked when the velocity fell below this threshold. Saccades interrupted by blinks were omitted by using a velocity threshold of 700 degrees/s. Different saccade parameters, such as amplitudes, peak velocity(PV), reaction time (RT), etc., were computed for further analysis.

Targets were placed at different screen positions with amplitudes ranging from 3 to 25 degrees on either the left or right visual fields, based on the RFs of the neurons that were being recorded in each session. Hence, the saccades had to be normalized to compare the velocities and amplitudes across sessions, which was accomplished by normalizing saccade durations.¹²³ Normalization involved re-estimating the average saccade displacement at every 5% of the saccade duration, such that each session had 20 data points going from 0 to 100%, irrespective of the number of data points in the original duration. A polynomial interpolation was used to estimate the saccade displacements at each of the 20 data points, and the same function was differentiated to calculate the angular velocities at the same points. This normalization was carried out separately for high-valued and low-valued object trials.

Pre-processing neural data

The neural data acquired through the Plexon system were processed using the MATLAB-based package KILOSORT¹²⁴ to perform automated spike sorting on the signals acquired from 24 independent V-probe channels. We manually curated the auto-detected units using PHY (https://github.com/cortex-lab/phy), a software that allows visualization of identified units, merging similar units into one, or splitting the activity of one identified unit into different units. A total of 854 units were identified and manually verified in this manner from 33 sessions recorded from two different animals.

Spikes of well-isolated neural units identified by the KILOSORT algorithm were used for all analyses. Spike density functions (SDF) were generated using Gaussian kernels with a bandwidth of 10ms after aligning the activity on the target onset or saccade onset. Neurons were included in the analysis if they had at least 5 valid trials in each task condition and fired more than 20 spikes during the analysis time window.

Data analysis

Functional classification of SC neurons

The activity of individual SC neurons evoked by single objects in the RF was used to classify cells into functional categories. Windows corresponding to baseline activity (-200 to 0ms), Phasic activity #1 (0 to 80ms), and Phasic activity #2 (80 to 160 ms) were defined relative to target onset (Figure 1D left). Visual activation was defined as a significant increase in the phasic#1 and phasic#2 activities compared to baseline. Two windows corresponding to pre-saccadic intervals (-100 to -50ms ) and (-40ms to 0ms) were defined relative to saccade onset (Figure 1D right). Motor activity was defined as a significant increase in the firing rates in the later pre-saccadic interval compared to the earlier. A Kruskal-Wallis test was used to assess statistical significance between the firing rates in each window.

Visual neurons showed significantly higher activation during phasic#1 and phasic#2 compared to the baseline activity when aligned to the target onset. In addition, when aligned on saccades, there was no significant difference in activity between the two pre-saccadic intervals. Cells that fulfilled these criteria were defined as visual neurons.

Motor neurons are cells that show a significant increase in activity in the later pre-saccadic interval compared to the earlier pre-saccadic interval. In addition, when aligned on target onset, the phasic#1 activity was significantly lower than the phasic#2 activity. Cells that fulfilled both these conditions were defined as motor cells.

Visuomotor cells had significant activation during the phasic#1 and phasic#2 epochs relative to the baseline period when aligned on target onset. In addition, when aligned on saccades, there was a significant increase in the later pre-saccadic interval compared to the earlier. Cells that fulfilled both these criteria were defined as visuomotor neurons.

Tonic neurons are defined as cells that showed significant activity reduction during the phasic#1 or phasic#2 epochs compared to the baseline. In contrast to visual and motor cells, tonic neurons have high baseline firing rates that are suppressed by target onset in the RF.

520 units out of the total 854 identified units were classified into four functional subtypes using the above criteria. The rest 334 units were not classified and excluded from subsequent analysis because 1) the cells did not respond to the task in the manner specified above, 2) the cells responded in less than 5 trials, or did not fire more than 20 spikes during the task (n=170) 3) the cells were task modulated but did not meet the statistical criteria used for classification in this study (n=164).

Value modulation

We studied the effect of value in each functional subtype of SC neurons separately. To visualize this modulation, SDFs were generated separately for trials in which high and low objects were presented in the neurons' RF. To quantify neuronal modulation by object value, we measured each neuron's response to 4 different high objects by summing the number of spikes within the window 80-160ms after object presentation. We then compared it to the response of the same neuron in the same window when four low objects were presented in the RF of the neuron. The statistical significance of this activity modulation at the level of single neurons was assessed using the Wilcoxon signed-rank test. Additionally, AUROC values were computed using the neural activity in the time interval 80-160ms after target onset to parametrize the value modulation into a uniform scale ranging from 0-1.

We also determined the time course of value modulation. To determine when the differences between high and low objects became significant, we initially computed the difference between the firing activities of these two conditions (SDF_High-SDF_Low). We then computed the mean and the variability of this difference during the baseline period (-300 to 0ms). These measures were used to identify a threshold (mean + 3 SD) for each neuron separately. The time when the difference between the activity profiles of high and low objects crossed this threshold and remained above it continuously for 20ms was identified. This time point was considered the onset of value modulation for that particular neuron. Using this approach, the onset of value modulation was determined for each cell separately, and the mean onset times were compared between the different subtypes.

We also applied the same onset detection procedure on the averaged population responses of each neuronal subtype. Compared to the mean onset times determined from individual neurons, the onsets detected from the population responses were approximately 10 ms earlier.

Behavioral modulation

We also quantified the contribution of the SC neurons to the observed saccadic behavior by computing Pearson’s correlation coefficients to assess the linear relationship between saccade behavioral parameters and the firing activity of well-isolated SC neurons. The observed neural activity was aligned on target onset, and the average firing rate in the time window 80 to 160 ms was calculated for each trial. Only trials with at least 5 spikes were included in this analysis. The Pearson correlation coefficient was computed between the average firing rate in each trial and the corresponding behavioral measure (RT or PV). Only those neurons with at least 15 valid trials are included in this analysis. The calculated correlation was used as an index to quantify the sensitivity of SC neurons to the upcoming saccade behavior. We computed two metrics for each identified neuron: the Reaction Time Index (firing rate vs. RT) and the Peak Velocity Index (firing rate vs. PV). The correlation was considered significant if the p-value was less than or equal to 0.05.

Addressing collinearity of measured variables

Previous studies have shown that firing rates in the colliculi are correlated with the RT and PV of the upcoming saccade. In addition, high-valued objects elicit saccades with greater vigor, resulting in colinear dependencies between the parameters of interest. Hence, we employed multiple approaches to test whether the observed modulation in the firing rates of SC neurons had a value component independent of saccade vigor.

Vigor matching

To ensure that differences in observed neural activity were not confounded by saccadic vigor, we matched the distributions of RT and PV across high and low-valued trials within each session. For RT and PV separately, all valid trials were sorted, and differences between value conditions were quantified using the standardized mean difference (SMD). If the SMD exceeded 0.2 (Cohen’s small effect threshold), we iteratively removed extreme trials from the condition with more available data until the distributions converged. Sessions requiring excessive removal (>40% of trials from either condition) were excluded. After matching, RT and PV distributions did not differ across value conditions (tested with t test and KS test), ensuring that subsequent neural comparisons reflected value-related modulation rather than differences in saccade vigor. The vigor matching reduced the number of valid neurons in each SC subtype by ∼20%.

GLM models

To quantify the unique contributions of value and saccade vigor-related parameters to firing-rate variability at the population level, we constructed a pooled generalized linear model (GLM) using data from all neurons across sessions. For each trial, the number of spikes within the time epoch 80 to 170ms, and corresponding behavioral predictors—object value (coded as +1 for high value and -1 for low value), RT, PV, and saccade amplitude—were extracted. All variables, including the firing rate, were z-scored within each neuron. The GLM assumed a Gaussian error model and took the form:

$$FR={\beta }_0+({\beta }_{value}\ast V)+({\beta }_{RT}\ast RT)+({\beta }_{PV}\ast PV)+({\beta }_{amp}\ast AMP)+\epsilon$$

Model performance was evaluated using 10-fold cross-validation with non-overlapping trial partitions, and R-square was computed on the held-out predictions. The statistical significance of each coefficient was determined from the model’s t-statistics. To test whether the value coefficient (β_value) reflected true value coding rather than spurious correlations, a control analysis was implemented in which value labels were systematically manipulated: 1) 50% shuffle, where value labels were randomly reassigned resulting in ∼50% of trials to change labels (leaving the remaining 50% unchanged), thereby degrading the value structure while preserving overall trial composition. 2) 100% flip – where all value labels were inverted (high ↔ low), thereby fully reversing the value structure. The separation between the real and control β_Value confirmed that the value term captured a genuine, non-random component of the firing-rate variance that could not be explained by RT, PV, or amplitude.

Bootstrapping procedure

The experiment performed in this study allowed us to directly assess the effect of value on the response of different types of SC neurons. It is unclear whether this modulation affected visual or presaccadic motor processes encoded in SC neurons. To disentangle the visual and motor stages in the neural responses of these neurons, as done previously.^57,58 we employed a hierarchical bootstrapping procedure (Figure S2) on the data. The following steps were used to bootstrap the experimental data separately for each functional subtype of SC neurons.

• 1)The trials in each experimental session were divided into three quantiles (Figure S2A) based on the saccadic RT. The firing rates associated with each neuron’s activity were likewise divided into three quantiles (Figure S2B) based on the same behavioral quantiles.
• 2)To simulate the neural activity for a hypothetical trial belonging to one of the quantiles (single bootstrap replicate), we first created a pseudo population of 25 neurons. For creating one of these 25 neurons, we first resampled the 33 recording sessions with replacement. Then, within the selected session, we randomly selected one of its constituent neurons, and within the selected neuron, we randomly chose one trial from the relevant behavioral quantile. The neural responses from these 25 neurons were pooled together to generate the SDF for the hypothetical trial. (Figure S2C). This approach ensured that the nested structure of the dataset (session-neurons-trials) was preserved in the bootstrapped samples as well.
• 3)The firing rate and the time to peak in multiple time epochs were calculated for each bootstrapped trial. 10–35 such trials were simulated (Figure S2D) for each behavioral quantile. We then compared these metrics between the three RT quantiles and computed an F-statistic. These steps constitute a single repeat of the bootstrapping process

This bootstrapping procedure was repeated 10000 times. The mean time to peak and its 95% confidence intervals were identified and tabulated in the supplementary tables. This bootstrapping process was repeated with their RT labels shuffled to create a null distribution. The F-statistic, computed from the real data, was then compared to this null distribution to obtain the p-value for the relevant comparison. The p-value obtained from this bootstrap-based permutation test is reported in the text and tabulated in the Table S2.

The bootstrapping procedure was performed separately for high- and low-object trials, and the mean firing rates, along with 95% confidence intervals, were calculated for the three time epochs. The firing rates between high and low conditions are compared in each epoch, and the p-value for these comparisons was estimated directly from the confidence intervals and tabulated in the Table S3.

Statistical analysis

For population analysis, firing rates, quantified separately for each neuron in each condition, were pooled, and a paired t test was used to compare high and low object responses. Error bars in all plots show the standard error of the mean (SEM) unless otherwise noted. For the analysis involving bootstrapped samples, a Bootstrap-based permutation test was used to determine the p-values. While comparing value modulation (high vs low) in the bootstrapped samples across the different phases and RT quantile (short, medium, and long), p-values were calculated directly from the bootstrapped confidence intervals. Significance thresholds for all tests in this study were α=0.05. The P values are denoted in figures using asterisks. Triple asterisks (∗∗∗) denote p < 0.001, double asterisks (∗∗) denote p < 0.01, single asterisk (∗) denotes p < 0.05, and ns denotes p ≥ 0.05.