All behavioral and neurophysiological data were preprocessed with MATLAB 2022b (MathWorks).

In the choice task, saccade onsets toward good and bad objects were defined as the eye speed exceeding 40°/s within 400 ms of the target onset. During the fixation task, saccades to the presented objects were detected offline. They were considered fixation break errors.

Welch's t tests were performed for each scene condition to compare the saccade reaction times for the good and bad objects (Fig. 2A). During this analysis, only the reaction times of the saccades toward bad objects (return go) were included, and those for returning from the bad target (return back) were not included. To determine whether the proportion of monkeys choosing to stay for a bad object was higher for Scenes 1 and 2, where the object values remained constant, Fisher's exact test was conducted (Fig. 2B).

Behavioral results. A, Raincloud plots of saccade reaction times for good or bad objects in Scenes 1–4 during the choice task of the two monkeys, C and S. In each panel, the top “cloud” illustrates the probability distribution of reaction times, and the bottom “rain” shows the raw plots of individual reaction times. In all the scenes, the reaction times of the two monkeys toward good objects are significantly shorter than those toward bad objects (p = 0 for all scenes for both monkeys, Welch's t tests). B, Proportions of selected actions, Accept, Return, Stay, Other, or fixation break error (Fix break) for bad objects in Scenes 1–4 during the choice task. FP, fixation point; ITI, inter-trial interval; Obj, object; Tgt, target.

The sample size was not determined by statistical methods. Instead, the number of recorded neurons was determined based on a previous study (Watanabe and Munoz, 2010), which recorded the activities of striatal neurons of macaques while they performed an antisaccade task. All statistical analyses were performed with the preprocessed data using the R software (version 4.2.2).

We employed the k-means clustering method to categorize neurons based on their activities when good or bad contralateral objects were presented. This process involved standardizing the activity of each neuron using a Z-transform. The average neuronal activity in response to the presentation of either a good or a bad object in the contralateral direction was computed. This calculation had a 200 ms window, extending from 100 to 300 ms after the onset of object presentation. To quantitatively determine the optimal number of clusters (K) for k-means classification, we simulated the silhouette values 5,000 times and examined which K (up to K = 6) produced the largest average silhouette value. We then performed a one-way analysis of variance (ANOVA) and conducted post hoc pairwise t tests with Bonferroni’s corrections to confirm the optimal K value (Fig. 3A,B). Statistical significance for the post hoc test was set at α = 0.05/10 with Bonferroni’s correction.

Scatterplots of clustering and recording sites in the striatum. A, Result of simulation of silhouette analysis to determine the optional K for clustering. The y-axis represents the averaged silhouette number for each of the 5,000 silhouette analysis ran with K ranging from 2 to 6, and the error bars represent the standard deviation. B, Two-dimensional scatter plots of the Z-transformed mean firing rate for individual neurons during 200 ms after 100 ms at the onsets of presentation of the good (x-axis) and bad (y-axis) objects. Using these two-dimensional data, the neurons are divided into three groups using clustering methods. C, Reconstruction of recording site of task-related striatal neurons of the two monkeys. Cd, caudate nucleus; Put, putamen.

The data from each neuron was aligned with the initiation of events (scene, target, and saccade onset). The time course of the neuronal response to event onset for each condition was investigated by calculating peristimulus time histograms (PSTHs) in 1 ms bins. It was smoothed with a spike density function using a Gaussian filter (σ = 20 ms). To visualize the neuronal activity of individual neurons on the color map, the Z-transformation of the activity in each neuron was performed to visualize the activities of individual neurons (Adler, et al., 2012; Kaplan et al., 2020). First, the firing rate of the baseline was calculated by averaging the firing rate within 500 ms before the scene, target, or saccade onset. Then, this baseline was subtracted from the smoothed PSTH and aligned with the initiation of events. Subsequently, a Z-score transformation was conducted for each PSTH; the baseline mean was subtracted from it, and the difference was divided by the standard deviation (SD).

To prevent false positives, generalized linear mixed-effects models (GLMMs) were used in this study to compare neuronal activities across conditions (Yu et al., 2022). First, we compared the full model containing explanatory variables as fixed effects for all statistical tests using the GLMM. All monkey, neuron, and experimental session IDs were included as random effects, which constituted the null model. A parametric bootstrap method was conducted to assess the goodness of fit of the models by performing 10,000 iterations and computing the p value based on the difference in deviance between the two models. If the full model showed a significant fit, we conducted post hoc pairwise t tests with Bonferroni’s corrections to further explore these differences. For GLMM, we used the lme4 (Bates et al., 2014), pbkrtest (Halekoh and Højsgaard, 2014), and emmeans (Lenth et al., 2019) packages in RStudio.

Figure 5, B, G, and K, shows the comparison of neuronal activity at scene onset in the choice task. A GLMM was employed to evaluate differences in neuronal activity at scene onset. The full and null models used for the comparison are as follows:FullModel:NeuronalActivity∼Scene+(1|monkey_ID)+(1|monkey_ID:Neuron_ID),NullModel:NeuronalActivity∼(1|monkey_ID)+(1|monkey_ID:Neuron_ID), where NeuronalActivity is the mean PSTH of individual neurons during a 200 ms period, ranging from 100 to 300 ms after scene onset; Scene (1–4) is the fixed effect; and monkey_ID and Neuron_ID are the random effects.

The level of statistical significance for the model comparison was set at α = 0.05.

Figure 5, D, H, and L, shows the comparison of neuronal activities at target onset for the four scenes during the choice task. A GLMM was conducted to examine the differences in neuronal activity at the target onset. The full and null models used for the comparison are as follows:FullModel:NeuronalActivity∼Scene×Value×Direction+(1|monkey_ID)+(1|monkey_ID:Neuron_ID),NullModel:NeuronalActivity∼(1|monkey_ID)+(1|monkey_ID:Neuron_ID), where NeuronalActivity is the mean PSTH of individual neurons within 200 ms from 100 ms after the target onset; Scene (1–4), Value (good vs bad), and Direction (contralateral vs ipsilateral) are the fixed effects; and monkey_ID and Neuron_ID are the random effects.

Statistical significance for the model comparison was set at α = 0.05. We performed six pairwise t tests to compare the mean neuronal activities for the good and bad objects during the choice task for each scene and six pairwise t tests for comparison across scenes. Therefore, statistical significance for the post hoc test was set at α = 0.05/12 with Bonferroni’s correction.

A sliding time window approach was employed to determine when the neuronal responses to good and bad objects began to diverge significantly in Clusters 1–3 during the choice task. A window size of 50 ms was used, and the window was advanced in 1 ms increments from 575 ms before to 575 ms after the object onset. For each window, the normalized mean neuronal activity toward good and bad objects of individual neurons was calculated, and a paired-sample t test was performed between the two conditions. The significance level was set at α = 0.05. A significant divergence in neuronal activity was defined as occurring when the p value remained below 0.05 for at least 10 consecutive windows. This criterion was applied to reduce the likelihood of false-positive errors. Finally, the time points of significant differences in neural activity were visualized as magenta-colored markers in the panel for each condition, with the first time points that reached significance for 10 consecutive windows shown as vertical dotted lines (Fig. 5C,F,J).

Figure 6, B, E, and H, shows the comparison of neuronal activity at saccade onset for Scene 1 during the choice task. Given that comparisons of the mean neuronal activity during the postobject onset period showed no significant differences for the four scene conditions, analysis of the perisaccade onset period was only performed for the Scene 1 condition. We employed a GLMM to investigate differences in neuronal activity during saccade onset. The full and null models used for the comparison are as follows:FullModel:NeuronalActivity∼Value×Direction+(1|monkey_ID)+(1|monkey_ID:Neuron_ID),NullModel:NeuronalActivity∼(1|monkey_ID)+(1|monkey_ID:Neuron_ID), where NeuronalActivity refers to the mean PSTH of individual neurons during 200 ms beginning 100 ms after saccade onset in Scene 1; Value (good vs bad) and Direction (contralateral vs ipsilateral) are the fixed effects; and monkey_ID and Neuron_ID are the random effects. The level of statistical significance for the model comparison was set at α = 0.05. We performed six pairwise t tests for comparison of mean neuronal activity for good and bad objects during the choice task, and statistical significance for the post hoc test was set at α = 0.05/6 with Bonferroni’s correction.

For further analysis, a series of neuronal data analyses were conducted to ascertain the involvement of the recorded neuronal activity in facilitating saccades (Fig. 7). Initially, the data recorded from each neuron were segmented into four groups based on the saccade reaction times in each trial under the four conditions (contralateral/ipsilateral vs good/bad object). These groups were arranged in ascending order of saccade latency, with Group 1 having the shortest latency and the subsequent groups having progressively longer latencies. The average neuronal activity of each group was calculated. This calculation had a 200 ms window, from 100 to 300 ms following object presentation, and the resulting average neuronal activity was standardized using a Z-transformation. The correlation coefficients for individual neurons was computed to explore the relationship between saccade latency and neuronal activity. The coefficients were designed to assess the correlation between the group order (from Groups 1 to 4) and mean neuronal activity. Finally, Wilcoxon signed-rank tests were used to determine whether the median of the correlation coefficients differed significantly from zero across the datasets to establish the statistical significance of our findings.

Figure 8, E, H, K, N, Q, and T, shows the comparison of neuronal activities at target onset while monkeys rejected bad objects by suppressing the saccade to a presented bad object (stay) or returning the gaze to the original center point without fixation for >400 ms after the saccade to the target (return) during the choice task and at object onset during the fixation task. By comparing the neural activities for these two cases, we determined whether the observed neural activity was related to the inhibition of the saccade or the rejection of the bad object. It was also determined whether the observed neural activity was related to proactive or reactive inhibition by comparing it with the neural activity during the fixation task. During the fixation task, the same objects were used (their values were constant during Scene 1 of the choice task). We compared the average neural activity for Scene 1 during the choice task when subjects rejected bad objects by return and stay and the neural activity when the objects were presented during the fixation task using GLMM. A GLMM was conducted to test the differences in neuronal activity for the target or saccade onset. The full and null models for the comparison are as follows:FullModel:NeuronalActivity∼Scene×Value×Direction+(1|monkey_ID)+(1|monkey_ID:Neuron_ID),NullModel:NeuronalActivity∼(1|monkey_ID)+(1|monkey_ID:Neuron_ID), where NeuronalActivity is the mean PSTH of individual neurons during a 200 ms period, beginning 100 ms after target onset; Scene (1–4), Value (good vs bad), and Direction (contralateral vs ipsilateral) are the fixed effects; and monkey_ID and Neuron_ID are the random effects.

The statistical significance for the model comparison was set at α = 0.05. Considering that six pairwise t tests were performed to compare the normalized neuronal activities for the good and bad objects during the choice task, statistical significance for the post hoc test was set at α = 0.05/6 with Bonferroni’s correction.