Prefrontal spatial working memory network predicts animal's decision making in a free choice saccade task

Abstract

While neurons in the lateral prefrontal cortex (PFC) encode spatial information during the performance of working memory tasks, they are also known to participate in subjective behavior such as spatial attention and action selection. In the present study, we analyzed the activity of primate PFC neurons during the performance of a free choice memory-guided saccade task in which the monkeys needed to choose a saccade direction by themselves. In trials when the receptive field location was subsequently chosen by the animal, PFC neurons with spatially selective visual response started to show greater activation before cue onset. This result suggests that the fluctuation of firing before cue presentation prematurely biased the representation of a certain spatial location and eventually encouraged the subsequent choice of that location. In addition, modulation of the activity by the animal's choice was observed only in neurons with high sustainability of activation and was also dependent on the spatial configuration of the visual cues. These findings were consistent with known characteristics of PFC neurons in information maintenance in spatial working memory function. These results suggest that precue fluctuation of spatial representation was shared and enhanced through the working memory network in the PFC and could finally influence the animal's free choice of saccade direction. The present study revealed that the PFC plays an important role in decision making in a free choice condition and that the dynamics of decision making are constrained by the network architecture embedded in this cortical area.

neurophysiological investigations of the prefrontal cortex (PFC) have shown that neurons in the lateral PFC exhibit a persistent activation during the delay period (delay-period activity) when the monkey is remembering a particular spatial location in memory-guided saccade tasks (Constantinidis et al. 2001a, 2001b; Funahashi et al. 1989, 1990, 1991, 1993b; Goldman-Rakic et al. 1990). This delay-period activity has been proposed to be a neuronal correlate of active maintenance of visuospatial information. The role of the PFC in information maintenance has been further supported by lesion studies in monkeys (Funahashi et al. 1993a; Sawaguchi and Iba 2001) and humans (D'Esposito and Postle 1999; Mottaghy et al. 2002; Müller et al. 2002) and by functional brain imaging studies (Courtney et al. 1998; Sakai et al. 2002; Zarahn et al. 1999, 2000). Thus the maintenance of task-relevant spatial information could be one of the key features that could help us to understand the function of the PFC (Fuster 2008).

Spatially selective activity of PFC neurons has also been proposed to be related to decision-making processes such as response selection (Rowe et al. 2000) and spatial attention (Lebedev et al. 2004; Messinger et al. 2009). Especially in human neuroimaging studies, the PFC has been reported to play a role in self-initiated behavior and internally driven decision making (Frith et al. 1991; Haynes et al. 2007; Hyder et al. 1997; Lau et al. 2004a, 2004b; Soon et al. 2008). Therefore, the known characteristic activation of spatially selective PFC neurons could also be related to decision making under these situations. Previous studies have investigated the neuronal underpinning of decision making under these situations in the PFC (Watanabe et al. 2006; Watanabe and Funahashi 2007) and other related areas such as frontal eye field, supplemental eye field, and lateral intraparietal cortex (Coe et al. 2002). These studies reported that the activity of spatially selective neurons is related to the animal's own decision about saccade direction. However, in these studies, a fixed set of spatial locations were repeatedly presented as options for a saccade in a block of trials. Under this setup, presentation of the spatial cues was less informative, since the available saccade directions were obvious without seeing the actual cue presentation. The monkey could indeed decide the saccade direction before the presentation of the visual cues. In addition, because of the predetermined and less various configuration of the cues, precise analysis about the relationship between the neuron's receptive field and chosen location was limited. Therefore, further experimentation that enables a close examination of how prefrontal working memory network represents multiple spatial information and how the activity of PFC neurons is related to the animal's own decision about saccade direction is needed.

In the present study, we established a free choice memory-guided saccade task in which the monkeys chose by themselves the direction of the saccade among multiple locations changing from trial to trial. By varying the options for a choice in each trial, we could examine the precise time course of the decision-making process taken in the network of spatially selective PFC neurons. We found that the activity of lateral PFC neurons could predict the animal's decision about subsequent eye movement direction even before cue presentation. This suggests that preexisting activation state of PFC neurons immediately before cue presentation influenced the construction of spatial representation and eventually biased the animal's subsequent choice of the saccade direction. Furthermore, we found that the impact of neuron's activity on the animal's choice was stronger in neurons that showed greater persistent activity in a control spatial working memory task. In addition, although PFC neurons tended to represent unchosen spatial location more weakly from the beginning of the trial, this suppression of unchosen location was modest when chosen and unchosen locations were placed in the same side of the visual field. This finding was in accord with the known contralateral organization of spatial working memory network in the PFC. These results indicate that the role of PFC neurons in a free choice of saccade direction is linked to their firing property and network background as a neuronal underpinning of spatial working memory function. Our present study showed a possible overlap of cellular mechanisms for maintenance and decision making of spatial information, and offers a clue to a further investigation on the nature of the spatial information processing taken place in the PFC.

MATERIALS AND METHODS

Animals

We used two female Japanese monkeys (Macaca fuscata; monkeys O and E). The monkeys were housed in individual stainless steel home cages. Water intake was restricted in the home cage and provided as a reward in the laboratory. Additional vegetables and fruits were provided to fulfill the daily requirement of water intake if necessary. All experimental procedures were conducted in accordance with the guidelines provided by the Primate Research Institute of Kyoto University and were approved by the Animal Research Committee at the Graduate School of Human and Environmental Studies, Kyoto University.

Apparatus

During experimental sessions, the monkey sat in a primate chair in a dark sound-attenuated room with its head movements restricted by a head-holding apparatus. We used TEMPO software (Reflective Computing, Olympia, WA) for task control and data acquisition. Visual stimuli were presented on a 20-in. CRT monitor (Dell UltraScan D2026T-HS; Dell, Round Rock, TX) that was placed 40 cm from the subject's face. A scleral search coil system (Enzanshi Kogyo, Tokyo, Japan) was used to monitor the monkeys' eye movements (Judge et al. 1980; Robinson 1963).

Tasks

We used two types of memory-guided saccade tasks (Fig. 1A): an instructed choice task (ICT) and a free choice task (FCT). The tasks were similar to those used in previous studies (Mochizuki and Funahashi 2014; Watanabe et al. 2006; Watanabe and Funahashi 2007). In both tasks, a trial started with the presentation of a fixation point (white cross, 0.5° in visual angle) at the center of the monitor. After the monkey maintained fixation on the fixation point for 1.0 s (fixation period), eight peripheral targets (white cross, 0.75°) were presented at an eccentricity of 13° (0°-315°, separated by 45°). The monkey had to neglect these targets and keep watching the fixation point for another 1.0 s (precue period). Next, one or two visual cues (filled white circle, 2.5°) were briefly blinked over the peripheral targets for 0.5 s (cue period). In the ICT, one cue was presented at one of the eight target locations. In the FCT, two identical cues were simultaneously presented at two peripheral locations. After the cues disappeared, the monkey had to maintain fixation for a 1.5- to 3.0-s random length of delay (delay period). At the end of the delay period, the fixation point was turned off, and the monkey was required to make a memory-guided saccade toward the cued location within 0.5 s. The reward was delivered after the monkey maintained fixation on the correct target location for 0.3 s. In FCT trials, a saccade to either of the two locations was regarded as correct. Every correct response was rewarded by a drop of juice, and there was no difference in the amount of reward regardless of the monkey's choice in the FCT or the type of the task.

Fig. 1.

A: task configuration. Schematic illustrates the 2 tasks. In the instructed choice task (ICT), the monkey was required to make a memory-guided saccade toward the cued location. In the free choice task (FCT), the monkey needed to choose 1 of 2 cued locations before making a saccade. B: preference indexes (bars) and response times (lines) of each monkey averaged across 8 bins of absolute directions. Preference index was expected to be 0.25 if the monkey does not exhibit directional preference. C: choice proportions and response times in the FCT for rotated cue locations relative to the receptive field of the neurons (T_in, neuron's preferred direction; T_ipsi and T_contra, perpendicular directions ipsilateral and contralateral to T_in, respectively; T_opp, opposite direction 180° away from T_in). Proportion of T_in choice (left) was calculated for each of the T_in vs. T_ipsi, T_in vs. T_contra, and T_in vs. T_opp pair conditions, along with the total proportion (“all”) calculated by collapsing the pair conditions. Response times (middle) and normalized response times (right) were averaged for each response direction, along with the grand average, by collapsing all the directions.

The location of the cue in ICT trials was randomly determined as one of the eight peripheral target locations. Possible cue locations in FCT trials were limited to four locations (0°, 90°, 180°, and 270°) to reduce the number of combinations of cue locations. In an FCT trial, cues were presented at two of these four possible locations. Accordingly, trials consisted of eight cue conditions in the ICT and six pair conditions in the FCT. ICT and FCT trials were intermingled in random order.

During a recording session, we first presented only ICT trials using the eight possible cue locations as explained above. After we isolated the activity of a single neuron, we examined whether it had a directionally selective task-related activity during performance of the ICT. We collected, on average, 8.3 trials for each of the eight direction conditions for this screening. We then quantitatively analyzed the activity of the neuron during several task epochs: cue (0–500 ms after the onset of the cue), early delay (0–1,000 ms after the start of the delay), late delay (1,000–500 ms before the end of the saccade), early response (300-0 ms before the end of the saccade), and late response (0–300 ms after the end of the saccade). If the neuron exhibited a significantly different firing rate during any of these epochs compared with the baseline period (0–1,000 ms before the onset of the cue, Dunnett's test for multiple comparisons, P < 0.05), we categorized it as a task-related neuron. The neuron was then further tested for directional selectivity. We used a modified circular normal distribution (von Mises distribution) as a tuning curve to evaluate the modulation of the neuron's firing rate across the eight cue conditions:

$$\mathit{\text{f}}(d|\mu ,\beta ,B,R)=B+R\cdot \frac{\exp \left[\beta \cdot \cos \left(d-\mu \right)\right]}{\exp \left(\beta \right)},$$

where the firing rate of a neuron (f) was determined as a function of the direction of the cue (d), based on the baseline (0 ≤ B) and magnification (R) factors for the firing rate and the location (μ) and concentration (0 ≤ β) factors for the von Mises distribution. The peak direction, estimated as the μ parameter by fitting of the tuning curve during the epoch in which the firing rate was highest, was regarded as the neuron's preferred direction. The size of the receptive field was also quantified from the estimated parameter as 1/$\sqrt{\beta }$, which can be regarded as an analog of standard deviation parameter of a normal distribution (σ). If the fitting did not converge, the neuron was considered to lack directional selectivity.

Once the neuron's preferred direction was determined, we rotated the cue locations so that one of the eight possible locations was placed at the neuron's preferred direction. The monkey then performed randomly intermingled ICT and FCT trials using these rotated cue locations. We only used the data recorded in these postscreening trials with rotated cue locations, except for the estimated receptive field of the neurons, which was calculated from the activity during screening ICT trials. The four orthogonal cue locations for the FCT are now referred to as T_in, T_ipsi, T_contra, and T_opp, where T_in is the neuron's preferred direction, T_ipsi and T_contra are the perpendicular directions ipsilateral and contralateral to T_in, respectively, and T_opp is the opposite direction 180° away from T_in. The directions other than T_in (i.e., T_ipsi, T_contra, and T_opp) were also collectively referred to as T_out.

Since the focus of the present study was to determine how the spatial representation in the PFC was involved in the animal's own decision making in choosing saccade directions, we only analyzed FCT trials that included T_in, where the neuron being recorded was responsible to represent, as one of the two cues. Therefore, only three pair conditions (T_in vs. T_ipsi, T_in vs. T_contra, and T_in vs. T_opp) of six possible pair conditions were considered in the present analysis. For the ICT, we only used the data for trials in which the visual cue was presented at one of the four locations that appeared in the FCT. Each neuron's directional selectivity was confirmed by a postrecording offline analysis as a larger firing rate in T_in cue trials than in T_out cue trials in the ICT with rotated cue locations (t-test, P < 0.05). Neurons that did not show higher activation in T_in than in T_out cue trials in the ICT were excluded from further analysis.

Surgery and Training Procedure

We implanted a stainless steel head-holding device and a scleral search coil in the monkeys. A scleral search coil was implanted onto the right eye globe by dissecting the conjunctiva (Judge et al. 1980). The monkeys were first anesthetized by an intramuscular injection of ketamine hydrochloride (10 mg/kg) and then an intravenous injection of pentobarbital sodium (10–15 mg/kg). Heart rate and respiration were monitored during the surgery. Stainless steel screws were put into the skull to ensure firm adhesion of the head-holding device. The connector for the search coil and the head-holding device were fixed to the skull with dental acrylic. All of the surgical procedures were performed under aseptic conditions.

After the monkeys recovered from surgery, we started training of the tasks. We first trained the monkeys with the ICT. When the monkeys learned to perform the ICT (about 85% correct for more than 5 consecutive experimental sessions), we started to intermingle FCT trials with ICT trials.

After we completed the task training, we performed the second surgery to implant a stainless steel cylinder (MO-903E; Narishige, Tokyo, Japan) for the recording of neuronal activity. The monkeys were anesthetized with the same procedure as the first surgery and then fixed to the stereotaxic apparatus. We made a small hole (20 mm in diameter) on the skull with a trephine. The stereotaxic coordination of the center of the hole was set approximately 30.0 mm anterior from the interaural plane and 15.0 mm lateral from the midline, and determined by referring structural magnetic resonance imaging (MRI) pictures of the monkey's brain. We attached the stainless steel cylinder to the hole with stainless steel screws and dental acrylic. All of the surgical procedures were performed under aseptic conditions. After the monkeys recovered from surgery, we started neuronal recordings.

Data Collection

We recorded single-neuron activity from the cortex within and surrounding the principal sulcus. The area of the recording in the lateral PFC was determined on the basis of MRI pictures of the brains. We used glass-coated Elgiloy microelectrodes (0.5–3.0 MΩ at 1 kHz) to record single-neuron activity. An electrode was advanced with a hydraulic microdrive (MO-95; Narishige). Raw neuronal activity was amplified using an amplifier (DAM80; WPI, Sarasota, FL) and monitored on an oscilloscope (SS-7802; IWATSU, Tokyo, Japan) and an audio monitor. During experiments, we isolated single-neuron activity from raw activity using a window discriminator (DIS-1; BAK Electronics, Mount Airy, MD) and monitored the isolated single-neuron activity together with raw activity using an oscilloscope. Single-neuron activity and task events were stored as a data file on a laboratory computer.

Data Analysis

All statistical analyses and data plotting were performed using the statistical software R 3.2.1 (R Core Team 2015). Before testing the difference in central values among groups, we performed Shapiro-Wilk tests to examine normality of the data in each group. We also performed Bartlett's test or the Fligner-Killeen test to examine the homogeneity of variances. Based on results of these tests, we selected a nonparametric test when appropriate. We used Holm's correction method for P values on statistical results taken from a set of multiple comparisons unless otherwise noted.

Behavioral analysis.

The proportion correct was calculated separately for the ICT and FCT by dividing the number of trials with correct target capture by the number of trials in which the animal reached the response period.

To examine the animal's preference toward four directions in the FCT, we defined preference indexes based on the proportion of chosen direction. For a given direction, we calculated the proportion of trials in which that direction was chosen by the animal from the total number of trials in which that direction was available in the FCT. The calculated four proportions were then divided by their sum. We call these normalized proportions of choosing each direction as preference indexes. Preference index was expected to be 0.25 if the animal chose each direction equally in the FCT. Preference indexes were separately calculated for the behavior obtained during recordings of each neuron because the absolute angles of the four directions differed based on the location of the receptive field of neurons.

To compare the behaviors in different recording sessions with different cue configurations, we grouped the absolute directions of responses by eight bins of 45° width. We then averaged the preference indexes categorized into each bin. We used the same bins to calculate averaged response times in the ICT and FCT with different response directions. Response times were measured as the latency from disappearance of the fixation point to the onset of a saccade detected by the method in a previous study (Martinez-Conde et al. 2000). We further tested the animal's task performance in the FCT on the basis of the relative directions from each neuron's receptive fields. For each of the T_in, T_ipsi, T_contra, and T_opp directions, we calculated the mean response times in correct FCT trials. Response times for these relative directions could vary, reflecting the difference in motor execution processes toward different absolute directions. Therefore, for each of the T_in, T_ipsi, T_contra, and T_opp directions, we also calculated the normalized response times by subtracting the mean response time in the ICT from that in the FCT, and then dividing it by the standard deviation of the response times for that direction in the ICT. This tested whether the animal's speed of responses was different among the four response directions in the FCT, canceling out the effect of the difference in motor processes for different absolute saccade directions.

Task-related activity and receiver operating characteristic analysis.

We used a 100-ms time window sliding in 25-ms steps to make perievent time histograms to examine task-related activities of the neurons. Constructed histograms were then averaged across neurons to create population histograms. We also used a receiver operating characteristic (ROC) analysis to compare the strength of neuronal activity between two different trial conditions (Britten et al. 1992; Shadlen and Newsome 1996). For each time window, we constructed an ROC curve and calculated the ROC value (area under the ROC curve) using 100 criterion firing rates. To evaluate the onset of ROC elevation, we repeatedly tested the significance of differences in the ROC values of the neurons from 0.5 (1-sample t-test, α = 0.05). If the ROC values were larger than 0.5 in five consecutive bins, the time of the first bin was regarded as the onset of ROC elevation.

We applied an ROC analysis to the data from both the ICT and FCT. In the ICT, we compared the neuronal firing between T_in cue trials and T_out cue trials. Therefore, the calculated ROC value is an index of traditional memory-related activity that encoded the spatial location of the cue instructed in that trial. In the FCT, we compared the neuronal firing between T_in choice trials and T_out choice trials in each pair condition. Therefore, the calculated ROC value is an index of decision-related activity that encoded the subsequently chosen spatial location from the same set of cues.

Baseline sustainability of firing.

Previous studies have suggested that the dynamics of the spontaneous fluctuation in neural activity reflect the background structural and functional architecture of the network (Kenet et al. 2003; Tsodyks et al. 1999). In the present study, we were particularly interested in the relationship between the persistence of spontaneous activity and the neuron's role in memory and decision-making functions. To quantify the persistence of a neuron's activity at a baseline state, we examined the temporal correlation of firing rates within a trial (Ogawa and Komatsu 2010). We divided the first 800 ms of the precue period (1,000–200 ms before cue onset) into eight successive 100-ms time bins. We calculated the trial-to-trial variation in activity within each bin by subtracting the mean firing rate of the given bin across trials from the firing rate for each trial in the same bin. We then calculated the Pearson's correlation coefficient of these values between two different bins interposed by a given length of interval. Seven intervals (0–600 ms in 100-ms steps) were available depending on the combination of the bins, where “0-ms interval” meant two successive bins and “600-ms interval” meant the longest interval between the first and the last bins of the 800-ms period used in this analysis. Different pairs of bins with the same interval were pooled to calculate a single correlation coefficient for each interval length. Therefore, seven correlation coefficients were calculated, one for each of the interval lengths, for each neuron. We refer to the calculated correlation coefficient as “baseline sustainability,” since it reflects how the activity within a time bin could be sustained until another temporally distant bin.

Serial correlation of the interspike interval.

We also measured the sustainability of the activity of each neuron by calculating a serial correlation of the interspike interval (ISI). In this analysis, we first calculated the ISIs of a neuron by using all of the collected data, including those from nontask epochs such as the intertrial interval. Next, we calculated Pearson's correlation coefficient between the lengths of successive ISIs. Since the ISI is a measure of the momentary level of activation, a stronger serial correlation of ISI indicates that the activation state once achieved by the neuron tended to persist for a while.

Dimensional reduction in population activity.

We used a dimensional reduction technique with a principal component analysis (PCA) to compare the activation patterns of PFC neurons among different task conditions (Briggman et al. 2005; Broome et al. 2006; Churchland et al. 2007; Shenoy et al. 2013). For each neuron, we first calculated the average firing rate in each task condition (4 conditions for the ICT and 6 conditions for the FCT) during ±2,000 ms from the cue onset. We used 50-ms time bins sliding in 25-ms steps to calculate the mean firing rates. We then stacked these averaged firing rates for each neuron and each condition into an M × N matrix, where M is the number of bins in a trial multiplied by 10 (total number of task conditions) and N is the number of neurons under interest. We applied a PCA to this matrix. The first three principal components were used to create the principal component state space. The activation state and its transition were represented as a trajectory inside the state space. To evaluate how the neuronal activation patterns differed between the conditions, we calculated the Euclidean distances between trajectories.

Correlation analysis between tasks.

To investigate how spatial representation was constructed in the network of PFC neurons during an FCT trial, we applied a correlation analysis to the activation patterns of PFC neurons in different tasks. In this analysis, we tested the similarity of the neuronal activation patterns during the FCT to those at the end of the delay period of the ICT. At the end of the delay period in an ICT trial, spatially selective PFC neurons were expected to represent a sole spatial location to which a saccade was going to be directed soon thereafter. Therefore, the activation pattern of PFC neurons in this period could be regarded as a built template when the network had already finished representing a single spatial location. On the other hand, the pattern of neuronal activation in the cue period of an FCT trial should be more ambiguous because two spatial locations are represented in the network. As the delay period progressed in an FCT trial, the activation pattern should gradually become similar to that in the ICT, since the monkey was required to prepare a saccade toward only one of the two locations. By testing how the neuronal activity was similar between these different periods in different tasks, we tried to examine how the spatial information needed for a subsequent saccade was constructed from the two locations presented in the FCT.

For the ICT, we used a 500-ms time bin in a preresponse period ranging from −1,000 to −500 ms from the end of the saccade. For the FCT, we used 250-ms time bins sliding through a trial in 1-ms steps. In a given time bin, we first calculated each neuron's average firing rates in each task condition. We then subtracted each neuron's grand-average firing rate among task conditions in that bin from its firing rates in each task condition, which gave the discrepancies of each neuron's firing rates in different conditions from its average. Finally, we calculated the rank correlation (Kendall's tau) between each time bin of the FCT and the preresponse period of the ICT between task conditions in which the monkey made the same response (e.g., T_ipsi cue trials in the ICT and T_out choice trials in the T_in vs. T_ipsi pair condition in the FCT). To evaluate the onset of significant correlation, we tested the significance of the correlation in each bin with α = 0.05. If there was a significant correlation between the FCT and preresponse period activity in the ICT, and if the correlation remained significant until the last time bin of the pericue period in the FCT, we regarded the onset of the first bin of these periods as the onset of significant correlation.

We also examined the correlation between the activation patterns in T_out choice trials in the FCT and those in T_in cue trials in the ICT. These trial conditions differed with respect to the final saccade direction and thus were expected to result in different activation patterns of directionally selective PFC neurons.

RESULTS

Behavioral Performance

We analyzed the behavioral performance of the animals during the recording sessions. The average proportion of correct performance in the ICT and FCT was 98.0% and 98.3% for monkey O and 99.8% and 99.9% for monkey E, respectively. There was no statistically significant difference in task performance between the ICT and FCT in either monkey (paired t-test, corrected P = 0.50 and 0.13). In correct ICT and FCT trials, the mean response time from the disappearance of the fixation point to the onset of a saccade was 263 and 265 ms for monkey O and 231 and 232 ms for monkey E, respectively. There was also no significant difference in the response time between the tasks (paired t-test, corrected P = 0.42 and 0.50).

We further examined the relationship between the animals' behavior and response directions in the tasks. Figure 1B shows the response times (lines) and preference indexes (bars) for each direction. Two-way ANOVA on each animal's response times revealed a significant main effect of direction (uncorrected P < 0.001 for both monkeys), but there were no main effects of the type of the tasks (uncorrected P = 0.12 and 0.24 for monkeys O and E, respectively) or the interaction between task and direction (uncorrected P = 0.70 and 0.35). In addition, there was no significant effect of direction on the preference indexes calculated from the proportion of choices for each direction in the FCT (1-way ANOVA, uncorrected P = 0.13 and 0.10). Therefore, the observed differences in response times for each direction were more likely to be attributed to the difference in motor execution process rather than the effect of the animal's unequivalent motivation for responses toward each direction.

Neuronal Database

We recorded neurons in and around the principal sulcus during performance of the tasks. Of 444 neurons recorded, 107 exhibited directionally selective activation during at least one epoch in the screening ICT trials. These neurons were further recorded in randomly intermingled ICT and FCT trials with rotated cue locations (see Tasks in materials and methods). Eighty-four neurons had at least five correct trials for each of the six FCT conditions (3 pair conditions × 2 choice results) and were confirmed to have directional selectivity in the postrecording offline analysis. We used these neurons for further analysis. We only used the activity of PFC neurons recorded during intermingled ICT and FCT trials for the analysis below.

Based on the rotated cue locations (T_in, T_ipsi, T_contra, and T_opp) determined by the receptive field of each neuron, we further tested the animals' task performance in the FCT for each direction (Fig. 1C). There was no difference in the proportion of T_in choice in all of the three pair conditions (1-way ANOVA, P = 0.11). The proportion of T_in choice was not significantly different from 0.5 (1-sample Wilcoxon rank sum test, corrected P > 0.05 for all the pair conditions). Also, there was no difference in the response times (1-way ANOVA, P = 0.11) and the normalized response times (Kruskal-Wallis test, P = 0.11) for each response direction. The normalized response times were not significantly different from zero (1-sample Wilcoxon rank sum test, corrected P > 0.05 for all the directions), meaning that the responses toward each of the four relative directions in the FCT were comparable to those to the same direction in the ICT. These results indicate that observed characteristics in neuronal activity reported below could not be attributed to the animal's preference toward a particular direction or to the difference in the degree of motor preparation toward each direction.

Choice-Predictive Activity

Figure 2 shows the activity of two representative neurons. Both neurons exhibited a larger firing rate in T_in than in T_out cue trials (t-test, 0–500 ms from cue onset, P < 0.05) during the cue period of the ICT (Fig. 2, top row, ICT trials). Thus they were more activated during the cue period when the visual cue was presented at T_in. We examined whether the activity of these neurons was related to the monkey's choice in the FCT (Fig. 2, bottom 3 rows, FCT trials for 3 pair conditions that included T_in). The neuron shown in Fig. 2A exhibited activation in response to the presentation of cues in FCT trials. The magnitude of cue-period activity was almost identical in trials in which the monkey chose T_in (T_in choice trials) and trials in which the monkey chose T_out (T_out choice trials). This is not surprising, because one of the two cues was always presented at the T_in location (neuron's preferred direction) in all three of these pair conditions. Therefore, this neuron was likely to exhibit a similar magnitude of cue-period activity when the visual cue was presented at T_in regardless of whether the monkey was going to choose T_in or T_out later in that trial.

Fig. 2.

Activity of 2 representative neurons. Neuronal firing in each condition was plotted separately for the ICT (top row) and 3 pair conditions that included T_in as 1 of the 2 cues in the FCT (bottom 3 rows: T_in vs. T_ipsi, T_in vs. T_contra, and T_in vs. T_opp conditions). The left half of each panel is aligned to the time from cue onset, and the right half is aligned to saccade offset. Different colors in the histograms and rastergrams correspond to different directions of saccades. Both neurons exhibited directionally selective transient activation to presentation of the cue in the ICT. In each pair condition of the FCT, the left neuron (A) exhibited nearly equivalent activation to the presented cues regardless of which cue location was chosen later in that trial. In the right neuron (B), the strength of the transient response to the cues in the FCT was significantly larger when it was followed by the animal choosing the neuron's preferred direction, even though the T_in cue was presented along with the T_out cue in all 3 pair conditions.

The other neuron, in Fig. 2B, showed cue-period activity that was related to the animal's subsequent choice in the FCT. In all three pair conditions, the strength of the transient response to the same two cues was significantly different depending on the monkey's subsequent choice. Whereas the neuron was strongly activated during the cue period in T_in choice trials, this activation was not observed in T_out choice trials, even though one of the cues was simultaneously presented at T_in (t-test between T_in and T_out choice trials, 0–500 ms from cue onset, P < 0.05, pair conditions collapsed). We refer to this firing pattern of PFC neurons (i.e., strong activation in T_in choice trials compared with T_out choice trials in the FCT) as “choice predictive.” In every FCT trial with a given pair condition, the monkey was presented with two physically identical cues at the same spatial locations regardless of which of them was chosen later in that trial. Therefore, choice-predictive activity cannot be explained as a mere reflection of the physical stimuli. Rather, the strong correlation between neuronal activity and the animal's subsequent choice suggests that PFC neurons play an active role in the free choice of a spatial location. Choice-predictive activity was also observed in the precue period (Fig. 2B). The activity of the neuron was slightly, but significantly, higher before the start of the cue period when the monkey was going to choose the neuron's preferred direction in the current trial (t-test between T_in and T_out choice trials, 1,000–0 ms before cue onset, P < 0.05, pair conditions collapsed).

Population Activity

We confirmed the presence of choice-predictive activity in the cue and precue periods of the FCT in a population analysis. Figure 3 shows population histograms and ROC transition of 59 PFC neurons that exhibited directionally selective cue-period activity. Differential firing in response to the cues presented in the FCT was consistently observed between T_in choice trials and T_out choice trials (Fig. 3B). We analyzed the activity of these neurons during the cue period in each task condition (0–500 ms from cue onset). In the ICT, cue-period activity in T_in trials (average 20.2 spikes/s) was significantly stronger than that in T_ipsi (12.0 spikes/s), T_contra (11.6 spikes/s), and T_opp (9.8 spikes/s) trials (paired t-test, corrected P < 0.001 for all comparisons). In the FCT, cue-period activity in T_in choice trials was also significantly stronger than that in T_out choice trials in all three pair conditions (corrected P < 0.001 for all pair conditions). In comparisons of different FCT pair conditions, the activity in T_in choice trials were comparable among the T_in vs. T_ipsi (19.4 spikes/s), T_in vs. T_contra (20.2 spikes/s), and T_in vs. T_opp (20.6 spikes/s) pair conditions (corrected P = 0.94, 0.89. and 0.94, respectively). However, in the T_out choice trials, neurons tended to be more activated in the T_in vs. T_ipsi pair condition (16.9 spikes/s) than in the T_in vs. T_contra (14.9 spikes/s, corrected P = 0.060) and T_in vs. T_opp (14.9 spikes/s, corrected P = 0.065) pair conditions, whereas there was no significant difference between the latter two pair conditions (corrected P = 0.97).

Fig. 3.

Population histograms and the change in receiver operating characteristic (ROC) values in directionally selective cue-period neurons. A and C: averaged histograms and the ROC transition in the ICT. Fifty-nine neurons exhibited a significant directionally selective transient response to the cue in the ICT. Different colors indicate different directions of saccades. In the ICT, neuronal activity decreased with presentation of the cue at a location other than the neuron's preferred direction (T_in), but the strength of the suppression was equivalent in the 3 T_out conditions. B and D: averaged histograms and the ROC transition in the FCT. In the FCT, the neurons exhibited differential activation that predicted the animal's subsequent choice of saccade direction. Different colors indicate the three pair conditions under investigation in the FCT. Solid and dotted lines in the histograms indicate the choice of the T_in and T_out directions later in that trial, respectively. The difference between solid and dotted lines with the same color (choice-predictive activity) was evident in the cue period but actually started to appear before cue onset. An ROC analysis showed the same result. The increase in the ROC values from 0.5 started 750 ms before the start of the cue period.

We also performed an ROC analysis on the cue-period activity of these neurons in T_in and T_out choice trials in the FCT. In all the three pair conditions, average ROC values (0.57, 0.61, and 0.63 for each of the 3 pair conditions) were all significantly larger than 0.5 (1-sample t-tests, corrected P < 0.001 for all of the conditions) in the cue period (0–500 ms from cue onset). By examining the change in the ROC value throughout the entire trial epoch using a sliding window, we observed early elevation of the ROC value that started before the presentation of the cues (Fig. 3D). A significant increase in the ROC value from 0.5 was observed 750 ms before the onset of the cues. The same analysis of activation during the ICT revealed that an elevation of the ROC value (calculated between the T_in and T_out cue trials) was observed 150 ms after the onset of the cue when the direction of the saccade was instructed (Fig. 3C). These results indicate that the choice-predictive activity of PFC neurons in the precue period of the FCT was not an artifact of the task structure, but rather reflected the influence of these neurons on the animal's decision making regarding the saccade direction when the choice was left to the animal.

Relationship Between Choice-Predictive Activity and Persistent Delay-Period Activity

To further investigate the role of PFC neurons in the decision making regarding the saccade direction, we compared the activities and firing properties of neurons with and without choice-predictive activity in the FCT. We categorized a neuron as choice predictive if it exhibited a differential activation between the T_in and T_out choice trials during the cue and precue periods (−1,000 to 500 ms from cue onset) in at least one of the three FCT pair conditions (t-test). Figure 4 shows population histograms of neurons with and without choice-predictive activity. PFC neurons with choice-predictive activity also showed directionally selective persistent delay-period activity. On the other hand, neurons without choice-predictive activity were activated only during cue presentation and did not exhibit persistent delay-period activity.

Fig. 4.

Population histograms of neurons with and without choice-predictive activity in the precue and cue periods. Conventions for the histograms are the same as those in Fig. 2. Activity for T_in (solid lines) and T_out choice trials (dotted lines) in the 3 FCT pair conditions were plotted separately for choice-predictive (A; n = 38) and unpredictive neurons (B; n = 21). Choice-predictive neurons also exhibited directionally selective activity during the delay period.

Based on this difference in task-related activity between neurons with and without choice-predictive activity, we further compared these two groups in terms of the persistence of activation (Fig. 5). We first quantified the strength of directionally selective persistent activity during the delay period of the ICT for each neuron by calculating the ROC value between T_in and T_out cue trials at the middle of the delay (1,000–1,500 ms from the start of the delay period of the ICT). When compared among all of the directionally selective neurons (Fig. 5A, n = 84), the strength of directionally selective persistent activity in the ICT was closely correlated (Pearson's R = 0.345, P < 0.001) with the strength of the choice-predictive difference in activity in the cue and precue periods of the FCT (quantified by the ROC value calculated between T_in and T_out trials within −1,000 to 500 ms from cue onset, pair conditions collapsed). When compared between groups, choice-predictive neurons had stronger directional selectivity in the delay period in the ICT than choice-unpredictive neurons (Fig. 5C, Wilcoxon rank sum test, P < 0.001). In addition, the choice-predictive neurons showed a higher baseline sustainability of activation even in the precue period (Fig. 5D, t-test, P < 0.05) and a stronger serial correlation of the ISI (Fig. 5E, Wilcoxon rank sum test, P < 0.05). Choice-predictive neurons were characterized by a higher sustainability of activation even between temporally distant time bins (Fig. 5B). The persistence of activation as measured by the baseline sustainability and the serial correlation of the ISI could be important for retention of the spatial information as sustained firing during the delay, and thus can be regarded as a key feature of PFC neurons in spatial working memory function. The coupling of these measures to the presence of choice-predictive activity in the early task epochs of the FCT suggests that the firing properties of PFC neurons that are essential to the memory function might also lead to a distinctive role of these neurons in the selection of spatial locations in a free choice condition.

Fig. 5.

Characteristics of firing properties of choice-predictive neurons. A: correlation between persistent directionally selective activity and choice-predictive activity. The strength of directionally selective activity in the delay period of the ICT (transverse axis, ROC values between T_in and T_out trials in 1,000–1,500 ms after the start of the delay period) was closely correlated with the strength of choice-predictive activity in the precue and cue periods of the FCT (vertical axis, ROC values between T_in and T_out trials in −1,000 to 500 ms from cue onset). B: comparison of baseline sustainability between choice-predictive (black) and choice-unpredictive (gray) neurons using different lengths of intervals. Choice-predictive neurons were characterized by a greater sustainability of activation even when the 2 bins were separated by a long interval. C–I: comparison of firing properties and animal's behavior between choice-predictive and unpredictive neurons. C: strength of persistent delay-period activity in the ICT. D: baseline sustainability for 400-ms interval. E: serial correlation of interspike interval (ISI). F: absolute direction of the receptive field (μ). Dots indicate the center of the receptive field (RF), and error bars indicate the size. Zero corresponds to horizontal direction contralateral to the recorded hemisphere, and positive and negative values indicate upper and lower direction from the horizontal meridian, respectively. G: size of the RF (1/$\sqrt{\beta }$). H: proportion of T_in choice in each pair condition and total proportion of T_in choice by collapsing pair conditions. I: difference of response times between T_in choice and T_out choice trials. Choice-predictive neurons were characterized by their high persistence of activation (C–E) compared with choice-unpredictive neurons, without differences in absolute direction and size of the RFs (F and G) and the animal's behavior (H and I).

We examined the possible effect of behavioral difference as well as the difference in receptive field properties between choice-predictive and unpredictive neurons. However, there were no difference between the two groups of neurons in the distribution of the preferred directions (Fig. 5F, Watson's test for homogeneity of circular data, P > 0.10) or the size of the receptive fields (Fig. 5G, Wilcoxon rank sum test, P = 0.21). Also, two-way ANOVA on the proportion of T_in choices revealed no main effects of neuron groups (choice-predictive/unpredictive neurons) and pair conditions, and no interaction of these two factors (Fig. 5H, P > 0.05 for both main effects and the interaction). One-sample t-tests revealed that the proportion of T_in choices was not significantly different from 0.5 in any of the neuron groups and pair conditions (corrected P > 0.05 for all the 2 × 3 combinations of neuron groups and pair conditions). The same comparison on the difference of the response times between T_in choice and T_out choice trials also revealed no significant main effects or interaction of neuron groups and pair conditions (Fig. 5I, P > 0.05 for both main effects and the interaction). One-sample t-tests revealed that the difference of the response times between T_in choice and T_out choice trials was not significantly different from zero in any of the neuron groups and pair conditions (corrected P > 0.05 for all the combinations).

Comparison of the Neuronal Activation Pattern in a State Space

To gain further insight into how spatial representations are held and integrated in the network of the PFC to perform a final saccadic response in the FCT, we used a dimensional reduction technique. The change in neuronal activation in each task condition in the ICT and FCT was expressed as a trajectory in a principle component state space (Fig. 6). In the ICT (Fig. 6A), the four trajectories that represented the neuronal activation patterns in T_in, T_ipsi, T_contra, and T_opp cue trials remained within the neighboring area until the time of cue onset. The trajectory for T_in cue trials then started to diverge from those for the other three cue conditions, which reflected a strong transient activity of directionally selective neurons in response to cue presentation (Fig. 3A). In the FCT (Fig. 6B), the trajectories for T_in choice and T_out choice trials reached slightly distant points in the space even at the beginning of the cue period. After the presentation of the cues, the trajectories for T_in choice trials further deviated from those for T_out choice trials and tracked similar paths to the trajectory for T_in cue trials in the ICT. Conversely, the trajectories for T_out choice trials returned to the initial state during the cue and delay periods, in a similar manner to T_out cue trials in the ICT.

Fig. 6.

Dynamics of neuronal activation using the state space based on a principal component analysis. A and B: the activity of prefrontal cortex (PFC) neurons in each of the ICT (A) and FCT conditions (B) are shown as trajectories inside a 3-dimensional principal component space. The activity around the cue and delay periods (from the start of the fixation period to 1,500 ms after the start of the delay period) in both the ICT and FCT was collectively used to construct a state space. Letters in panels show the start of the precue (P), cue (C), and delay (D) periods. In the ICT (A), the activity of PFC neurons were indistinguishable at the start of the cue period. The trajectory in the T_in cue condition then started to diverge from that in T_out cue conditions. In the FCT (B), the trajectories for T_in choice trials took similar courses to those for the T_in cue condition in the ICT, whereas those for T_out choice trials resembled those for T_out cue conditions in the ICT. However, there was little separation between the trajectories for the T_in and T_out choice trials in the T_in vs. T_ipsi pair condition compared with the other pair conditions. dim, Dimension. C: the distance between the trajectories for T_in choice and T_out choice trials in the state space. The trajectories for trials with different choices immediately diverged from each other in the cue period in the T_in vs. T_contra/T_opp conditions, but not in the T_in vs. T_ipsi condition.

However, in the T_in vs. T_ipsi pair condition, the trajectory for T_out choice trials remained relatively adjacent to that for T_in choice trials, compared with the other two pair conditions. To evaluate the difference in neuronal activation patterns, we calculated the distance between the trajectories for the T_in choice and T_out choice trials in each of the three pair conditions in the FCT (Fig. 6C). In the T_in vs. T_contra and T_in vs. T_opp pair conditions, the distance between the trajectories for T_in choice and T_out choice trials increased in the cue period. In the T_in vs. T_ipsi pair condition, the distance between T_in and T_out choice trials remained relatively small in the cue period and gradually increased during the delay period. The average distance between T_in and T_out trajectories during the cue period was 24.7, 66.5, and 72.8 for the T_in vs. T_ipsi, T_in vs. T_contra, and T_in vs. T_opp pair conditions, respectively. Paired t-tests revealed that there was a significantly less distance between the trajectories for T_in and T_out choice trials in the T_in vs T_ipsi pair condition than in the other two pair conditions (corrected P < 0.001). The distance between the T_in and T_out trajectories was also greater in the T_in vs. T_opp pair condition than in the T_in vs. T_contra condition (paired t-test, corrected P < 0.001), but this difference was small. These results support the observation in the population histograms and ROC analysis that choice-predictive activity was established more slowly when the two cues were presented in the same hemifield (T_in vs. T_ipsi pair condition).

Effect of Cue Configuration

Previous analyses suggested that the time course of the establishment of choice-predictive activity was dependent on the configuration of the cues. To investigate how the final spatial representation was constructed in each pair condition of the FCT, we used a between-task correlation analysis on the firing patterns of directionally selective PFC neurons. In this analysis, we examined the correlation of neuronal activity between the FCT and the preresponse period of the ICT. For each task condition and each time bin in the FCT, we calculated the correlation coefficient between the activation of directionally selective PFC neurons (n = 84) in that time bin and that in the preresponse period (1,000–500 ms before the end of the saccade) of the ICT. As the decision-making process regarding the subsequent saccade direction progressed in the FCT, the correlation coefficient was expected to increase because the activation pattern of PFC neurons should have been similar to that when the animal was ready to make a saccade in the ICT. By measuring the transition of correlation coefficients in each task condition, we tried to clarify how the configuration of the cues might influence the dynamics of visuospatial decision making in FCT trials.

Figure 7 shows the correlation in the neuronal activation pattern between the FCT and the preresponse period of the ICT. We first examined the correlation between the ICT and FCT in which the monkey eventually made a response to the same direction (correlation between T_in cue/choice trials for Fig. 7A, T_out cue/choice trials for Fig. 7B). In the T_in vs. T_contra and T_in vs. T_opp pair conditions, the correlation coefficients started to rise from zero at around the beginning of the cue period. In the T_in vs. T_contra pair condition, a significant correlation started to be observed 142 ms before the start of the cue period in T_in choice trials and 48 ms after cue onset in T_out choice trials. In the T_in vs. T_opp pair condition, a significant correlation started to be observed 90 ms before and 113 ms after cue onset in T_in choice and T_out choice trials, respectively. However, in the T_in vs. T_ipsi pair condition, the development of a correlated activation pattern was weak. In T_in choice trials, the onset of significant correlation was 408 ms after cue presentation. In T_out choice trials, significant correlation that lasted stably during the delay period was not observed. This result indicates that, in the network of spatially selective PFC neurons, a sole spatial representation was constructed from the presented two locations more slowly when the two cues were located in the same hemifield, consistent with the observation in the previous analyses on population activity (Fig. 3D) and the neuronal state space (Fig. 6C). Especially, the slower development of choice-predictive activity in the T_in vs. T_ipsi pair condition may be the result of the disarranged construction of spatial representation in T_out choice trials in this pair condition.

Fig. 7.

Construction of spatial representation in the FCT. A and B: each line shows the correlation coefficients between the neuronal activation pattern at each time bin in a given FCT condition and that of the preresponse period in a corresponding ICT condition in which the monkey made the same response. Data are separately plotted for T_in choice (A) and T_out choice trials (B). Different colors indicate different pair conditions in the FCT. Thick solid lines show ranges of significant correlation in each task condition. Triangles at top show the onset of significant correlation that lasted through the delay period. In the T_in vs. T_contra/T_opp pair conditions, significant correlation began around the start of the cue period. In the T_in vs. T_ipsi pair condition, significant correlation was observed at the end of the cue period in T_in choice trials and was not observed in T_out choice trials. C: results of a similar correlation analysis calculated between T_out choice trials in the FCT and T_in cue trials in the ICT. The neuronal activation pattern in FCT trials with T_out choice started to diverge from that in ICT trials with a T_in cue before the start of the cue period in the T_in vs. T_contra/T_opp conditions. However, a significant negative correlation was not observed until the delay period in the T_in vs. T_ipsi pair condition.

For T_out choice trials in the FCT, we also examined the correlation of the neuronal activation pattern with that in T_in cue trials in the ICT (Fig. 7C). In all three pair conditions of the FCT, T_in cue was presented during the cue period. However, since the final saccade directions were not in the neurons' receptive fields in T_out choice trials, the activity of PFC neurons were expected to gradually become different compared with when T_in was instructed in the ICT. We confirmed this prediction as an early negative correlation of the neuronal activation patterns in the T_in vs. T_contra and T_in vs. T_opp pair conditions. In T_out choice trials, the population activity of PFC neurons started to differ from that in T_in cue trials in the ICT at 296 and 158 ms before the onset of the cues in the T_in vs. T_contra and T_in vs. T_opp pair conditions, respectively. In contrast, the onset of a significant negative correlation was 941 ms after cue onset (441 ms after the start of the delay period) in the T_in vs. T_ipsi pair condition. This means that the activation state of PFC neurons in T_out choice trials in the T_in vs. T_ipsi pair condition remained indistinguishable from that in T_in cue trials in the ICT for a longer period. This result also suggested that the construction of spatial representation for T_out was slower in the T_in vs. T_ipsi pair condition than in the other pair conditions.

DISCUSSION

In the present study, we investigated the role of spatially selective PFC neurons in an animal's decision about saccade direction in a free choice condition. When T_in was later chosen as a saccade direction, PFC neurons were strongly activated by cue presentation despite the presence of another cue outside their preferred direction. Choice-predictive activity was distinct from the very beginning of the cue period and was observed even before cue onset.

The positive correlation between the strength of the delay-period activity in the ICT (memory-related activity) and the strength of the choice-predictive activity in the cue and precue periods of the FCT (decision-related activity) revealed that PFC neurons with stronger memory-related activity in the ICT tended to show stronger choice-predictive activity in the pericue periods of the FCT. The stronger sustainability of firing in neurons with choice-predictive activity suggested that memory and decision functions are supported by a common feature of PFC neurons to sustain their activation state within the circuitry. In addition, when to-be-chosen T_out was located in a hemifield different from that in which T_in was located, the transient response to cue presentation was weak. However, when to-be-chosen T_out was in the same hemifield as T_in, the cue-period activity was stronger, even though the neuron's preferred direction was not going to be chosen. This indicates that unnecessary spatial information tended to be suppressed from the beginning of its representation in the PFC, but this adaptive modulation was modest if the two spatial locations were in the same hemifield. Our present study revealed that the role of the PFC in the decision-making process is closely linked to its role in information maintenance process, and these different processes share the same functional characteristics that emerged from the underlying cellular mechanism.

Activity of PFC Neurons Related to the Animal's Decision

In the present study, we found that the strength of neurons' activity in precue and cue periods was correlated with the animal's subsequent decision regarding the saccade direction (choice-predictive activity). However, in the FCT, one of the six pair conditions was randomly assigned in each trial. Also, FCT trials were randomly intermingled with ICT trials. Only after cue presentation could the animal know whether they were allowed to choose the saccade direction by themselves or where the options for the choice would be. Therefore, it was impossible for the animal to make a reasonable decision before cue onset. We propose that this early choice-predictive activity can be explained as an influence of fluctuating neuronal firing before the start of a trial. In each trial, the activity of directionally selective neurons can randomly fluctuate during the precue period. This fluctuation of activity can be regarded as baseline random noise in spatial representation in the PFC. If the activity of neurons that are responsible for a particular direction happen to be elevated during the precue period of an FCT trial, these neurons should be able to more quickly respond to the presentation of cues, one of which appeared at their preferred direction. In the network of the PFC, neurons responsible to different spatial locations have inhibitory connections so that the PFC retains only the most relevant spatial information in a winner-take-all manner (Compte et al. 2000; Rao et al. 1999; Wang et al. 2004). Therefore, the faster construction of a spatial representation will disturb the formation of other spatial representations through mutual competition, resulting in the adoption of the prematurely biased location as the direction of the saccade in the current trial. As a result, trials in which directionally selective PFC neurons show slightly stronger activation before the cue period should be overrepresented among trials in which their preferred direction was later chosen by the animal.

Previous studies have reported the activity of neurons in the PFC, frontal eye field, supplemental eye field, and lateral intraparietal cortex using memory-guided (Watanabe et al. 2006; Watanabe and Funahashi 2007) or visually guided (Coe et al. 2002) free choice tasks. However, in these studies, a same set of fixed locations were repeatedly presented in a block of trials. Therefore, a detailed investigation of the time course of neuronal activity and the interpretation of its role in decision making were difficult because the animal could easily predict the available options independently of the progress of a trial. Also, in those task setups, the animal could exhibit a strong tendency or strategy to repeatedly choose the same option. Therefore, the previous studies used particular reinforcement rules to prohibit the animal from choosing the same option repeatedly and forced the animal to choose different options. This procedure can be regarded as a trained allocation of choices administrated by a reward schedule. In the present study, we intermingled ICT and FCT trials and changed the combination of the options for a decision. In addition, the absolute locations of the four options changed randomly for the animal depending on the receptive field of the neuron. As a result, the monkeys were presented with different decision contexts from trial to trial, and their choices were substantially varied among options without restrictive rules for decision. We propose that our current experimental design is more appropriate for the investigation of the neuronal mechanisms of internally driven decision making compared with the designs of previous studies.

In other cortical areas, a biasing effect of baseline fluctuation of the neuronal activity on the subsequent animal's behavior has been reported (Platt and Glimcher 1999; Shadlen and Newsome 2001). For example, Shadlen and Newsome (2001) reported that the activity of primate lateral intraparietal neurons before the onset of random-dot motion stimulus was higher when motion coherency was weak and the neurons' preferred motion direction was going to be chosen. They argued that this was because the existing neuronal fluctuation before stimulus presentation biased the subsequent competition between the representations of different motion directions. Rolls and Deco (2011) recently reported that this kind of bias based on random fluctuation could actually take place in an integrate-and-fire attractor network model. They confirmed the relationship between preexisting random fluctuation in spontaneous activity and the result of the subsequent neuronal competition in an artificial network in which there was no potential confounding such as a subject's specific behavioral strategies. Our present results are in accord with these previous reports.

Relationship Between Decision Making and Memory Maintenance

We found that neurons with choice-predictive activity during cue and precue periods showed higher sustainability of firing such as an elevated delay-period activity (Figs. 4 and 5). The persistent delay-period activity of PFC neurons while the monkey is remembering a particular spatial location is thought to be the neural basis of spatial working memory (Funahashi et al. 1989, 1990; Fuster 2008; Goldman-Rakic et al. 1990; Miller and Cohen 2001). An elevated firing rate sustained during several seconds of delay is not likely to be supported only by subcellular mechanisms and is instead attributed to the network property of the PFC with recurrent feedback inputs (Constantinidis and Wang 2004; Wang 2013). We propose that the strong correlation between memory-related activity (persistent delay-period activity in the ICT) and decision-related activity (choice-predictive activity in the FCT) is a consequence of this network property of the PFC. Since PFC neurons are mutually interconnected, the incidental activation of a group of neurons before the start of a trial could persist for some time through this network. Heterogeneity in the activation level among neurons might further be amplified during the precue period through mutual facilitation and competitive inhibition of neurons with the same and different directional selectivities, respectively. The premature imbalance of activation will then result in a difference in the strength of cue representations in the cue period and the animal's final choice toward the strongly represented direction in the FCT.

In recent electrophysiological research, there has been a debate about the role of the lateral PFC in spatial information processing. Several studies have proposed that lateral PFC is more related to the spatial attention than spatial working memory (DeSouza and Everling 2004; Everling et al. 2002; Lebedev et al. 2004; Lennert and Martinez-Trujillo 2011, 2013; Messinger et al. 2009; Tremblay et al. 2015). For instance, by using a behavioral task in which a location to remember and a location to allocate a visually guided attention are separated, Lebedev et al. (2004) showed that large proportion of lateral PFC neurons are selective to attended rather than remembered spatial location. In their study, spatial working memory process was depicted as maintenance memory and separated from the attention process by preventing the animal from paying attention to the remembered location. However, the concept of working memory includes both active maintenance and manipulation of information (Baddeley 1986, 2003; Baddeley and Hitch 1974). Working memory tasks in animals (Dudchenko 2004) and humans (Miyake et al. 2000) typically consist of attentional shifting, updating, or inhibitory control of the maintained information. This is because a mere maintenance of information is unlikely in a variety of cognitive operations, and the maintenance of task-relevant information necessarily requires attentional control. This joint formularization of memory and attention is the essence of the working memory, and the validity of working memory concept in clinical, developmental and experimental psychology (Baddeley 2003; Conway et al. 2003; Kane and Engle 2003; Saperstein et al. 2006) suggests that maintenance and attentional manipulation of information cannot be dissociated as independent processes.

Based on these psychological backgrounds, the elucidation of how maintenance and manipulation of information are simultaneously and jointly performed in the activity of cortical neurons is essential for the understanding of the neuronal mechanism of working memory. Therefore, in the present study, we used a traditional memory-guided saccade task combined with a subjective decision-making process. As a result, we observed that fluctuation of the activity of PFC neurons before cue presentation induced an early bias in the representation of spatial cues and eventually influenced the animal's decision. Importantly, this task-related activity was correlated with more fundamental firing characteristic of PFC neurons to sustain its activation state for relatively longer period. If there is no such persistence in neuronal activity, the premature fluctuation of activity before a trial could not survive until presentation of the cues and should never influence the animal's behavior. In this sense, the capability of the PFC network to maintain spatial information played a pivotal role in the decision-making process under a free choice condition. This is a succinct example that the network property of the PFC that enables the maintenance of information can be regarded as a key feature in understanding the PFC's roles in other cognitive processes (Procyk and Goldman-Rakic 2006; Wang 2008). Especially, the effect of preexisting neural state on subsequent decision making is a prevailing subject in recent noninvasive electrophysiological studies in human (Bengson et al. 2014; Bode et al. 2012; Hesselmann et al. 2008). Our present report directly demonstrates the neuronal underpinning of such phenomena from the viewpoint of the known function and characteristics of PFC neuron's activity and also shows that the influence of preexisting fluctuation takes place on the order of hundreds of milliseconds in task-related neuronal activity. Our report provides a clue for integrated understanding of lateral PFC's role in spatial decision making and working memory functions, from a viewpoint of basic characteristics of neuronal firing in this cortical area.

Competition of Spatial Representations Within or Between Hemifields

In previous studies regarding the role of the PFC in working memory function, a single visual cue was used to inform the animals of the location to be remembered for an upcoming saccade (Boch and Goldberg 1989; Funahashi et al. 1989, 1990; Rainer et al. 1998). In these studies, PFC neurons with mnemonic visuospatial activity tended to have directional selectivity toward locations contralateral to the side of the hemisphere being recorded. A unilateral lesion to the PFC was reported to result in disrupted performance of the memory-guided saccade to the contralateral hemifield (Funahashi et al. 1993a). These findings suggest that the PFC is organized to participate in the processing of spatial information in the contralateral hemifield (Funahashi 2013).

The activity of PFC neurons during the performance of a memory-guided saccade task in which the monkeys chose the saccade direction by themselves has been previously reported (Watanabe et al. 2006; Watanabe and Funahashi 2007). However, in those experiments, the locations of the cues were fixed to the four perpendicular directions and all four cues were repeatedly presented in a block of trials. Therefore, the influence of contralateral organization of the PFC on the representation of multiple pieces of spatial information could not be examined. In the present study, we compared the time course of the emergence of choice-predictive activation among the three FCT pair conditions. We found that choice-predictive activity developed more slowly in the T_in vs. T_ipsi pair condition than in the T_in vs. T_contra/T_opp conditions (Fig. 6). The slow construction of the final spatial representation was especially obvious in T_out choice trials in the T_in vs. T_ipsi pair condition (Fig. 7). We propose that this stronger representation of the unchosen direction when the two cues are located in the same hemifield is the result of the contralateral organization of the PFC. Since neurons with directional selectivity toward a particular side of the visual space are assembled in the contralateral hemisphere of the PFC, they may be more tightly interconnected through local circuits than neurons with preferences for different hemifields, which are more likely to be distributed in different hemispheres and thus require callosal connections to interact with each other. A recent study on the concurrent memorization of multiple spatial locations also suggested a stronger interaction between spatial representations within the same hemifield (Matsushima and Tanaka 2014). Through this stronger local connection, the spontaneous fluctuation of neuronal activity can be more frequently shared by neurons with a preference for the same hemifield. The shared fluctuation between neurons before the start of a trial can be regarded as unbiased preexisting spatial representation, which leads to strong representations of both subsequently chosen and unchosen locations in response to the presented cues. In contrast, the preexisting activation levels may vary between neurons with preferences for locations in different hemifields, which can then cause suppression of the representation of the subsequently unchosen location by amplifying the premature bias. Therefore, the difference in the time course of the development of choice-predictive activity among pair conditions can be explained by uneven lateralization of directionally selective neurons in the PFC. Future studies with simultaneous recordings of PFC neurons are needed for quantitative investigation of correlated fluctuation in the spontaneous activity of spatially selective neurons.

GRANTS

This research was supported by MEXT of Japan Grants-in-Aid for Scientific Research 21240024 and 25240021 (to S. Funahashi) and by a Japan Society for the Promotion of Science (JSPS) Grant-in-Aid for JSPS Fellows 23·7155 (to K. Mochizuki).

DISCLOSURES

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

AUTHOR CONTRIBUTIONS

K.M. and S.F. conception and design of research; K.M. performed experiments; K.M. analyzed data; K.M. and S.F. interpreted results of experiments; K.M. prepared figures; K.M. and S.F. drafted manuscript; K.M. and S.F. edited and revised manuscript; K.M. and S.F. approved final version of manuscript.