Frontal eye fields in macaque monkeys: prefrontal and premotor contributions to visually guided saccades

Abstract

Neuronal spiking was sampled from the frontal eye field (FEF) and from the rostral part of area 6 that reaches to the superior limb of the arcuate sulcus, dorsal to the arcuate spur when present (F2vr) in macaque monkeys performing memory-guided saccades and visually guided saccades for visual search. Neuronal spiking modulation in F2vr resembled that in FEF in many but not all respects. A new consensus clustering algorithm of neuronal modulation patterns revealed that F2vr and FEF contain a greater variety of modulation patterns than previously reported. The areas differ in the proportions of visuomotor neuron types, the proportions of neurons discriminating a target from distractors during visual search, and the consistency of modulation patterns across tasks. However, between F2vr and FEF we found no difference in the magnitude of delay period activity, the timing of the peak discharge rate relative to saccades, or the time of search target selection. The observed similarities and differences between the 2 cortical regions contribute to other work establishing the organization of eye fields in the frontal lobe and may help explain why FEF in monkeys is identified within granular prefrontal area 8 but in humans is identified within agranular premotor area 6.

Introduction

We understand how the occipital, parietal, and temporal lobes are divided into distinct areas with unique intrinsic structure, connectivity, and functional properties (reviewed by Kaas 2017). Although we know that the frontal lobe can also be divided into more or less distinct areas, we lack comprehensive knowledge about the boundaries and functional transitions between neighboring areas. This information is necessary for reliable comparisons across species. Here we address this general issue through comparison of neighboring areas in lateral frontal cortex.

For more than a century, an area known as the frontal eye field (FEF) has been identified with the guidance and control of eye movements (reviewed by Schall 2015; Schall et al. 2017). In macaque monkeys, much is known about the properties of FEF neurons and connectivity with cortical and subcortical structures; however, the boundaries of FEF and functional transitions with neighboring areas remain uncertain. In monkeys, FEF has been defined as the subregion at the caudal end of Brodmann’s area 8 in which low current microstimulation elicits saccades (Robinson and Fuchs 1969; Bruce et al. 1985; Huerta et al. 1987; Stanton et al. 1988; Schall 1991). In humans, though, FEF has been described as being located in Brodmann’s area 6 (reviewed by Paus 1996; Amiez and Petrides 2009), even though some of the original descriptions of human FEF (e.g. Foerster 1931, 1936) referred to an area 8αβγ as defined by Vogt and Vogt (1919). This is still a matter of debate, as others describe 2 FEF in humans and ascribe homology between the dorsal human FEF and macaque FEF and between the ventral human FEF and macaque area 45B (reviewed by Borra and Luppino 2021).

Several other observations raise questions about the caudal boundary of FEF in macaque monkeys. Premotor cortex in area 6 is traditionally identified with the guidance of limb movements (Wise 1985; Wise et al. 1992; Wise et al. 1996; Kalaska et al. 1997; Kalaska et al. 1998; Cisek and Kalaska 2005; Thura and Cisek 2014; Neromyliotis and Moschovakis 2017b; Neromyliotis and Moschovakis 2018). Nevertheless, neuronal spiking associated with reaching or grasping can be modulated by gaze angle (Baker et al. 1999; Boussaoud et al. 1993; Boussaoud 1995; Mushiake et al. 1997; Boussaoud et al. 1998; Cisek and Kalaska 2002; Baker et al. 2006). Moreover, multiples lines of evidence associate area 6 with saccade production. First, microstimulation with low currents caudal to the arcuate sulcus in agranular area 6 elicits saccadic eye movements, albeit at slightly higher thresholds than in FEF in the rostral bank of the arcuate in area 8 (Fujii et al. 2000; Neromyliotis and Moschovakis 2017a). Second, neuronal spiking in premotor regions modulates in association with gaze shifts produced during saccade tasks (Kurata 2017; Neromyliotis and Moschovakis 2017b; Neromyliotis and Moschovakis 2018). Third, 2-deoxyglucose uptake associated with saccades is observed in area 6 in regions oligosynaptically connected with the abducens nucleus (Moschovakis et al. 2004). Fourth, functional magnetic resonance imaging (fMRI) activation is observed in the rostral premotor region of macaques during saccade tasks (Koyama et al. 2004; Moschovakis et al. 2004; Baker et al. 2006; cf. Ford et al. 2009).

Based on these characteristics, the term premotor eye field has been coined to refer to this caudal, postarcuate region (Amiez and Petrides 2009; Savaki et al. 2015; Neromyliotis and Moschovakis 2017b, 2018). The label “eye field” may be misleadingly restrictive given the prevalence of arm movement-related activity in this region (Hoshi and Tanji 2000; Raos et al. 2004). Also, it is not the only premotor eye field; other regions of premotor cortex are associated with gaze behavior (Boussaoud et al. 1998; Hoshi and Tanji 2004; Neromyliotis and Moschovakis 2017a). However, its interconnectivity with supplementary eye field (Huerta and Kaas 1990; Stepniewska et al. 2006) and its proximity to FEF motivate a more direct comparison between this premotor eye field and FEF during gaze behavior.

Although the contribution of FEF to more cognitively demanding tasks, such as memory-guided saccades and both overt and covert visual search is well known (e.g. Wardak et al. 2006; Schall 2015; Fiebelkorn and Kastner 2020; Gaillard and Ben Hamed 2020), parallel contributions of the premotor region have not been tested. Thus, we compared single neuron discharge profiles sampled in both banks of the arcuate sulcus during a memory-guided saccade task and a shape singleton visual search task. Because of suggested anatomical and physiological distinctions between dorsal (PMd) and ventral (PMv) premotor cortex (F2 and F5; see Hoshi and Tanji 2007 for review) and between rostral and caudal divisions of dorsal premotor cortex (Geyer et al. 2000; Luppino et al. 2003), as well as a corresponding distinction within dorsal premotor cortex where cognitive functions are represented more rostrally and motor functions more caudally (Abe and Hanakawa 2009; Nakayama et al. 2016), we targeted the rostroventral portion of the dorsal premotor cortex (F2vr) in the caudal bank (Fig. 1A). A preliminary analysis of some of these data has been presented previously in abstract form (Zinke et al. 2015).

Figure 1.

Recording sites. (A) Regions of the peri-arcuate macaque frontal lobe, using a mixture of nomenclatures. Brodmann’s area 8 is divided into areas 8B dorsally and 8A near the arcuate sulcus. 8A is further divided into areas 8Ar on the convexity and areas 8 m and 8 l as the medial and lateral aspects, respectively, of the rostral bank of the arcuate sulcus. Together, areas 8 m and 8 l comprise FEF. More laterally lie areas 44, 45A, and 45B. Caudal to the arcuate sulcus lies Brodmann area 6 which has been subdivided into area 6Vr (F5), 6Vc (F4), and 6D. Area 6D has itself been divided to rostral (F7, 6DR) and caudal (F2) regions which correspond to supplementary eye field (SEF) and dorsal premotor cortex (PMd). PMd itself has been divided into rostroventral (F2vr) and dimple (F2d) aspects. We targeted our recordings to FEF (in red) and this rostroventral region of the dorsal premotor cortex F2vr (blue). (B) Sulcal patterns for each of the 3 monkeys are illustrated. Principle and arcuate sulci are labeled for Ga in the top left. Patterns are mirrored for monkey Da such that all 3 illustrations depict anterior to the left, posterior to the right, medial to the top, and lateral to the bottom. Neuron counts for each location are depicted by the size of the circle, with larger circles indicating a larger neuron count. FEF recording locations are red, F2vr recording locations are blue. Neuron counts for the memory-guided saccade task are shown on the right and for shape singleton search on the right.

Materials and Methods

Monkeys

Data were collected from 3 male macaque monkeys (Macaca radiata) weighing approximately 8.0 kg and ranging in age from 7 to 9 years. All procedures were in accordance with the National Institutes of Health Guide for the Care and Use of Laboratory Animals and approved by the Vanderbilt Institutional Animal Care and Use Committee.

Recording Techniques

Magnetic resonance imaging (MRI) compatible headposts were placed on the midline and recording chambers were placed over the arcuate sulcus. Surgery was conducted under aseptic conditions under isoflurane anesthesia. Antibiotics and analgesics were administered postoperatively. Details about the surgical procedure have been described previously (Schall et al. 1995a; Sato and Schall 2001; Cohen et al. 2009). Electrophysiological data were obtained from linear electrode arrays, either a 24-channel Plexon U-probe (monkeys Ga, He) or a 32-channel Neuronexus Vector Array (monkey Da). Both types of probes had a 150 μm recording contact spacing. Data were streamed to a data acquisition system: Multi-Neuron Acquisition Processor (40 kHz, Plexon, Dallas, TX—monkeys Da, Ga, and He) or a Tucker Davis Technologies System 3 (25 kHz, monkey Da). Single units were identified online using a window discriminator (Plexon) or principal component analysis (TDT).Units recorded from the TDT system were sorted offline using Kilosort (Pachitariu et al. 2016). Eye position was recorded using EyeLink 1000 (SR Research). Eye position was calibrated daily and streamed to the data acquisition system and stored at 1 kHz.

FEF was localized using anatomical MRI reconstructions with a chamber grid (Crist instruments) projected onto the reconstructions. We identified grid locations whose trajectory passes through the rostral bank anterior to the genu of the arcuate sulcus, the anatomical location of the functionally defined FEF (Stanton et al. 1988). Neuronal responses were typical of FEF neurons and microstimulation with low currents (<50 μA) elicited saccades. F2vr was identified in a similar manner, but in grid locations having trajectories passing through the caudal bank of the arcuate sulcus. For the 2 monkeys with a clear arcuate spur (Da and Ga), grid locations were chosen with a trajectory passing through the dorsal bank of the spur (F2vr). For the monkey without a spur (He), the locations were chosen in the caudal bank directly posterior to FEF. Microstimulation was not used to elicit saccades in these recording locations. Locations of the recording sites for neurons recorded in each task are projected onto each monkey’s sulcal pattern in Fig. 1B.

Behavioral Tasks

Monkeys performed 2 tasks. One task was a memory-guided saccade task (Fig. 2A). The details of this task have been described previously (Lowe and Schall 2018). Briefly, after fixating a central point for a period of 500 ms, a peripheral target stimulus was briefly presented at one of 8 locations, evenly spaced around an invisible circle beginning at the horizontal meridian, at 8° eccentricity while monkeys maintained fixation. After a variable delay period (range 300 to 800 ms), the fixation point disappeared, and the monkey was rewarded for making a saccade to and maintaining fixation on the cued location. To provide a fixation stimulus, the peripheral target was reilluminated when the monkey attained fixation on the remembered location. A juice reward was delivered if the monkey successfully fixated on the remembered location for 500 ms. The trial was aborted and a 2000 ms timeout was delivered if the monkey broke central fixation prematurely, made a saccade to an incorrect location, or broke fixation on the peripheral target after an initially correct saccade.

Figure 2.

Task diagrams. (A) Memory-guided saccade task. The task begins with the onset of a fixation spot. After a 500 ms fixation period, a single square target is briefly presented in the periphery. After the target disappeared, the monkey maintains fixation on the fixation spot for a variable delay period of 300 to 800 ms. After this delay, the fixation spot disappears and the monkey makes a saccade to the cued location. When the saccade is made to the correct location, the target reappears to provide feedback for the correct location for fixation. If the target is successfully fixated for 500 ms, reward is delivered. (B) Shape singleton search task. An example search array (top) is shown with a singleton T among 7 distractor Ls. Trial sequence (below) indicates key trial events. The trial begins when a fixation spot is shown. After a delay, the fixation spot disappears and a search array is shown. A saccade is made to the shape singleton and fixation is maintained on it. If the target is successfully fixated for 200 ms, reward is delivered.

The second task was shape singleton visual search that has been described in detail previously (Cosman et al. 2018). Briefly, after fixation of a central spot, monkeys were presented an array of 8 stimuli comprising either gray Ts or Ls (Fig. 2B), located at the same 8 candidate locations for the memory-guided saccade task (i.e. evenly spaced around a circle beginning at the horizontal meridian and 8° eccentricity). Within 1 session these stimuli were presented with one of 4 possible orientations. One stimulus was a shape singleton; 1 T could be presented among 7 Ls or 1 L among 7 Ts. After a variable fixation period of 300 to 800 ms, the stimulus array was presented, and the monkey was required to make a saccade to the singleton shape. A juice reward was delivered if the monkey successfully fixated the target for 200 ms. Trials were aborted if the monkey broke fixation prematurely, made a saccade to an incorrect stimulus, or broke fixation from the target prematurely after an initially correct saccade. On a subset of trials, a color singleton distractor was presented within the array; one of the distractors was chromatic. Singletons and salient distractors were presented exclusively on the horizontal or vertical meridians. Results with trials using a salient distractor were reported elsewhere (Cosman et al. 2018); these trials were excluded from the present analysis.

Data Analysis

Spike density functions (SDFs) were calculated by convolving spike trains with a kernel that resembles the postsynaptic potential elicited by an action potential (Thompson et al. 1996). For averaging across neurons, SDFs were normalized by z-scoring across the full trial and subtracting the mean activity during the prestimulus period as baseline (Lowe and Schall 2018). We took the time between −300 and −100 ms before array presentation as the prestimulus baseline period. This method of scaling responses reduces the skewness of the SDF across the population and generates a comparable range of activity across neurons without erroneously scaling neurons with little modulation (Lowe and Schall 2018). Because these measures are not strictly Z-scores, they will be referred to as arbitrary units (AUs).

Preferred locations were defined as the location in which the response modulation was greatest for a correct saccade made to that location. For visually responsive neurons, this would correspond to the visual receptive field (RF), whereas for movement-related neurons, this is more comparable to the movement field. Target selection times (TST) were determined from the difference of SDFs of trials where the search target was presented within and trials where the target was presented opposite to the preferred location. For each neuron, we calculated a Z-scored difference function. We first calculated the difference in response modulation for targets presented at the 2 locations. Then, we divided this difference function by the standard deviation of that function during the prestimulus baseline period (300 ms before the onset of the search array). Finally, we subtracted the mean of the prestimulus baseline difference. TST was defined as the earlier of 2 times (1) the time the Z-scored difference function exceeds 2 and continues to exceed 6 for at least 20 ms continuously or (2) the time the difference function exceeds 2 standard deviations of the baseline difference for at least 50 ms continuously. Visual latency was calculated in a similar fashion but the SDFs themselves aligned on array onset were required to meet the above criteria. To assess the onset of presaccadic activity, the SDF was smoothed with a 20 ms uniform kernel. Then, starting at the time of the saccade, the correlation of the smoothed SDF across time was calculated in a 100 ms window. This window was moved backward in time until the correlation was no longer significant. The end of the window that produced the first nonsignificant correlation was taken as the time of presaccadic activity onset. Peak presaccadic activity was defined as the time at which the smoothed SDF reached its highest value. Neurons for which the algorithm identified time of peak presaccadic activity as preceding the presaccadic activity onset by more than 20 ms are excluded from analysis.

For neuronal classification, the presence or absence of visual activity was determined by the visual latency. If the visual latency was <200 ms, the neuron was considered to have visual activity. Similarly, if the neuron had elevated activity at the time of the saccade using an identical algorithm as well as a positive correlation of response across time in the 20 ms before the saccade, the neuron was considered to have movement-related activity.

Subsequent classification was accomplished using a consensus clustering algorithm (Lowe and Schall 2018). First, we applied the algorithm to all neurons in the sample regardless of region. For this clustering, we set a minimum cluster size of 20; Lowe and Schall set a minimum cluster size of 10 but applied the algorithm to 1 area, so we doubled the minimum cluster size to accommodate the second region. Then, we applied the algorithm to each region separately. For this clustering, we kept the minimum cluster size as 10. Importantly, we also included responses to stimuli outside the neurons’ preferred locations. Otherwise, the clustering was identical to Lowe and Schall (2018).

To assess the consistency of clustering in the memory-guided saccade task and the visual search task, we calculated a signed χ² for each pairwise combination (Lowe and Schall 2018). That is, to account for either over- or underrepresentation of category combinations as opposed to raw counts, we calculated the difference between the observed count and the count expected by the marginal probabilities of the memory-guided and visual search clusters, then normalize by that expected count. To account for differences in category size and marginal probabilities, we shuffled category assignments 1000 times and recomputed the signed χ² for each combination and iteration. This provided a distribution of signed χ² values for each category combination when the structured relationship between tasks was destroyed. From these distributions, we calculated a bootstrapped Z-score. This Z-score quantifies whether a particular combination was as common as expected from incidental combinations with no structure (Z ≈ 0), more common (Z > 0), or less common (Z < 0) than expected by chance.

Results

Neuron Types

To compare neuronal modulation in the 2 eye fields, neurons were classified according to FEF criteria observed during memory-guided saccades (Schall 1991; Cohen et al. 2009). This task creates the same testing conditions as Bruce and Goldberg's (1985) learned saccade task and has been used by many others (e.g. Hikosaka and Wurtz 1983). Specifically, based on increased activity during the visual, presaccadic, both, or neither period of modulation, neurons were classified as visual, movement, visuomovement, or unclassified. The average SDFs of the neurons of each type in each region are shown in Fig. 3.

Figure 3.

Traditional neuron classes. FEF neurons are frequently categorized as visual, visuomovement, or movement neurons. The present sample of neurons from each area was categorized according to these criteria. The mean ± standard error of mean (SEM) of the normalized SDF aligned on target onset (left) and saccade (right) for the neurons are plotted in red for FEF and blue for premotor cortex. Visual neurons are in the top row, visuomovement neurons in the second row, movement neurons in the third row, and uncategorized neurons in the bottom row. Number of neurons in each category and their proportions are shown. On the right, proportions of each neuron type in each area are shown.

In FEF, 32.2% of neurons were classified as visual, 16.9% as purely movement, 35.0% as visuomovement, and 16.0% were unclassified. In F2vr, 25.4% of neurons were classified as visual, 23.0% as purely movement, 24.2% as visuomovement, and 27.5% were unclassified. These proportions were significantly different across cortical areas (Contingency test χ²(3) = 21.4, P < 0.001). Relative to FEF, in F2vr, fewer neurons had visual responses and more neurons had exclusively presaccadic activity or were unclassified. This was true for monkey Da (χ²(3) = 12.9, P = 0.005) and Ga (χ²(3) = 13.5, P = 0.004), but not for He (χ²(3) = 0.9, P = 0.397).

Neuronal Modulation Timing

Given that neurons with visual, delay, and presaccadic activity are present in samples from both areas, we measured visual latency, magnitude of delay period activity, beginning of presaccadic activity, and time of peak presaccadic activity from neurons in each region during memory-guided saccades. Cumulative distributions of these values are shown in Fig. 4. Across areas, the latencies of visual responses to stimuli in the preferred locations in the memory-guided saccade task were effectively indistinguishable (Fig. 4A). In FEF, the median response latency of neurons with visual activity was 51 ms. In F2vr, the median response latency of neurons with visual activity was 55 ms. The distributions of response latencies were not significantly different (Wilcoxon rank-sum Z = −1.29, P = 0.198).

Figure 4.

Neuronal characteristics during memory-guided saccades. Cumulative distributions of visual response latency (A), time at which presaccadic movement activity begins (B), time of the peak of movement activity (C), normalized magnitude of delay period activity when the target was in the RF (D), normalized magnitude of delay period activity when the target was out of the RF (E), and the difference in delay period activity between the in and out of RF conditions (F) with medians shown as vertical dashed lines for FEF (red) and F2vr (blue). Values of test statistics and P values associated with tests are shown inset. In (B), the time at which presaccadic movement activity begins, inferences differed for different monkeys, so statistics for individual monkeys is also shown.

The median onset of presaccadic activity before saccades toward the preferred locations was 88.5 ms in FEF and 68.0 ms before saccades toward the preferred locations in F2vr. The distributions were significantly different (Fig. 4B; Wilcoxon rank-sum Z = −5.52, P < 0.001). However, this was only true for one of the monkeys, monkey Ga (Z = −2.22, P = 0.027), but not Da (Z = −1.25, P = 0.213) or He (Z = −0.57, P = 0.571), so this finding should be taken with caution.

The median time when presaccadic activity peaked relative to saccade initiation in FEF was −1.0 ms and in F2vr was −5.0 ms. The distributions were not significantly different (Fig. 4C; Wilcoxon rank-sum Z = 1.39, P = 0.166).

The mean delay period activity in FEF when the singleton was in the preferred locations was 0.01 AU, and that in F2vr was −0.02 AU. The distributions were not significantly different (Fig. 4D; Wilcoxon rank-sum Z = 1.43, P = 0.153). However, the mean delay period activity in FEF when the singleton was not in the preferred locations was −0.18 AU, and that in F2vr was −0.37 AU. The distributions were significantly different (Fig. 4E; Wilcoxon rank-sum Z = 2.12, P = 0.034). Combining these measures, the magnitude of selectivity during the delay period in FEF was 0.29 AU, and in F2vr was 0.37 AU. These distributions were not significantly different (Fig. 4F; Wilcoxon rank-sum Z = −0.40, P = 0.689).

Spatial Characteristics of Neuronal Modulation

The previous section described the temporal characteristics and delay period activity of FEF and F2vr neurons. Few major differences were found. Another dimension that could differentiate the 2 areas is the spatial tuning of neuronal responses. Microstimulation of FEF preferentially generates contraversive saccades, and most neurons in FEF are more active after visual stimuli appear in the contralateral hemifield and in association with contraversive saccades (Bruce et al. 1985; Bruce and Goldberg 1985; Schall 1991, 2015). The spatial tuning of FEF neurons during the visual and saccade epochs generally overlaps, though visual neuron receptive fields tend to be smaller than saccade neuron movement fields (Bruce and Goldberg 1985; Khanna et al. 2020). Similarly, microstimulation of the dorsal premotor cortex preferentially generates contraversive saccades (Fujii et al. 2000), and neuronal responses tend to be greatest in association with contralateral gaze shifts (Hoshi and Tanji 2004; Savaki et al. 2015). However, neurons in area F2 do respond for ipsilateral movements, albeit with smaller magnitude and longer latency (Hoshi and Tanji 2004). Also, neurons in area F5 are active for both contralateral and ipsilateral saccades (Neromyliotis and  Moschovakis 2018). With this background, we compared the distributions of preferred visual stimulus and saccade directions in FEF and F2vr.

To determine the preferred stimulus/saccade direction and associated tuning width, we used a vector averaging approach and a von Mises fit, respectively, to neural spike rates during visual (50–150 ms after stimulus presentation) and saccade (from 75 ms before to 25 ms after saccade initiation) epochs. Preferred direction was calculated by first determining the average spike rate for each stimulus location, and then calculating the direction of the vector average of these per-location responses. Tuning width was first quantified as the full width at half-max for the fit von Mises distribution, which was then converted to visual field span using the law of cosines based on the eccentricity of the stimuli. During the visual epoch, we found in FEF a statistically significant concentration of contralateral preferred stimulus direction in FEF (183.9° ± 4.3°; Rayleigh Z = 22.87, P < 0.001) and in F2vr (184.5° ± 6.1°; Rayleigh Z = 10.77, P < 0.001). These distributions were not statistically different (Fig. 5A; Watson-Williams F_1,688 = 0.030, P = 0.873). The receptive field width in FEF (11.3° ± 0.2°) was not significantly different from that in F2vr (11.6° ± 0.2°) (Fig. 5B; Wilcoxon rank-sum Z = 0.69, P = 0.489). For comparison with a recent description of the superior colliculus (Hafed and Chen 2016), we compared receptive field size of neurons with receptive fields in the upper and lower visual fields in both areas and found no differences. Thus, the visual responses of neurons in FEF and F2vr are indistinguishable in spatial tuning characteristics.

Figure 5.

Spatial characteristics of neurons during memory-guided saccades. (A) Polar histograms of direction of center of visual receptive fields calculated from response 50–150 ms after array presentation for FEF (red) and F2vr (blue) in 10° bins. Rayleigh’s test for circular nonuniformity was conducted, and P values for FEF and F2vr are written in red and blue, respectively. The P value for a Watson-Williams test comparing the 2 distributions is written in black. (B) Cumulative distributions of the width of the receptive fields in the upper (thick) and lower (thin) visual fields in FEF and F2vr. The P value for a Wilcoxon rank sum test of equal means is shown in black. (C) Polar histograms of preferred directions of saccade movement fields, calculated from spike rates 75 ms before until 25 ms after saccade initiation, for FEF and F2vr. Conventions as in (A). (D) Cumulative distributions of width of movement fields in the upper (thick) and lower (thin) visual fields in FEF and F2vr. The P value for a Wilcoxon rank sum test of equal means is shown in black. (E) Polar histograms of the difference in direction of the centers of visual receptive and saccade movement fields, where equivalent preferred locations are to the right and opposite preferred locations are to the left. The P values for a V test of nonuniformity centered on 0° for FEF and F2vr are shown in red and blue, respectively. Receptive and movement field centers varied in direction but fell in the same hemifield. (F) Distributions of differences in tuning width are shown for FEF and F2vr. Negative values indicate narrower tuning in the visual epoch and positive values indicate a narrower tuning in the movement epoch. The P values from paired t-tests comparing visual and movement tuning widths for FEF and F2vr are written in red and blue text, respectively. The P value from a two-sample t-test comparing response field width differences between FEF and F2vr is written in black.

During the saccade epoch, we found in both FEF and F2vr most neurons preferring horizontal saccade. In FEF, a statistically significant concentration of neurons had contraversive preferred directions (172.2° ± 4.6°; Rayleigh Z = 4.37, P = 0.013). In F2vr, on the other hand, neurons with contraversive and ipsiversive preferences were balanced (188.2° ± 6.7°; Rayleigh Z = 2.36, P = 0.095), and so significantly different from the distribution in FEF (Fig. 5C; Watson-Williams F_1,688 = 42.26, P < 0.001). The movement field widths in FEF (11.2° ± 0.2°) were not significantly different from those in F2vr (10.0 ± 0.33) at the median (Fig. 5D; Wilcoxon rank-sum Z = 0.00, P = 0.998). However, at the distributional level, the movement field widths were significantly different (Kolmogorov–Smirnov KS = 0.24, P = 0.015). Movement fields occupying a quadrant or less were smaller in F2vr than those in FEF. For comparison with a recent description of the superior colliculus (Hafed and Chen 2016), we compared movement field size of neurons with receptive fields in the upper and lower visual fields in both areas and found no differences.

We next compared the spatial tuning measured in the visual and measured in the saccade epochs. The direction of the best visual and saccade responses overlapped with a mean difference of 0° in FEF (V-test V = 116.72, P < 0.001) and in F2vr (V-test V = 45.14, P < 0.001). However, the variation in best direction in the 2 epochs was significantly broader in F2vr relative to FEF (Fig. 5E; Watson-Williams F_1,688 = 10.61, P = 0.001). Although visual receptive field width and saccade movement field width were not statistically different in FEF (t₇₆ = −1.48, P = 0.143) nor in F2vr (t₂₅ = 1.2, P = 0.242), the sharpening of field widths in the movement epoch relative to the visual epoch trended toward being more pronounced in F2vr than in FEF (t₁₀₁ = −1.92, P = 0.058).

In sum, the spatial tuning properties of FEF and F2vr neurons are similar but not identical at all times. Visual receptive fields are similar sizes and concentrations in the contralateral hemifield. Saccade-related movement fields are also of similar size, but whereas contraversive movement fields were most common in FEF, no laterality was found in F2vr. These findings are all consistent with the previous literature concerning direction preferences in FEF and F2vr.

Modulation During Visual Search

The characteristics of FEF neurons during visual search tasks have been described (Thompson and Bichot 2005; Bisley 2011; Schall 2015). We obtained parallel measures of the modulation of neurons in F2vr during a visual search task. Monkeys earned fluid reward for shifting gaze from central spot to a singleton T among Ls or a singleton L among Ts (see Methods). The visual latencies of these neurons were comparable to the latencies observed in the memory-guided saccade task; median visual latency of FEF neurons was 50 ms and of F2vr neurons was 47 ms, and the distributions were not significantly different (Wilcoxon rank-sum Z = −0.342, P = 0.7323). In addition to the nonspecific visual response, FEF neurons exhibited the frequently reported target selection; responses initially did not discriminate target from distractors but eventually became greater when the singleton shape was in the preferred locations and reduced when a distractor was in the preferred locations (Fig. 6A). Surprisingly, F2vr neurons also exhibited target selection during visual search. This was observed in a substantial sample of the neurons in F2vr, albeit a statistically significantly lower fraction than in FEF (F2vr n = 134, 48.2%; FEF n = 219, 58.6%; Contingency test χ²(1) = 6.89, P = 0.009).

Figure 6.

Neuronal modulation during visual search. (A) Mean ± SEM of SDFs of neurons that select the location of a visual search target in FEF (red, left) and premotor cortex (blue, right). Saturated colors indicate trials with the search target in the RF and desaturated colors with a nontarget distractor within the RF (insets). Cumulative distributions of TST are inset between −.5 and + 0.5 AU with a vertical arrow indicating median TST. Neuron count and proportion of neurons within area are labeled. (B) Cumulative distributions of TST in FEF (red) and premotor cortex (blue). TST were effectively identical across areas. (C) Distribution of TST differences in simultaneously recorded pairs of neurons in FEF and premotor cortex. By convention, when an FEF neuron precedes a premotor cortex neuron, the difference is negative. The mean difference is shown as a vertical red line.

Of the neurons that did exhibit target selection, the timing of selection was statistically indistinguishable between the 2 regions (Fig. 6B). The median TST in FEF was 118.5 ms (mode = 106 ms) and in F2vr was 111.0 ms (mode = 101.8 ms). The distributions were not significantly different (Wilcoxon rank-sum Z = 1.08; P = 0.280).

The interpretation of this observation must be qualified by appreciating how the times of saccade target selection vary across sessions in proportion to variation of response times. To control for such extraneous factors, we compared directly the TST of neurons recorded simultaneously in FEF and F2vr. The distribution of differences in TST for simultaneously recorded neurons in the 2 regions is shown in Fig. 6C. The mean difference in TST (TST_FEF—TST_F2vr) between such neuron pairs was −1.17 ms (mode = 0.25), which was not significantly different from zero (t₁₃₅ = −0.40, P = 0.689). Thus, saccade target selection arises simultaneously in FEF and in F2vr.

We next assessed whether the magnitude of the modulation signaling target selection was similar in the 2 regions. The average neuronal response 150 to 200 ms after array presentation was calculated for each neuron when the target was in (r_in) or outside (r_out) of the preferred location. A modulation index was defined as the difference in the 2 trial conditions divided by their sum, (r_in − r_out)/(r_in + r_out). The median selection index for FEF neurons (0.134) was greater than that for F2vr neurons (0.094), but the distributions of the modulation indices were not significantly different (Wilcoxon rank-sum Z = 1.62; P = 0.104). Thus, although a lower fraction of F2vr neurons relative to FEF neurons signal the location of the target during visual search, they do so at the same time and to the same degree.

Consensus Clustering

To perform an unbiased comparison of the diversity of neuron types within FEF and F2vr, we applied a consensus clustering algorithm to the mean SDFs of each neuron in both areas (Lowe and Schall 2018). If the outcome of the clustering algorithm distinguishes the areas, this would be evidence that the areas are populated by neurons with different patterns of modulation. Alternatively, if the outcome of the clustering algorithm does not distinguish the areas, this would be evidence that the areas are populated by neurons with common patterns of modulation. This method has not been applied to data from >1 cortical area.

We employed the algorithm by including modulation patterns when the target was both within and outside of the preferred location. To apply consensus clustering across cortical regions, we will explore multiple analysis pipelines. The first approach will determine clusters of units combined across areas. This will reveal the similarities of functional neuron categories across the 2 areas and assesses the relative proportion attributed to each category. The second approach will determine clusters of units for each area separately and then use a classifier to assign units from the 1 area to the categories identified in the other. This will both refine the categories for each region and assess the similarities and differences across areas.

Clustering Across Areas

To quantify functional similarity across cortical areas, we performed consensus clustering for units combined across areas (Fig. 7A). To disambiguate results from different tasks, categories are labeled with subscripts for the task used: MG for memory-guided saccade task and VS for shape singleton visual search task.

Figure 7.

Consensus clusters for both regions. A consensus clustering algorithm was applied to identify different categories of neurons in FEF and premotor cortex. (A) Schematic of recording and clustering pipeline. Recordings were performed from both FEF (red electrode) and premotor cortex (blue electrode). Clustering was applied to the sample of neurons across regions. (B) Clustering results are shown for the memory-guided saccade task. Category numbers are arbitrarily assigned on a roughly visual to motor spectrum. Mean SDFs for target in (thick) and out of (thin lines) RF are shown for FEF on the left (red) and premotor cortex on the right (blue). SDFs are aligned on target onset (left column of each pair) and saccade (right column of each pair). Number and proportion within area are labeled for each category and area combination. Significantly different proportions of categories between areas are indicated with a horizontal bar and asterisk. The region with the larger proportion of that category is indicated with a bold, italic number, and proportion text. Category proportions are also shown to the right as bar plots, with proportions of FEF neurons in red and F2vr neurons in blue. Significantly different proportions are indicated with an asterisk. (C) Clustering results are shown for the shape singleton search task. Conventions are as in (B).

During the memory-guided saccade task, the algorithm identified 6 categories (Fig. 7B), which were arbitrarily numbered accordingly roughly to the level of visual activity. Categories 1_MG and 2_MG demonstrated visual activity and not presaccadic activity. They were distinguished by the presence (1_MG) or absence (2_MG) of delay period activity and an earlier (1_MG) or later (2_MG) time of peak visually evoked activity. Category 3_MG also demonstrated visual activity with a relatively early time of peak visual activity, but was distinguished from category 1_MG by the presence of postsaccadic activity, at least in the FEF neurons in this category. Category 4_MG demonstrated visual activity, but of lower magnitude than categories 1 MG–3_MG, and pronounced presaccadic buildup activity. Categories 5_MG and 6_MG demonstrated suppression of activity after the visual stimulus and were distinguished by whether the activity remained below baseline through the trial (5_MG), or whether the activity ramps and peaks at the time of saccade (6_MG). In this way, these categories could be identified as fixation neurons and pure movement neurons, respectively.

The 6 categories were distributed unequally across regions (Contingency test χ²(5) = 25.2, P < 0.001). Post hoc contingency tests indicate that categories 1_MG, 5_MG, and 6_MG were distributed unequally across regions, whereas categories 2_MG, 3_MG, and 4_MG were found in similar proportions across regions. Category 1_MG was more prevalent in FEF than in F2vr, and categories 5_MG and 6_MG were more prevalent in F2vr than in FEF.

We also performed this analysis for neuronal responses during shape singleton search and again found 6 categories of neurons (Fig. 7C). Two categories, categories 1_VS and 2_VS, demonstrated visual responses without presaccadic responses. These were distinguished by the magnitude of visual response and the magnitude of spatial selectivity; both features were of greater magnitude in category 1_VS than category 2_VS. Category 3_VS and category 4_VS demonstrated the most spatial selectivity and were the 2 categories whose responses increased and peaked around the time of saccade. They were distinguished by whether the activity increased in time only (3_VS) or decreased briefly before ramping in a preexcitatory pause (4_VS; Sato and Schall 2001). Categories 5_VS and 6_VS were characterized by suppressed responses and were distinguished by whether this suppression was brief (5_VS) or extended through the trial (6_VS).

These categories were also distributed unevenly between the 2 regions (Contingency test χ²(5) = 21.9, P < 0.001). Post hoc contingency tests indicate that categories 1_VS and 6_VS were distributed unevenly across the 2 regions, whereas categories 2_VS, 3_VS, 4_VS, and 5_VS were distributed similarly across the 2 regions. Category 1_VS was more prevalent in FEF than in F2vr, whereas category 6_VS was more prevalent in F2vr than in FEF. These results were consistent with the between-area differences observed in the memory-guided clusters, which were also consistent with the traditional categorization.

Of the 670 neurons recorded in the memory-guided saccade task and 652 neurons recorded in the visual search task, 338 had sufficient data to analyze in both tasks, and 320 were categorized in both tasks. We found a significant relationship between the clustering categories produced from neuronal signals recorded in the 2 tasks (χ²(25) = 57.26, P < 0.001). This was true for both FEF (χ²(25) = 44.47, P = 0.001) and F2vr (χ²(25) = 39.14, P = 0.036), although the χ² value was smaller for F2vr than for FEF, indicating more heterogeneity of clustering of neurons in F2vr relative to FEF. This will be addressed more below.

To confirm the statistical significance of the observed relationships given the large number of category combinations, we performed a bootstrapping procedure by randomly shuffling category assignments for MG and VS clusters separately and recalculated the χ² test statistic. Of 1000 iterations of random shuffling, the empirical χ² statistic for FEF was greater than the value for the shuffled labels 989 times (P = 0.011). Likewise, the empirical χ² statistic for F2vr was greater than the value for the shuffled labels 979 times (P = 0.021), once more indicating more heterogeneity in F2vr than in FEF.

To assess the qualities of these associations between the MG and VS clusters, we calculated a signed χ² for each category combination and converted this value to a Z-score after 1000 iterations of randomly reassigning category labels (Fig. 8). This afforded visualization of whether particular category combinations were more common than expected by chance (Z > 0), less common than expected by chance (Z < 0), or as prevalent as expected given the marginal probabilities of each category (Z ≈ 0). As a measure of confidence in the bootstrapped Z-score, we assessed the expected counts of neurons. When observed counts are fewer, small differences in neuron counts can deviate more from expectation—a 1 neuron increase constitutes a 100% change if only 1 neuron is expected but only a 5% change if 20 neurons are expected.

Figure 8.

Cluster combinations across tasks. (A) Legend matrix and the 2 extreme possibilities for consistency of clusters across tasks. The color of each row represents the bootstrapped Z-score of a signed χ² quantifying the consistency of clustering of neurons across tasks. Observed neuron counts for common clustering across tasks were compared with the counts expected by chance given the counts observed in each cluster obtained individually from memory-guided and visual search tasks. If more neurons than expected were found in a particular cluster combination, that cell in the matrix is highlighted green. If fewer than expected were found, the cell is highlighted magenta. If just the expected number of neurons was found, the cell is highlighted black. The cross-hatch of each column represents the number of neurons contributing from each area to each cell of the matrix, with FEF counts indicated by right-slanted hatches, and F2vr counts indicated by left-slanted hatches. These hatches signify the confidence in interpretability of each combination. The confidence in interpretability of each combination is proportional to the number of neurons in a given combination. If few neurons are expected, a small difference is registered as a large deviation. If many neurons are expected, a small difference is registered as a smaller deviation. Cells in the matrix associated with larger counts appear more saturated. The outcomes of this analysis are bounded between 2 extremes. The first is a 1:1 mapping between cluster assignments identified during the memory-guided saccade task and the visual search task (lower left matrix). A 1:1 mapping is signaled by only 1 column per row (or row per column) highlighted green and all remaining columns (or rows) highlighted magenta. The second is a random mapping between cluster assignments (lower right matrix), which is signaled by black highlight in all cells. (B) Cross-task clustering combinations for FEF. The form of the matrix indicates a result intermediate between the 1:1 and the random mappings. To assist in visualizing the relationships, the average SDFs for the particular MG and VS clusters are redrawn from Fig. 7. (C) Cross-task clustering combinations for F2vr. The form of the matrix indicates a result less like the 1:1 mapping and more like the random mapping. Details in text.

To contextualize this analysis and the visualization of the results in this particular sample of neurons, we demonstrate the 2 extreme alternative outcomes: a perfect 1:1 mapping and complete random independence. First, though, to develop the intuition for this analysis approach, which is to our knowledge entirely novel, consider 2 numerical categorizations of people: alphabetical rank by last name and numerical rank by street address number. If we assume that all people in a household have the same last name and that all last names are unique, then we would expect a 1:1 mapping between categorization according to name and categorization according to address. That is, the people with last name rank 1 (e.g. Aarons) all live at address rank 1 (e.g. 101 Main), and the people with the last name rank (e.g. Zimmerman) all live at the last address rank (e.g. 999 Main). However, these rankings need not be in the same order. The Zimmermans could live at 102 Main, and the Aarons at 998 Main. Thus, the order of the list of last names need not correspond to the order of the list of addresses. Nevertheless, because all last names are unique, there is a 1:1 mapping between last names and addresses. However, a 1:1 mapping would not be observed if people with different last names live at one address. In this case, there is neither an ordinal relationship between rank values nor a predictive relationship between names and addresses. Thus, because cluster numbers are arbitrary, a 1:1 mapping does not entail that membership in cluster 1_MG necessarily equates to membership in cluster 1_VS. Rather, it entails that membership in cluster 1_MG equates to membership in 1 and only 1 VS cluster. A strict 1:1 relationship would describe the interdependencies of all rows and columns. In a weaker but not random relationship, the category relationships are not commutative but are instead directional. If the Browns move in with the Aarons at 101 Main, then both names are uniquely mapped to an address, but that address maps to 2 names. The most restrictive sense of 1:1 mapping, unique relationships between all rows and columns, sets an implausibly high bar for these particular data. Empirically, neurons in FEF and F2vr follow an intermediate pattern, so results will be described in the less restrictive sense of 1:1 mapping. We found some relationships consistent with 1:1 mapping in which category identity in a row is associated with 1 and only 1 category identity in a column, or vice versa. We also found violations of strict 1:1 mapping in which category identity in a row is associated with >1 category identity in a column, and vice versa.

To illustrate a strict 1:1 mapping between MG and VS clusters, we generated the Z-scored signed χ² plot after equating all of the VS cluster and MG cluster IDs (Fig. 8A, lower left). Then, to highlight that the numerical category value is arbitrary, category labels were shuffled for the VS cluster IDs. In each row, 1 column contains a category combination occurring significantly more common than expected by chance (Z > 0, green). All other category combinations were significantly less common than expected by chance (Z < 0, magenta).

To illustrate random mapping, we reassigned both MG and VS categories according to the marginal probabilities associated with each category observed in both tasks. For example, 35.6% of neurons belonged to category 1_MG, so 114 neurons (35.6% of the 320 neurons recorded in both tasks) were reassigned to category 1_MG. Of these, 23 were reassigned to category 1_VS and 15 were reassigned to category 2_VS. This corresponds to 20% and 13.4% of the 114 neurons in category 1_MG, respectively, dictated by 20% and 13.4% of the full sample belonging to categories 1_VS and 2_VS. Similarly, 7.5% of the full sample belonged to category 2_MG, so 24 neurons (7.5% of 320) were reassigned to category 2_MG, and of these 5 and 3 (20% and 13.4% of 24) were reassigned to categories 1_VS and 2_VS. Because these reassignments were constrained by the respective marginal means of MG and VS clusters, the resulting combinations are exactly as prevalent as expected by chance (Z = 0, black), both on and off the diagonal (Fig. 8A lower right).

Considering first FEF, this analysis demonstrates nonrandom relationships between some MG and VS clusters (Fig. 8B). Two positive associations were clearest: 1_MG = 1_VS and 3_MG = 3_VS. Also, 6_MG = 6_VS, but with few neurons. In other words, the modulation patterns of these neurons did not change across tasks. The units identified as 1_MG = 1_VS consist of visually responsive neurons without movement-related activity before memory-guided saccades with minimal visual search target selectivity. The units identified as 3_MG = 3_VS produced spatially selective visual responses and spatially selective postsaccadic activity. The units identified as 6_MG = 6_VS had suppressed visual responses and postsaccadic activity. Many associations were lower than expected by chance consistent with a 1:1 mapping. These include 1_MG ≠ 5_VS, 1_MG ≠ 6_VS, 2_MG ≠ 1_VS, 3_MG ≠ 1_VS, 4_MG ≠ 6_VS, 6_MG ≠ 1_VS, and 6_MG ≠ 3_VS. These dissociations are easy to understand upon examination of the SDF of each pair. Counterexamples to a 1:1 mapping were also found. The association 4_MG = 3_VS exceeded chance because both exhibited spatially selective visual response and saccade-related activity. Similarly, 5_MG = 6_VS perhaps because of the postsaccadic suppression. These counterexamples illustrate a many:1 mapping, where categories 3_MG and 4_MG are both associated with category 3_VS, and categories 5_MG and 6_MG are both associated with category 6_VS. The remaining pairwise associations were not different from that expected by chance. Overall, FEF neurons that are in the same cluster in 1 task tend to be clustered together in the other task as well.

Considering next F2vr, in general the mapping between MG and VS clusters is much less pronounced as indicated by less extreme positive and negative Z-scores relative to those for FEF. Two positive associations consistent with 1:1 mapping were evident in F2vr, 1_MG = 1_VS and 4_MG = 3_VS, though both associations were weaker than comparable associations in FEF. Also consistent with the 1:1 mapping were the following negative associations: 1_MG ≠ 6_VS, 4_MG ≠ 2_VS, 4_MG ≠ 5_VS, 5_MG ≠ 4_VS, 6_MG ≠ 1_VS, and possibly 6_MG ≠ 3_VS. In F2vr relative to FEF, more of the associations of clusters across tasks were at random levels. In general, these novel findings indicate that particularly pronounced activity patterns in FEF remain distinct across tasks, but the consistency of responses across tasks was weaker in F2vr.

Clustering Within Areas

To compare and contrast further the functional neuron types within and across FEF and F2vr, we employed a different approach consisting of the following steps: (a) apply consensus clustering to neurons sampled in 1 area, (b) train a classifier on the categories, (c) create composite categories by assigning neurons from the other area to the categories identified by the classifier, and (d) evaluate the similarities and differences. This method offers a different perspective on whether specific neuronal response profiles in 1 area exist in the other. To disambiguate results, we introduce a nomenclature whereby categories are labeled with superscripts for the area to which clustering was applied and subscripts for the task used. Thus, a neuron assigned to category 1 in a clustering of FEF neurons during memory-guided saccades will be written as 1^FEF_MG.

Consider first the assignment of neurons sampled in F2vr by a classifier trained on consensus clusters derived from FEF (Fig. 9). Among FEF neurons, during memory-guided saccades, 6 categories were identified (Fig. 9B). The composite categories assigned by the classifier were distributed unequally across FEF and F2vr (Contingency test χ²(5) = 25.6, P < 0.001). Post hoc contingency tests indicated that this difference was driven by the distributions of categories 1^FEF_MG, 2^FEF_MG, 3^FEF_MG, 5^FEF_MG, and 6^FEF_MG. Categories 1^FEF_MG and 2^FEF_MG both demonstrated visual responses without presaccadic activity and were distinguished by the presence of delay period activity (1^FEF_MG) or of postsaccadic activity (2^FEF_MG). Notably, the former was more prevalent in FEF than in F2vr, whereas the latter was more prevalent in F2vr than in FEF. Two categories of visuomovement neurons, 3^FEF_MG and 4^FEF_MG, were identified, but only 3^FEF_MG was more common in FEF than in F2vr. These categories were distinguished by the time course of the return to baseline after the saccade, being slower in category 3^FEF_MG and faster in category 4^FEF_MG, reminiscent of clipped movement-related neurons in superior colliculus (Waitzman et al. 1991) and FEF (Segraves and Park 1993; Lowe and Schall 2018). Another category, 5^FEF_MG, had clipped movement-related responses but little or no visually evoked activity. This category was more prevalent in F2vr than in FEF. Finally, category 6^FEF_MG, characterized by prolonged suppression, was more prevalent in F2vr than in FEF (see category 5_MG, category 5_VS, Fig. 7B–C).

Figure 9.

Assigning F2vr neurons to consensus clusters of FEF. (A) Schematic of sampling, clustering, classifier training, and evaluation pipeline. Neurons sampled in FEF (red electrode) underwent consensus clustering. A classifier was trained on the resulting clusters. Then, the classifier assigned neurons sampled in F2vr (blue electrode) to clusters. These were used to create a composite set of cluster IDs across both regions. (B) Clustering results are shown for the memory-guided saccade task. Conventions as in Fig. 7B. (C) Clustering results are shown for the shape singleton search ask. Conventions as in Fig. 7C.

Among FEF neurons, during the visual search task, 6 categories were identified (Fig. 9C). The composite categories assigned by the classifier were distributed unequally across FEF and F2vr (Contingency test χ²(5) = 166.9, P < 0.001). Post hoc contingency tests indicated that this difference was driven by the distributions of categories 1^FEF_VS, 2^FEF_VS, 3^FEF_VS, 4^FEF_VS, and 6^FEF_VS. Categories 1^FEF_VS and 2^FEF_VS both exhibited visual responses that did not distinguish target from distractor and no presaccadic activity. The 2 categories were distinguished by an earlier, larger (1^FEF_VS) compared with later, smaller (2^FEF_VS) visual response. The former was more prevalent in FEF, and the latter in F2vr. Category 3^FEF_VS comprises visually responsive neurons that distinguish the target from a distractor before the saccade. Consistent with the findings presented in Fig. 5, this category was much more prevalent in FEF than in F2vr. Category 4^FEF_VS, with a weak and nonselective visual response followed by no discharge or suppression until the saccade, more common in F2vr. This prevalence could arise by this category being populated by units that were unmodulated or did not fit in any of the FEF-based categories. Category 5^FEF_VS, exhibiting suppression persisting from after array presentation until after the saccade, was equally common in both areas. Finally, category 6^FEF_VS comprises neurons with no visual response and pronounced saccade-related activity. According to this analysis approach, this category was more prevalent in FEF than in F2vr, contrasting with the data shown in Fig. 3.

Consider next the assignment of neurons sampled in FEF by a classifier trained on consensus clusters derived from F2vr (Fig. 10). Among F2vr neurons, during memory-guided saccades, 6 categories were identified (Fig. 10B). The composite categories assigned by the classifier were distributed unequally across FEF and F2vr (Contingency test χ²(5) = 25.6, P < 0.001). Post hoc contingency tests indicated that this difference was driven by the distributions of categories 2^PM_MG, 4^PM_MG, and 5^PM_MG. Three categories were identified with visual responses and no presaccadic activity, 1^PM_MG, 2^PM_MG, and 3^PM_MG. They were distinguished by the presence (1^PM_MG, 2^PM_MG) or absence (3^PM_MG) of delay period activity, presence (3^PM_MG) or absence (1^PM_MG, 2^PM_MG) of postsaccadic suppression, and later (1^PM_MG) or earlier (2^PM_MG, 3 ^PM_MG) suppression relative to saccade initiation. Of these categories, only 2^PM_MG was differentially distributed across areas, being more prevalent in FEF. Two categories, 4^PM_MG and 5^PM_MG, exhibited weak visual responses and more pronounced presaccadic activity. They were distinguished by pronounced (4^PM_MG) or diminished (5^PM_MG) visually evoked activity and an earlier (5^PM_MG) or later (4^PM_MG) return to baseline after the saccade. Category 4^PM_MG was more prevalent in F2vr, and category 5^PM_MG was more prevalent in FEF. Finally, category 6^PM_MG, exhibiting a brief visual response followed by sustained suppression, was found in equal proportions across regions.

Figure 10.

Assigning FEF neurons to consensus clusters of F2vr. (A) Schematic of sampling, clustering, classifier training, and evaluation pipeline. Neurons sampled in F2vr (blue electrode) underwent consensus clustering. A classifier was trained on the resulting clusters. Then, the classifier assigned neurons sampled in FEF (red electrode) to clusters. These were used to create a composite set of cluster IDs across both regions. (B) Clustering results are shown for the memory-guided saccade task. Conventions as in Fig. 7B. (C) Clustering results are shown for the shape singleton search ask. Conventions as in Fig. 7C.

In F2vr during visual search, seven categories were identified (Fig. 10C). The composite categories assigned by the classifier were distributed unequally across FEF and F2vr (Contingency test χ²(6) = 77.2, P < 0.001). Post hoc contingency tests indicated that this difference was driven by the distributions of categories 1^PM_VS, 2^PM_VS, 3^PM_VS, 6^PM_VS, and 7^PM_VS. Categories 1^PM_VS, 2^PM_VS, and 3^PM_VS exhibited visual responses with little presaccadic activity. Category 1^PM_VS exhibited the most spatial target selectivity and presaccadic activity and was more prevalent in FEF. Category 2^PM_VS exhibited a brisk visual transient and was more prevalent in F2vr. Category 3^PM_VS exhibited a weak visual response, modest discrimination of target from distractor, pronounced postsaccadic suppression, and was more prevalent in FEF. Categories 4^PM_VS and 5^PM_VS exhibited clear discrimination of target from distractor with presaccadic activity peaking at saccade initiation. They were distinguished by the presence (category 5^PM_VS) or absence (category 4^PM_VS) of a preexcitatory pause (Sato and Schall 2001). These categories were equally distributed in FEF and F2vr, but target discrimination was stronger in FEF. Categories 6^PM_VS and 7^PM_VS were defined by suppression upon array presentation and no discrimination of target from distractor or saccade direction. They were distinguished by the latency and duration of the suppression, it being earlier and briefer for 6^PM_VS and later and prolonged after the saccade for 7^PM_VS. Both of these categories were more prevalent in F2vr.

In summary, the results of consensus clustering, whether neurons are combined or separated across areas for the clustering, were internally consistent. The identified categories are generally consistent with but more diverse than the traditional classification. The results consistently demonstrate that FEF has higher proportions of visually related responses, whereas F2vr has higher proportions of putative fixation neurons with suppressed activity. Neurons with purely movement-related responses were not consistently found in 1 region more than the other, but clustering the 2 regions individually revealed different motifs of activity in the 2 regions. Relative to those in FEF, movement-related neurons in F2vr have less spatial specificity and different time courses of peak activity.

Discussion

In macaques, FEF resides within area 8, but the caudal extent of FEF defined functionally has not been described in detail. Anatomically, area 8 is defined by a granular layer 4, which gradually thins caudally until it is absent around the fundus, marking the rostral limit of area 6 (Stanton et al. 1989). Functionally, neurons deep in the arcuate sulcus contribute to smooth pursuit more than to saccadic eye movements (MacAvoy et al. 1991; Tanaka and Lisberger 2002; Izawa et al. 2009; Lixenberg and Joshua 2018), and neuronal responses just caudal to FEF in the posterior bank of the arcuate sulcus, in area 6 or premotor cortex, have marked similarities to responses in FEF in simple oculomotor tasks (Kurata 2017; Neromyliotis and Moschovakis 2017b, 2018). These indiscrete anatomical and functional transitions call attention to the ambiguous caudal boundary of FEF. Still, the prefrontal and premotor regions are not identical, as saccade-related neurons in F5 discharge before FEF neurons, are more likely than FEF neurons to respond to ipsilateral stimuli, and respond before forelimb as well as eye movements, e.g. Neromyliotis and Moschovakis 2018. Here, we show that these similarities extend to a more complex task, a singleton shape visual search task, in the ventral portion of the dorsal premotor cortex (F2vr). However, we did identify several distinguishing characteristics in both simple and complex tasks that may reveal the functional distinctions between these 2 regions. And although the assignments of function can be formulated by computational means (e.g. Boucher et al. 2007; Purcell et al. 2010, 2012), the approach of this study was decidedly data driven, focused on whether these regions exhibit identifiable differences and to understand the types of neurons present in the 2 regions. To that end, we will first discuss the subdivisions of premotor cortex and the targeting of F2vr specifically to compare to FEF. Next, we discuss similarities between the 2 cortical regions. Then, we will discuss the differences and their potential for distinguishing functional specificity. Finally, we will discuss the relationship between these regions and the putative homologues in humans.

Subdivisions of Premotor Cortex

Premotor cortex is an anatomically and functionally diverse area, and its subdivisions have important implications for functional relationships to other areas. Primarily, the immediate postarcuate regions of the premotor cortex can be divided into dorsal (PMd or F2) and ventral (MPs or F4 and F5) divisions, in addition to F7 (which contains SEF), based on cytoarchitecture (Matelli et al. 1985; Geyer et al. 2000; Luppino et al. 2003) and connectivity (Ghosh and Gattera 1995; Tanné-Gariépy et al. 2002; Luppino et al. 2003). These anatomical subdivisions correspond with functional differences in visuomotor behavior (see Hoshi and Tanji 2007 for review). Specifically, PMd seems to be related more closely to dynamics and directions of motions as well as stimulus–response rule associations (Riehle and Requin 1989; Wallis and Miller 2003; Romo et al. 2004; Cromer et al. 2011; Yamagata et al. 2012; Coallier et al. 2015), whereas PMv seems to be more closely related to the visual information instructing a movement (Boussaoud and Wise 1993a, 1993b; Hoshi and Tanji 2000, 2004). Furthermore, within PMd, visual responses are preferentially located in its rostrolateral extent, corresponding to F2vr (Fogassi et al. 1999), and rostral PMd is associated with more cognitive functions, whereas caudal PMd is associated with skeletal motor functions (Abe and Hanakawa 2009; Nakayama et al. 2016). Thus, to identify whether the posterior bank of the arcuate sulcus is similar to FEF with respect to cognitive and visuomotor functions, we specifically targeted F2vr due to its putative cognitive specialization and action selection responses. This targeting was possible due to our use of MRI-guided electrode penetrations to ensure recordings were in this small, specific target region (Fig. 2).

Effector Specificity of FEF and F2vr

We compared the relative contributions of these 2 areas to gaze tasks. More would be learned by also testing with forelimb movement tasks. This is an important consideration given that FEF is not identified with the skeletomotor system, whereas F2vr, and premotor cortex in general, is typically identified with the skeletomotor system and investigated with forelimb movement tasks either in isolation or in combination with gaze movement. Here we consider the limitations in interpretation entailed by the lack of skeletomotor task.

Premotor cortex is generally considered to be involved in manual responses (Wise 1985; Wise et al. 1992, 1996; Kalaska et al. 1997; Kalaska et al. 1998; Cisek and Kalaska 2005; Thura and Cisek 2014; Neromyliotis and Moschovakis 2017b; Neromyliotis and Moschovakis 2018). Yet, neurons in premotor cortex have motor responses that are modulated by eye position (Boussaoud et al. 1993; Boussaoud 1995; Mushiake et al. 1997; Boussaoud et al. 1998; Cisek and Kalaska 2002). In contrast, FEF is generally considered to be involved in eye movements, with unresponsiveness or suppression during manual tasks (Thompson et al. 2005; Lawrence and Snyder 2009; Kurata 2017). And although some FEF neurons are modulated in association with hand movements per se (Kurata 2017), most evidence suggests that FEF neurons are not producing hand movements but rather are modulated by hand position (Thura et al. 2008, 2011), if they are modulated at all (Mushiake et al. 1996). Thus, premotor cortex and FEF seem more specifically involved in hand and eye movements, respectively, but are both modulated by the position of the other effector. However, relative to the number of neurons in FEF active during hand movements, more neurons in premotor cortex are active during eye movements, particularly in ventral premotor cortex and the specific rostroventral region of the dorsal premotor cortex that we targeted (Kurata 2017). This finding motivated our study of F2vr in cognitively demanding saccade tasks.

The specialization of cortical areas for effectors may explain several of the key differences between regions that we identified. The most robust finding in this study is the difference in proportions of neuron types between the 2 regions. We found this in the traditional classification scheme in which visual and visuomovement neurons were more common in FEF, whereas movement-related and unmodulated neurons were more common in premotor cortex (Fig. 3), in the proportion of target selective neurons in visual search (Fig. 6), and in the differences in proportions of categories in the cluster analyses (Figs. 7–10). These differences may be explained by the exclusive use of saccade tasks. If we had tested during a manual response task, we may well have seen more visually responsive neurons in F2vr. In FEF, visual responses are enhanced when a stimulus is a potential saccade target (Goldberg and Bushnell 1981; Thompson et al. 1997), so perhaps some of the unmodulated neurons would have been modulated if the visual stimuli were potential reach targets. Similarly, perhaps some of the spatially nonselective F2vr neurons in the visual search task may have become spatially selective if the targets required a reach response. Previous research as described targets selection in dorsal premotor cortex before reaching responses during visual search (Song and McPeek 2010).

Effector specialization differences may also explain the differences in correspondence between neuron clusters found during memory-guided saccades and visual search between the 2 regions. If the exclusive use of saccade tasks prevented some of the F2vr neurons from being optimally active, then perhaps these neurons belong to their respective clusters less strongly. The associations between weakly defined cluster membership would naturally confuse the associations between the 2 tasks. This seems likely, as some of the differences between FEF and F2vr associations across tasks are attributable to differences in responses within cluster and task across regions. For example, FEF neurons in cluster 3_MG have stronger postsaccadic activity and less spatial selectivity than F2vr neurons in cluster 3_MG which are strongly spatially selective and whose movement-related activity peaks before the saccade. Other categories (e.g. category 4^FEF_VS and category 5^FEF_MG) appear to absorb unmodulated F2vr neurons and also weaken specific associations between the clusters found in the 2 tasks. Unfortunately, the consensus clustering algorithm, while producing robust clusters, is highly abstracted and the agglomerative procedure is nonparametric, making it difficult to define a metric of cluster membership confidence to limit the analyses to central cluster members.

Our approach emphasizes distinctions between categories of neurons. This perspective has proven essential to understand the organization of the visual pathway (Stone 1983) and animates current research (Zeng and Sanes 2017). We anticipate that the functional differences described here will eventually map onto morphological, biophysical, and connectivity differences. Another perspective is referred to as mixed selectivity (Fusi et al. 2016). Our analysis of the variable consistency of cluster assignments across the visual search and memory-guided saccade tasks endorses both perspectives. On the one hand, many visually responsive neurons were members of the same functional cluster in both tasks. On the other hand, other neurons were assigned to different clusters in the 2 tasks. The consistency of cluster assignment was greater in FEF than in F2vr. This observation is somewhat inconsistent with the supposition that mixed selectivity is a principle of prefrontal cortex function. A supposed advantage of mixed selectivity among neurons is the opportunity for a linear readout mechanism. However, the utility of this mechanism assumes that the downstream readout mechanism has access to signals from all sampled neurons. Wherever such a readout mechanism might be, the well-known specificity of connectivity of neurons across the cortical layers will necessarily limit the input from any cortical area to a presumptive readout mechanism. Thus, the readout mechanism can only have access to the mixed selectivity of the report-out neurons with axon terminals in the readout mechanism. Thus, a research goal would be to distinguish the report-out neurons from all others, which is, after all, just categorizing neurons. Certainly, more can be learned by examining these data through the mixed selectivity lens, but the 2 approaches are more complementary than conflicting (Dubreuil et al. 2021). In sum, the primary differences between regions found in this study, which was the proportions of neuron types, may be explained by the absence of manual responses in this investigation. This can be verified in future investigation, but this difference in effector specificity contextualizes the marked similarities between the 2 regions found in these gaze tasks.

FEF and F2vr have Similar General Characteristics

For the purpose of identifying the contributions of F2vr to visual gaze control, we recorded from neurons during a shape singleton search task. We found that neurons in F2vr exhibit target selection, as has been repeatedly described for FEF (e.g. Schall et al. 1995a; see Schall 2015 for review) and also in dorsal premotor cortex (Song and McPeek 2010). As a population, we found that TST was not measurably different between the 2 regions. Although the similarity of movement-related neurons in the FEF and ventral premotor cortex (F5) has been investigated (Neromyliotis and Moschovakis 2017b; Neromyliotis and Moschovakis 2018), we now demonstrate that the ventral portion of the dorsal premotor cortex (F2vr) has visual and movement-related responses similar to those seen in FEF.

Though the TST may be similar in the 2 regions, the speed of calculating this discrimination may differ. We found no differences between visual latencies of the 2 regions. Thus, neither the time of target selection per se nor the timing of the target selection operation are different between F2vr and FEF. Furthermore, we found that the time of onset of presaccadic activity is also not different between the 2 regions in 2 of the 3 monkeys. Thus, the visuomotor transformation occurs on similar timescales in both regions. This similarity in presaccadic activity onset is at odds with the results of Neromyliotis and Moschovakis (2018). However, their recordings were preferentially directed toward ventral premotor cortex (F5), whereas the present recordings were directed toward dorsal premotor cortex (F2vr). Because a mediolateral distinction has been observed in FEF (e.g. Suzuki and Azuma 1983; Markov et al. 2014) and premotor cortex (for review, see Hoshi and Tanji 2007), the regional differences may account for the present results. Thus, we here demonstrate visual and motor responses in an oculomotor task exist in PMd as well as in PMv.

Other metrics, specifically delay period activity, the time of the peak saccade-related activity, the spatial distributions of preferred locations, and the width of preferred location tuning curves did not differ between the areas. Based on these functional properties, the 2 regions are surprisingly comparable.

Differences Between FEF and F2vr

Though several gross metrics of visuomotor function are very similar between FEF and F2vr, we did identify several key differences between the regions. Most notably, we found a difference in the proportion of neuron types across the regions. By classifying neurons according to a canonical visual, visuomovement, and movement classification scheme (Bruce and Goldberg 1985), we found that the distribution of these neuron types was significantly different between the 2 regions. Specifically, FEF has more visual and visuomovement neurons than F2vr, whereas F2vr has more movement related and unclassified neurons. These differences were reinforced by the results of the consensus clustering identifying additional nuances in neuronal diversity. The categories identified in the memory-guided saccade task, after including responses to stimuli outside the RF, were consistent with those previously identified by Lowe and Schall (2018) which used only responses to stimuli inside the RF from these neurons, indicating robustness of the algorithm with additional conditions included in the clustering as well as the ability of this algorithm to differentiate between highly similar brain regions, and increasing confidence in this general finding.

The advantage of the consensus clustering approach is to take an unbiased approach to the modulation dynamics of the neurons, revealing additional neuron types that are generally not considered with the classic neuron categories for FEF. Specifically, the clustering approach reveals a neuron type that shows a pronounced suppression in both tasks; neurons showing this suppression were found in the 2 areas but at a higher proportion in F2vr. The consistent identification of this neuron type across regions and tasks, albeit at different proportions between regions, invites a more thorough consideration of its role in the microcircuitry. If these neurons are fixation neurons as Lowe and Schall (2018) suggest, then their increased prevalence in F2vr as opposed to FEF provides additional evidence that F2vr as an area is more closely related to movement production than visual processing, although both areas contribute to both processes.

Another advantage of the consensus clustering approach is that neurons in 1 brain region can be classified according to criteria defined by another region, and vice versa, to compare regions on equal ground. Here, for the first time we applied the consensus clustering algorithm to neurons recorded from multiple regions and apply the clustering both across the whole sample as well as within regions. This approach sheds light on the nature of the differences of proportions of neurons. Specifically, using traditional criteria we found that movement-related neurons were more prevalent in F2vr. However, when clustering was applied to the entire sample from both regions, we saw only modest hints of this pattern. When clustering was applied to FEF neurons only in the VS task, it revealed a larger proportion of pure movement-related neurons (category 6^FEF_VS). This is reconciled by applying clustering to F2vr only. In doing so, we identified another category that was not identified by any other method, category 6^PM_VS, that had presaccadic ramping activity that was spatially nonspecific activity peaking well after the saccade. This category was more prevalent in F2vr than in FEF and would be considered movement-related using traditional criteria. Spatially selective movement-related neurons, category 5^PM_VS, were also found in roughly equal proportions in FEF and F2vr. In sum, by categorizing neurons in 1 region according to the criteria defined by another region, more subtle distinctions such as selectivity magnitude and fine-grained temporal differences are revealed in ways that are heretofore intangible.

One simple explanation for these differences in neuron proportion is the region of sampling; in FEF, fixation neurons are found deep in the sulcus (MacAvoy et al. 1991; Tanaka and Lisberger 2002; Izawa et al. 2009; Lixenberg and Joshua 2018) and these neurons as well as movement neurons are predominantly found in deep layers, whereas visual and visuomovement neurons are concentrated in upper layers (Segraves and Goldberg 1987; c.f. Sommer and Wurtz 2000). If a similar compartmentalization of neuron types exists in F2vr, we could have inadvertently sampled different compartments from the 2 regions.

Similarly, the proportion of neurons exhibiting target selection was greater in FEF than in F2vr. Thus, although both regions are involved in saccade target selection, FEF seems more specialized for this function. Interestingly, the proportion of neurons in each region with visual responses is almost exactly the proportion of neurons exhibiting target selection. Although not all visual neurons exhibit target selection (Schall et al. 1995a; Thompson et al. 1996), the majority of those that do are more closely related to visual processing than saccade planning (Thompson et al. 1996; Sato et al. 2001; c.f. Costello et al. 2013). Thus, the difference in proportion of neurons exhibiting target selection between the 2 regions may be related to the proportion of visually responsive neurons and explained by putative specificity of effectors. This supposition is further supported by the magnitude of target selection; among those neurons in each region that did exhibit significant target selection, the magnitude of this selection did not differ between regions. However, FEF has been shown to exhibit target selection when no overt response is required (Thompson et al. 1997) and even when an arm movement is required in lieu of an eye movement (Thompson et al. 2005b), suggesting that FEF may have a more general role in directing covert and overt attention. Further studies with multiple effectors are necessary to disambiguate the role of each region in effector specific and effector invariant target selection.

Because we did not have the monkeys perform any manual task, we could not identify neurons that select a response regardless of effector and thus we may be collapsing across dissimilar neuronal categories, and we cannot quantify the timing of intereffector responses. Furthermore, we may have found more similar proportions of response types. In FEF, responses to visual stimuli are attenuated if they are not potential saccade targets (Goldberg and Bushnell 1981), and a similar attenuation has been observed for stimuli that are not potential reach targets (Wise et al. 1992). Our laboratory certainly appreciates the impact of neuronal diversity (Lowe and Schall 2018), thus we realize that further direct comparisons would require tasks with manual responses.

Connectivity of Premotor Cortex and FEF

An important consideration for the differentiation between FEF and premotor cortex is the pathways by which the demonstrated visuomotor activity may arrive or be sent to different areas. In a seminal study examining the signals sent to the superior colliculus from FEF, Segraves and Goldberg (1987) demonstrated an example in which one site in the caudal bank of the arcuate was antidromically activated by stimulation in the superior colliculus. Thus, premotor cortex can supply visuomotor information to the superior colliculus (see also Fries 1984; Distler and Hoffmann 2015), which is one step closer to the brainstem saccade generator (e.g. Basso and May 2017). Furthermore, retrograde tracers injected into an extraocular muscle reveal oligosynaptic connectivity of the motoneurons with both dorsal and ventral subdivisions of the premotor cortex (Moschovakis et al. 2004). These findings demonstrate that FEF and premotor cortex are similarly situated to direct oculomotor activity. However, these findings should be taken with caution, as the particular regions of the colliculus that receive input from PMd have also been associated with reaching movements (Werner et al. 1997), indicating that these connections may underly oculomotor activity related more specifically to hand–eye coordination rather than for eye movements alone.

FEF and the sulcal region of premotor cortex also share similar afferents. In 1 study, anterograde tracers were injected into the lateral intraparietal (LIP) area which resulted in labeling of the caudal bank of the arcuate sulcus, though these labels were more sparse than in the rostral bank (Schall et al. 1995b). Interestingly, anterograde tracers injected into the tempo occipital area TEO did not show this labeling, but retrograde tracers in both areas showed labeling in both FEF and the sulcal region of premotor cortex. This suggests that both premotor cortex and FEF are reciprocally connected with the dorsal stream, whereas premotor cortex sends but does not receive projections from the ventral stream. FEF is reciprocally connected with the ventral stream. These general findings were also demonstrated for connections with V4 in the ventral stream (Ungerleider et al. 2008) and again for LIP in the dorsal stream (Petrides and Pandya 1984; Blatt et al. 1990). This suggests that premotor cortex may be primarily involved in spatial attention (e.g. Lebedev and Wise 2001) but not feature attention, whereas FEF is involved in both (Lowe and Schall 2019; for examples of feature attention, see Bichot et al. 1996; Xiao et al. 2006; Peng et al. 2008; Bichot et al. 2015). Again, these connections should be interpreted with caution, as the parietal connections of F2vr are strongest to caudal regions (Matelli et al. 1998; Gamberini et al. 2009, 2020) which are involved in reach tasks (e.g. Galletti and Fattori 2018) and eye–hand coordination (Diomedi et al. 2020). Additional tasks requiring feature information and/or arm movements may provide additional evidence differentiating these 2 eye fields.

Homology of Macaque and Human Eye Fields

A debate in assessing the homology of human and macaque FEF has been the common identification of FEF in human area 6 (Paus 1996; Petit et al. 1997; Tehovnik et al. 2000; Amiez and Petrides 2009; Neggers et al. 2012; Percheron et al. 2015; Schall et al. 2017) but in macaque area 8 (Huerta et al. 1987; Stanton et al. 1988). This difference is more apparent than real. The literature on the location of FEF in humans has referred rather exclusively to Brodmann’s cytoarchitectonic map (1909). His is not the only or last description, and it does not correspond to modern descriptions in many respects (Zilles & Amunts 2010; Nieuwenhuys 2013). In Brodmann’s map, area 6 occupies a large amount of the frontal lobe, but contemporaneous as well subsequent maps by other investigators subdivide Brodmann’s area 6 into many more areas. Myeloarchitectonic studies have distinguished the caudal end of the middle frontal gyrus from surrounding areas (e.g. Nieuwenhuys et al. 2015). Although other authors locate the caudal end of the middle frontal gyrus in area 6 (Sarkissov et al. 1955), the region has also been labeled area FB (von Economo and Koskinas 1925), area 4 s (von Bonin (1949), the boundary of FA and FB (Bailey and von Bonin 1951), and 8αβγ (Vogt and Vogt 1926). Penfield, with Förster, described the majority of stimulation sites eliciting eye movements in humans as being in 8αβγ (Penfield and Rasmussen 1950). The elaborate map of von Kleist (1934) identifies ocular adversive movements with Brodmann’s area 8.

The structure of the cortex occupied by FEF has been reevaluated more recently. Human FEF can be distinguished from surrounding areas by MRI myelin mapping (Glasser et al. 2016). In addition, a recent anatomical study reexamined the architecture of this region using modern chemo-architectonic methods in postmortem tissue from 6 subjects (Rosano et al. 2003; see also Schmitt et al. 2005). The histological structure of the superior precentral sulcus was distinct from adjacent rostral and caudal regions. A thin granular layer 4 was observed in sections labeled with neuronal nuclear protein, and the nonphosphorylated neurofilament triplet protein. Also, clusters of large, intensely immunoreactive pyramidal cells were located in deep layers 3 and 5. In sections labeled for calcium binding proteins, the 2 walls of the sulcus were characterized by higher density of calretinin-labeled interneurons, lower density of calbindin-labeled pyramidal neurons, higher density of calbindin-labeled interneurons in layers 2–3, and higher density of large parvalbumin-labeled interneurons in deep layer 3. These histological features resemble the macaque FEF more than agranular area 6. These immunohistochemistry methods highlighted distinctions across this cortical region that is obscured in Nissl-stained section. Based on this analysis of cytoarchitectural, myeloarchitectural, and histochemical studies, dorsal hFEF should be identified with area 8 and not area 6, whereas the ventral extension of hFEF can be identified with macaque area 45B (Gerbella et al. 2007; Petit and Pouget 2019; Borra and Luppino 2021). The present results offer further insights into the localization debate. Certainly, the similarity in single neuron activity in the 2 regions, both visual and motor, will make the 2 regions more confusable in gross activation measures of fMRI.

Functional Roles of FEF and F2vr

We found that some neural characteristics differed between the 2 regions. Specifically, we found that the activity maintaining spatial information during memory-guided saccades and the proportion of neurons that select targets in the visual search task are more pronounced in FEF than in F2vr. This is consistent with the role of FEF in other cognitively demanding tasks, e.g. antisaccade tasks, in both humans and monkeys, and with findings that PMd is not active in cognitively demanding visual tasks, e.g. visual search (Makino et al. 2004; Wardak et al. 2010). However, our finding that a considerable proportion of F2vr neurons, albeit a smaller proportion than in FEF, select targets during visual search is at odds with fMRI results indicating that PMd is not active in visual search.

This difference might be reconciled by the manner of responses. In our task, monkeys shifted gaze to the target stimulus, whereas the Makino and Wardak tasks used manual responses guided by but not directed toward the relevant stimuli. During a covert visual search with a manual level response, neural activity in FEF selected the target, but eye movement neurons were suppressed (Thompson et al. 2005). Thus, a dissociation between stimulus and response may result in less activation of premotor cortex, and the attentional components of PMd responses may be present but detectable by fMRI only with overt manual responses. Tasks requiring directed reaching may differentiate premotor and prefrontal oculomotor regions, but when tasks employ only eye movements, whether for theoretical or practical reasons, the 2 regions can be confusable. In sum, the unexpected similarities of neural modulation characteristics in FEF and F2vr indicate linkages in visuospatial attention and hand–eye coordination that merit further investigation.

Funding

This work was supported by National Eye Institute RO1-EY08890, P30-EY008126, National Institute of Child Health and Human Development U54-HD083211 and by Robin and Richard Patton through the E. Bronson Ingram Chair in Neuroscience.

Notes

We thank M. Feurtado, M. Maddox, M. Schall, and C. Suell for technical support. We thank T. Apple, J. Parker, K. Shuster, and L. Toy for expert animal care. We thank S. Motorny for computer and network systems support as well as P. Henry, B. Williams, and R. Williams for electronic and instrumentation support. Conflict of Interest: None declared.