Transition from Target to Gaze Coding in Primate Frontal Eye Field during Memory Delay and Memory–Motor Transformation

Approximate location of the FEF and the recorded sites in the two monkeys. A shows the anatomical location of the FEF, located at the anterior bank of the arcuate sulcus. B, Sites within the FEF from which neurons were recorded in each animal are plotted (circles) in the coordinates of the recording chamber with the center (0,0) approximately located at the stereotaxic coordinates corresponding to the FEF (see Materials and Methods). The black semicircle represents the edge of the recording chamber. The color code represents the neuron type recorded from each site. Low-threshold microstimulation at these sites evoked saccades ranging from 2° (at the most lateral sites) and 25° (at the most medial sites) in head-restrained conditions (Bruce and Goldberg, 1985).

An overview of the analysis methods for identifying the spatial code and sampling neuronal activity from a time-normalized activity profile. A shows an example analysis for identifying the spatial code. A-1, Here, activity from the early visual response (80-180 ms after target onset) was sampled for analysis. A-2 shows the T–G continuum, and three example RF plots are shown for the visual response corresponding to the demarked models (arrows) along the T–G continuum. T is the eye-centered target model and G is the eye-centered gaze model. In the RF plots, each circle represents firing rate data (diameter) for a single trial plotted over position data corresponding to the tested model (in this study, models spanning the target model, T, and the gaze model, G). The PRESS residuals are shown at the bottom of each RF plot. In each RF plot, the color code (blue–red scale corresponding to low-to-high) represents the nonparametric fit made to all data points. A-3 shows the mean PRESS (y-axis) as a function of tested spatial model along the T–G continuum (x-axis). For this example visual response the best-fit model or spatial code (lowest PRESS residuals) is the intermediate model one step away from T (toward G). Although A shows analysis only for a single sampling window, for the main analyses reported in this study we sampled activity at 16 half-overlapping time windows with the first starting at visual response onset and the last starting at gaze onset. For this, we normalized the time between visual response onset until movement onset so that we could collapse all trials together for analysis. B shows the raster and spike density plots corresponding to the classic visually aligned (B-1) and movement-aligned (B-2) neuronal responses, as well as the time-normalized spike density (B-3), and illustrates activity sampling based on each of these schemes.

A representative neuron with visual, delay, and movement responses, and results for the overall population. A shows the visually aligned (left) and movement-aligned (right) raster and spike density plots for a VM neuron with sustained delay activity. The visual response of this neuron is from 65 to 300 ms after target onset, and the movement response begins 30 ms before gaze onset. B shows the time-normalized activity profile corresponding to A with the period between visual response (VR) onset and gaze movement onset normalized for all trials. C shows the RF maps for four time steps (C1-C4) sampled from the time-normalized activity profile (B, pink shades) with the blue-to-red color gradient representing low-to-high neuronal activity levels. The best-fit model (i.e., spatial code) at each of these time steps is depicted by a red triangle placed on the T–G continuum (panels above the RF plots). For this neuron, there was a progressive but partial shift (3 of 10 steps) in spatial code toward G. D depicts the time-normalized spike density for the entire population (n = 74), including neurons with either visual or movement response, or both. Neurons with movement-related activity beginning at or after gaze onset are eliminated. E shows the mean (±SEM) of spatially tuned best-fits at 16 half-overlapping time steps from an early visual period (visual response onset for visually responsive neurons, and 80 ms after target onset if the neuron was not visually responsive) until gaze movement-onset time. The solid line shows the mean of the fits made to individual neuron data highlighting the change in the population spatial code along T–G continuum as activity progresses from vision to movement. The histogram in the bottom panel shows the percentage of neurons that exhibited spatial tuning (y-axis) at a given time step (x-axis).

Single-neuron example and population results for V neurons. A shows the time-normalized spike density profile for an example V neuron (top) and the data points corresponding to the spatially tuned time steps across 16 half-overlapping time steps (bottom). The RF plot corresponding to the highlighted time step (bottom panel, light red circle with green boarders; first time-step here) is shown with the spatial code highlighted above the plot. B shows the population time-normalized post-stimulus time histogram (mean ± SEM) and the mean (±SEM) of the spatially tuned data points at these time steps across the V population. Colored data points (bottom) correspond to time steps at which the population spatial coherence was significantly higher than the pretarget baseline and gray shades correspond to eliminated time steps, with spatial coherence indistinguishable from pretarget baseline. The histogram shows the percentage of neurons at each time step that exhibited spatial tuning. The baseline firing rate is calculated based on the average firing rate in the 100 ms pretarget period is shown by the solid horizontal lines in spike density plots (A, B, top). For reference, the approximate visual, delay, and motor epochs are depicted at the top of the panels.

Single-neuron example and population results for VM neurons. A, B, Same conventions as in Figure 5. C, The RF plots corresponding to time steps with highlighted data points (circles with a green border) in A (bottom) are shown, with the spatial code along the T–G continuum highlighted above each plot.

Distribution of best-fit models across the T–G continuum for the VM population through five time steps through visual, delay, and movement responses. A shows the distribution of best-fits for VM neurons for early-visual (1st time step from the time-normalized activity profile), early-delay (4th time step), mid-delay (9th time step), late-delay (13th time step), and perimovement (15th time step) intervals. Only neurons with significant spatial tuning are considered. The number of neurons contributing to each distribution is indicated on each panel (the number in brackets also includes best-fits outside of the presented range). B plots the spatial code (i.e, value of the fit to the T–G data) at each of the delay intervals (y-axis), versus the spatial code at the perimovement period (red dots). Here, only the 21 neurons that contributed to all five panels in A were plotted. Note the trend (from the early- to mid- to late-delay periods) for the data points to migrate toward the line of unity (i.e., toward their movement fits).

Spatiotemporal progression of neuronal code in VM neurons with sustained delay activity. A shows the results with time-normalized activity sampling, including visual and movement response using the same conventions as in Figure 5B (bottom). B shows the results for only the delay period, with visual and movement responses excluded. Specifically, activity was sampled from 12 half-overlapping steps from the end of the visual response (on average, 266 ms after target onset) until the beginning of the movement response (on average, 85 ms before gaze onset). This duration was on average 635 ms. C shows the spatial code at fixed times intervals relative to the following specific task events: target onset (left); the Go-signal (middle); and gaze onset (right). For target-aligned analysis (C, left), the time from 80 ms after target onset and the earliest Go-signal was divided into eight half-overlapping steps, resulting in a sampling window size fixed for any session but ranging between 80 and 150 ms, depending on whether the earliest Go-signal appeared at 450 or 700 ms relative to target onset for that session. The Go-signal-aligned analysis (C, middle) was performed using 100 ms half-overlapping windows starting at 150 ms before and extending to 150 ms after the Go-signal. The movement-aligned analysis (C, right) was performed using half-overlapping 100 ms sampling windows starting from 150 ms before and extending to 150 ms after gaze onset. Notice that, although there is no change in spatial code triggered by specific task events, there is a progressive change in spatial code from T toward G as we move away from the time of target presentation (left) to the time of gaze onset (right), which is in agreement with the trend seen in A and B.

Single-neuron example and population results for DM neurons. A and B follow the same conventions as in Figure 5. C follows the same convention as in Figure 6C. Since these neurons lacked a visual response neuronal activity, sampling started from 80 ms after target onset.

Single-neuron example and population results for M neurons. The same conventions as those in Figure 5 are used. Since these neurons lacked a visual response neuronal activity, sampling started from 80 ms after target onset.