Neural Dynamics in Monkey Parietal Reach Region Reflect Context-Specific Sensorimotor Transformations

Anti-reach task and control of behavioral parameters.

Two male rhesus monkeys (Macaca mulatta) were trained to perform a visually instructed, center-out anti-reach task in a frontoparallel plane. Two versions of the task were used, a memory task (Fig. 1 A) and a reaction time task. The memory task was used for neural recordings to analyze the PRR activity with respect to motor goal representations during the delay period (as opposed to motor representations during motor preparation/execution). The reaction time task was used to test for behavioral effects of the task rule (pro/anti) on the speed of sensorimotor transformations.

• Download figure
• Open in new tab
• Download powerpoint

Figure 1.

A , Memory-guided anti-reach task. The monkey had to keep ocular fixation on a small central red dot throughout the trial. The color of the central hand stimulus (green/blue) instructed randomly interleaved pro- or anti-reaches with the movement goal (dotted circles; not visible to the monkey) either at the location (left, right, up, or down) of the previously flashed cue (CUE) or opposite of it. Disappearance of the central hand stimulus after a variable memory period (MEM) instructed the movement (MOV). Visual feedback (FDB) appeared only after the monkey touched the correct goal on the screen (success) or after a wait time had expired short enough not to allow posture corrections (failure). B , Localization of recording sites (cross indicates estimated center) in the PRR in the medial intraparietal sulcus (IPS) of monkey TI. (Shaded areas in the left one-half of the axial view are imaging artifacts.) C , Typical delay period tuning in PRR. Outer panels, Response profiles (PSTHs; aligned with movement instruction at t = 0; smoothed with 50 ms Gaussian kernel) of a single example neuron strongly selective for the four different locations of the cue/reach goal during the standard memory-guided pro-reach. The inner panel shows the corresponding tuning curve with the resulting directional tuning vector (see Materials and Methods) for the late memory period (shaded area in PSTHs). Smoothed PSTHs or interpolated curves were only used for illustration; all quantitative analyses were performed based on direct spike counts.

In the memory task, the monkey initiated a trial by fixating a small central, red fixation spot and touching an adjacent green or blue hand target (fixation period). The green touch target instructed a pro-reach trial, and the blue target an anti-reach trial. After a fixed delay (1.0 s), a peripheral visual cue of the same color as the central hand target was briefly flashed (cue period, 0.2 s), indicating to the monkey where he later will have to reach. For a variable duration, the monkey had to keep the hand fixation (memory period, 1.0–1.5 s). When the central hand target turned off (go signal), the monkey had to reach (movement period, maximum, 1.0 s) either to the memorized location of the previously flashed cue (pro-reach; green cue) or to the diametrically opposed position (anti-reach; blue cue). The monkey received visual feedback about the correct movement goal (circular patch stimulus at the motor goal location) as soon as he acquired the desired position, or after the maximum movement period expired, in which case the trial was aborted. The monkey had to keep his hand at the reach target location (feedback period, 0.3 s) to successfully finish the trial and receive liquid reward.

In the reaction time task, the monkeys had to reach toward the movement goal immediately after the visual cue was presented. The memory period was omitted in this case, and the maximum allowed movement period (reaction plus movement time) was reduced to 0.7–0.8 s. In every other respect, the reaction time task was identical to the memory task described above. Reaction time was defined as the time between the movement instruction and the monkey's release of the touch screen. Movement time was defined as the time between the release and reacquisition of the touch screen at the target position.

Both versions of the anti-reach task consisted of eight different trial conditions (four target positions by two mapping rules, pro/anti). The four possible peripheral reach targets (right, 0°; up, 90°; left, 180°; down, 270° direction) were pseudorandomly interleaved from trial to trial. Eccentricity was always 9 cm (corresponding to 17–20° visual angles for screen distances of 25–30 cm for the two monkeys). Pro- and anti-reaches were also pseudorandomly interleaved; except for 12 of the 60 recording sessions total, when a block design was used (80 trials per block). The data from the block-wise recordings yield no differences to the randomized recordings and, therefore, will be presented jointly.

Eye position was registered with an infrared CCD camera (240 Hz; ISCAN, Burlington, MA). The monkeys had to keep ocular fixation throughout the course of the trial (tolerance, 3.5–4.0° visual angle). Hand position was registered with a 19 inch touch screen (IntelliTouch; ELO Systems, Menlo Park, CA). The hand stimulus had to be continuously touched within a tolerance window of typically 1.8 cm (3.5–4.0° visual angle at 25–30 cm screen distance) for the central, and 2.5–3.5 cm (4.8–8.0°) for the peripheral targets. Otherwise, the trial was immediately aborted without reward. Visual instruction stimuli were either presented on a liquid crystal display monitor (19 inch ViewSonic VX924; <5 ms off–on–off response time) mounted behind the touch-sensitive screen, or rear-projected from a digital light processing projector (U2–870; Plus Corporation, Tokyo, Japan) onto a tangent screen attached to back of the touch-sensitive screen. Visual display latencies were recorded with photo diodes for both systems and corrected for in the data analysis.

Animal preparation and recording procedure.

Both monkeys first were implanted with a head holder. After the monkeys had learned the task, recording chambers were implanted above the posterior parietal cortex, contralaterally to the handedness of each monkey, to allow access to the medial wall of the intraparietal sulcus. Structural magnetic resonance imaging was used to identify the position of the intraparietal sulcus for placement of the recording chambers and guidance of the recordings (Fig. 1 B). Surgical and imaging procedures were conducted under general anesthesia. All procedures were done in accordance with National Institutes of Health guidelines and were approved by the California Institute of Technology Animal Care and Use Committee.

Extracellular neural recordings were made from up to five microelectrodes simultaneously using a five-channel microdrive (“mini-matrix”; Thomas Recording, Giessen, Germany). The raw signal from each electrode was preamplified (20×; Thomas Recording), bandpassed, and amplified (154 Hz to 8.8 kHz; 400–800×; Plexon, Dallas, TX), before being subjected to on-line spike-sorting (Sort Client; Plexon). All spike waveforms were digitized (40 kHz) and recorded to disk for off-line control of sorting quality and stationarity (Offline Sorter; Plexon). In some sessions, a single-channel drive (Narishige, Tokyo, Japan) was used in combination with other electrodes (FHC, Bowdoinham, ME) and a 20× gain head stage (Plexon), but otherwise identical signal processing.

Single-cell spiking activity in the cortex beneath the recording chamber in each monkey was mapped for reach-related, delay-period activity to define a region of interest that conforms with the functional definition of PRR reported previously (Snyder et al., 1997). Recording sites were within 5–10 mm posterior and 4–7 mm lateral (Horsley-Clarke coordinates) at 3–6 mm subsurface, in the hemisphere contralateral to the arm used for reaching (monkey TI, left arm; TA, right). Data from this mapping procedure were not included in the current study. From independent subsequent recordings in this region, all units were included in the analysis that were sufficiently well isolated and stationary, regardless of their task-relatedness or specific tuning properties.

Neural tuning analysis.

Single-cell spiking activity from successful memory reach trials were analyzed for spatial selectivity on a cell-by-cell basis (tuning analysis), and on the population level (see below, Neural decoding analysis).

In the tuning analysis, the selectivity of single-cell spike rates for the location of the visual cue/motor goal was assessed at different times relative to stimulus or behavioral events occurring during the trial. Fixed, predefined periods (each 200 ms) were used for the tuning analysis: the last 200 ms before cue stimulus onset (baseline), the 200 ms of cue visibility (cue), the time between 100 and 300 ms after cue offset (early memory), the last 200 ms before the go signal (late memory), the last 200 ms before reach target acquisition (movement), and 200 ms after target acquisition during the visual feedback (feedback). Because the memory period and the arm movement itself had variable duration across trials, the neural signals were aligned either with respect to the cue onset, the end of memory period, or the time of the target acquisition to calculate averages. Spatial selectivity was quantified with a directional tuning vector (DTV), separately for pro- and anti-reaches. It is defined as vector average across all center-out cue directions u⃗ _i (unit vectors) weighted with the corresponding mean spike rates r_ij of neuron j as follows: The mean spike rate is calculated within each period and averaged across trials with identical conditions (∼15–20 repetitions per condition). The direction of the resulting tuning vector defines the preferred direction of a neuron (for an example, see Fig. 1 C). The DTV was calculated relative to the location of the cue (instead of the motor goal), which is an arbitrary choice that does not in any way affect the results. Significance of directional tuning was tested with a nonparametric one-way ANOVA (Kruskal–Wallis) with the four different visual cue directions as groups and sample sizes defined by the number of identical trial repetitions.

Four main tuning categories were defined based on the presence (or absence) of tuning in the pro- and the anti-reach condition. In each period, a neuron could either be tuned in both conditions (category I), in one of the two conditions [pro-only (II) or anti-only (III) tuning], or in none of the conditions (IV). For the pro and anti tuning category (I), three spatially defined subcategories were built based on a second factor, the relative preferred direction between pro- and anti-reaches. In case a neuron was tuned in the pro and anti condition, the difference of preferred directions between pro- and anti-reach tuning could be <45° (visual tuning), >135° (motor tuning), or in-between (spatially undefined tuning).

Neural decoding analysis.

In the decoding analysis, we used one-nearest-neighbor classification to quantify how informative the neural population response is about individual task parameters (Duda et al., 2001). We used classification procedures with different class definitions (class, set of trials with the task parameter in question kept constant) to predict different task parameters. With an eight-way classifier, trials were classified with respect to all possible task parameter combinations (“total decode,” four cue directions by two task rules). With a two-way classifier, trials were classified only with respect to the task rule, regardless of cue/motor goal direction (“context decode”). With a modified four-way classifier, trials were classified with respect to cue/motor goal direction. Trials with one task rule (pro-reaches) were used to train the classifier, whereas trials with the other task rule (anti-reaches) were used as test trials (“transconditional decode”).

Similarity of the neural population response patterns between test and training trials is used to predict to which class a test trial belongs. In the one-nearest-neighbor algorithm, class membership of a test trial is predicted by the class membership of the training trial whose neural response vector has minimal Euclidian distance (“nearest neighbor”) to the response vector of the test trial in the N-dimensional space. N corresponds to the total number of neurons (see below). The neural response vector r⃗_i (t) = (r _1j(t),…,r_Nj (t)) is time dependent, where r_ij (t) represents the spike rate of neuron i in the time interval t of trial j. The classification procedure was performed separately for each time interval t. Time intervals were 100 ms long with successive intervals overlapping by 50 ms.

Decoding performance is the percentage of correctly predicted test trials. The classifier was trained with six of seven of all relevant trials (training set) and predictions were made for the remaining one of seven of the trials (test set) for all seven permutations of the sets (sevenfold cross-validation; see below for the total number of trials). Initial assignment of individual trials to the seven subsets was done at random. Decoding performances were averaged across the seven individual cross-validation results. Deviation of decoding performances from chance level, or differences in performances between different classification conditions (e.g., decoding the motor goal in pro- vs anti-trials only), were tested with a t test. The significance criterion was Bonferroni-corrected for the number of nonoverlapping time windows (N_t = 15; 30 time steps with 50% window overlap) to account for multiple testing induced by the time-dependent analysis.

The population of neurons was collected over multiple experimental sessions with a variable number of trials recorded for each neuron (typically ∼20 trials per condition). For the classification analysis, the same number of trials (=15) for each of the eight possible classes (four direction by two task rules) was selected as a random subsample of all trials recorded for this neuron. Only neurons were included in the decoding analysis for which a minimum of 15 trials per trial parameter combination was recorded (N = 131 of 143 recorded neurons). We did not preselect the neurons in any other way, which means the population decode includes all well isolated neurons, even if they are not significantly spatially tuned or task related based on the single-cell analysis.

Spatial selectivity of the neural activity only emerged after a spatial cue has been presented in each trial. To test for latency differences in the onset of spatial selectivity in pro- versus anti-reaches on the population level, we compared the dynamics of the decoding performances P_P and P_A between the two task conditions (pro/anti). For this, we calculated a cross-correlation function between the two performance curves (see Fig. 5 A) as a function of time. The relative time shift Δt between the two temporal performance profiles was calculated as the time lag of the peak in the unbiased cross-correlation function: Arguments n and m denote indices for the time bins. To increase the temporal resolution for the calculation of the time shift, the performance profiles P_P and P_A were interpolated by a factor of 25 before cross-correlating (“ideal” low-passing via fast Fourier-transform; MATLAB function interp), leading to a nominal 2 ms resolution. The time lag measure is sensitive to small differences in the exact shape of the two performance curves. This shape can vary slightly depending on the random trial selection in the cross-validation procedure. To estimate the reliability of the time lag measure, we averaged across 100 repetitions of calculating the performance curves with different random trial selections.

Other classification algorithms, such as maximum likelihood or Fisher's linear discriminant analysis, were tested also, yielding qualitatively the same results. Because the general conclusions are not affected by the choice of the classification algorithm, we decided on the nearest-neighbor algorithm because of its simplicity, robustness, and lack of assumptions on the shape of spike rate distribution or directional tuning functions.