Neuronal Activity in the Dorsal Premotor Area and Ventral Premotor Area of Monkeys adapting to a New Dynamic Environment Jun Xiao1, Camillo Padoa-Schioppa2 and Emilio Bizzi McGovern Institute for Brain Research, Department of Brain and Cognitive Sciences, Massachusetts Institute of Technology
Present address: J529, Biology Department City College of New York
2
Present address: Department of Neurobiology, Harvard Medical School.
3
Present address: McGovern Institute for Brain Research, Massachusetts Institute of Technology
Abbreviated title: New internal models of movement dynamics in PMd and PMv
Manuscript information: 33 text pages; 11 figures; 6 tables, 172 words in the abstract; 497 words in the introduction; 1317 words in the discussion.
Correspondence: Emilio Bizzi, M.D. MIT room E25-526, 77 Massachusetts Avenue, Cambridge, MA 02139. Phone: 617-2535769. Fax: 617-2585342. Email:
[email protected]
Acknowledgments: We thank Margo Cantor and Sylvester Szczepanowski for technical assistance. This research was supported by the National Institute of Health (NIH grant MN481185).
Key words: dorsal premotor area, ventral premotor area, force field adaptation, motor learning, neuronal plasticity, internal model, movement dynamics, monkey.
Abstract In a series of experiments, we investigated how neurons in the different motor areas of the frontal lobe reflect the movement dynamics, and how their neuronal activity undergoes plastic changes when monkeys learn a new dynamics. Here we describe the results obtained in the dorsal premotor area (PMd) and ventral premotor area (PMv). Monkeys performed visually instructed, delayed reaching movements before, during and after exposure and adaptation to a perturbing force field. The experimental design allowed dissociating the neuronal activity related to the movement dynamics from that related to the movement kinematics, and to investigate the neuronal plasticity associated with motor learning. During movement planning (i.e., during an instructed delay that followed the cue and preceded the go signal), dynamics-related activity was found in PMd but not in PMv. In contrast, neurons in both PMd and PMv reflected the movement dynamics during movement execution. In both PMd and PMv, neurons displayed plastic changes associated with the acquisition of a new dynamics, similar to that previously observed in the primary motor cortex (M1).
Introduction Execution of visually guided reaching movements involves a sequence of computational stages, including processing of movement kinematics (i.e., movement trajectory and its time derivatives) and processing of movement dynamics (i.e., the forces that ultimate cause the movement). From a computational perspective, processing of the inverse dynamics is non-trivial. Previous work suggested that the central nervous system might face this operation through internal models describing the dynamic properties of the limb and the environment (Kawato, 1999; Desmurget and Grafton, 2000; Wolpert and Ghahramani, 2000). Psychophysical studies in humans showed that subjects exposed to new dynamic environments can learn novel internal models for the dynamics (Shadmehr and Mussa-Ivaldi, 1994). It was also observed that new internal models are consolidated in the hours following exposure (Brashers-Krug et al., 1996). Moreover, internal models for the dynamics are acquired independently of other internal models, e.g. internal models for the kinematics (Flanagan et al., 1999; Krakauer et al., 1999). With respect to the physiological underpinnings, previous work mostly focused on the primary motor cortex (M1). Numerous experiments confirmed the original observation of Evarts (1968) that the movement-related activity of neurons in M1 changes reflects the presence of external loads (Fromm, 1983; Kalaska et al., 1989; Alexander and Crutcher, 1990a, b). More recently, we showed that the activity of neurons in M1 changed when monkeys acquired a new internal model for the dynamics. Neurons initially not “committed” to the task became directionally tuned following exposure to a new dynamic environment (Gandolfo et al., 2000). In addition, neurons that were initially directionally tuned changed their activity when monkeys learned a new dynamics, and often maintained their newly acquired activity patterns after re-adaptation to the non-perturbed environment (Li et al., 2001). Thus, evidence suggests that the neuronal population in M1 may contribute to both function of motor performance and motor learning (Li et al., 2001). In contrast, much less is known on how other areas of the frontal lobe contribute to establishing and processing internal models for the dynamics. According to the traditional serial view, several “premotor” areas harbor early sensorimotor processes and funnel their output into M1, which directly controls the execution of movements. In
agreement with this view, studies comparing the activity of different areas reported some evidence for hierarchical organization. For instance, unlike neurons in M1, neurons in PMd activate before expected visual signals (Mauritz and Wise, 1986) and during motor preparation (Wise and Mauritz, 1985; Kurata and Wise, 1988). In addition, target-related activity prior to and during movement is more frequent in PMd (Shen and Alexander, 1997a, b). Recent anatomical work, however, does not support a strictly serial view. Specifically, it was found that direct corticospinal projections originate from multiple “premotor” areas, including PMd, PMv, the supplementary motor area (SMA), the cingulate motor areas, and M1 (He et al., 1993, 1995). Concurrently, physiological work generally found extensive functional overlaps between areas (Riehle and Requin, 1989; Alexander and Crutcher, 1990a, b; Crutcher and Alexander, 1990; Crammond and Kalaska, 1994, 1996; Johnson et al., 1996; Scott and Kalaska, 1997; Scott et al., 1997; Shen and Alexander, 1997a, b; Johnson et al., 1999). These observations suggest that several motor areas might provide parallel contributions to the control of movement (Alexander and Crutcher, 1990a; Dum and Strick, 1991; Prut and Fetz, 1999). The present experiments tested the hypothesis that multiple areas might participate in the processing of the movement dynamics (Padoa-Schioppa et al., 2002) in two respects. First, we investigated whether neurons in PMd and PMv reflect the movement dynamics during movement planning and during movement execution. Second, we studied the changes in activity of neurons in these two areas when monkeys acquired a new internal model for the dynamics. Previous work on isometric activity supported our hypothesis for both PMd (Werner et al., 1991) and PMv (Hepp-Reymond et al., 1994; Hepp-Reymond et al., 1999).
Materials and Methods The experimental setup, recording techniques and data analysis were essentially the same previously described (Li et al., 2001), with minor differences. The NIH guidelines for the use and care of animals in laboratory were followed throughout the experiments.
Behavioral paradigm Two young female rhesus monkeys (R and N), weighting 5-5.5 Kg, participated in the experiment. The monkeys sat on a chair in an electrically isolated enclosure. With their right arm, they held the handle of a two degrees of freedom robotic arm (the manipulandum). The manipulandum allowed free movements limited to a horizontal plane. A computer monitor, placed vertically 75 cm in front of the monkeys, indicated the position of the handle (3x3-mm square, ca 0.2° of visual angle) and the targets of the movements (12x12-mm squares, ca 1° of visual angle). All reaching movements were from a central location to one of eight peripheral targets, equally spaced along a circle and 45° apart from each other. Actual reaching movements were 6 cm in length. The monkeys performed an instructed delayed reaching task. In each trial, the center square appeared on the screen. The monkeys acquired the center square (beginning of the trial), and waited for further instruction for 1 sec. One peripheral square then appeared on the screen (the cue), indicating the target of the reaching movement. Monkeys had to maintain the cursor still in the center target for a delay period of randomly variable duration (1.1-1.9 sec). At the end of the delay, the center square was extinguished (go signal), prompting the monkeys to move. Monkeys had to acquire the peripheral target within 1.8 sec, and to maintain the cursor within the peripheral square for an additional 1 sec to receive a juice reward (rew). After an inter-trial interval (iti) of 1 sec, the sequence started over again. During the movement, the monkeys had to maintain the trajectory within an angle of 60° on both sides of the straight line passing through the center and the peripheral square. The trial was immediately aborted if the monkey made an error, and another trial started, after the iti. Peripheral targets were pseudo-randomly chosen. Two motors attached at the base of the robotic arm allowed turning force fields on and off. In the experiment, we used one of two force fields, described by F=BV, where B is an anti-diagonal 2x2 rotation matrix B=[0, -b; b, 0] and V is the instantaneous velocity vector. Thus, the forces were in strength proportional to the velocity (viscous) and in direction orthogonal to the velocity (curl). Depending on the sign of “b”, this defines one of two force fields, clockwise (CK) or counterclockwise (CCK). For the intensity of the forces, we used b=0.07 N sec/cm. In each experimental session, the monkeys performed
in three subsequent behavioral conditions: Baseline (no force, ca 160 trials), Force (ca 160 trials), Washout (no force, ca 160 trials). In the Force condition, one of the two force fields (CK or CCK) was introduced. The same timing and spatial constraints were maintained throughout the session. During the training (4-6 months), the monkeys performed in non-perturbed conditions. The force fields were introduced only during the recordings. In total, monkey R performed in 28 and 27 sessions with the CK and CCK force fields, respectively. Monkey N performed 4 and 5 sessions with the CK and CCK force fields, respectively. For both monkeys, sessions with the two force fields were interspersed.
Surgery, microstimulation and gross anatomy Before the training, and under aseptic stereotaxic surgery, we implanted a headrestraining device on the skull of the monkeys. At the end of training, we performed a second aseptic stereotaxic surgery and implanted a recording chamber (18 mm, inner diameter) over the left hemisphere. For both monkeys, we centered the chamber on (A=16, L=-15). During the surgery, we could clearly observe the genu of the arcuate sulcus. After surgery, the monkeys were given antibiotics and pain medications. In the days prior to the first recording session and at the end of several recording sessions, we performed electrical microstimulation. We used a train of 20 biphasic charge balanced pulse pairs (pulse width=0.1msec, train duration=60 msec), delivered at 330 Hz and variable amplitude (20-120 µ Amp). Subsequent recording were concentrated in areas were arm movements could be elicited. Both monkeys were euthanized at the end of the experiment. They were given an overdose of pentobarbital sodium and then perfused transcardially with heparinized saline, followed by buffered Formalin. The recording locations were marked with electrodes dipped in black ink. The brain was then removed from the skull, and photographed.
Recordings The recording procedures were that previously described (Li et al., 2001). The trajectories (position and velocity) were recorded at the frequency of 100 Hz. For the
neuronal recordings, vinyl-coated tungsten electrodes (1-3 MΩ impedance) were manually advanced with a set-screw system (approximate depth resolution of 30 µm ). Electrical signals were acquired, passed through a head stage (AI 401, Axon Instruments) and an amplifier (Cyberamp 380, Axon Instruments), filtered (high and low cutoffs of 10 kHz and 300 Hz, respectively), and displayed on a computer monitor (sampling frequency of 20 kHz) using a commercial software (Experimenter’s WorkBench 5.3, DataWave Technology). Action potentials –detected by threshold crossing –were saved to disk (waveforms of 1.75 msec duration) for subsequent analysis. Eight electrodes were used in each recording session. No effort was made to locate the cortical layer of the recordings.
Data analysis: psychophysics For each movement, we defined the movement onset (mo) and the movement end (me) with a threshold-crossing criterion (4 cm/sec) on the speed. The psychophysics was analyzed using a correlation coefficient (CC), as previously described (Shadmehr and Mussa-Ivaldi, 1994; Li et al., 2001). For each trial, the CC is defined as the normalized co-variance between the actual speed profile (v) and an ideal speed profile (u), so that CC(v,u)=Cov(v,u v u)). In essence, the CC is a measure of similarity between the actual and the ideal speed profiles. The values of CC range between –1 and +1, and are close to +1 when the actual speed profile is close to ideal. In order to compare the neuronal activity across trials with similar kinematics, we disregarded the first four successful trials in each movement direction. Only the remaining trials were considered for further analysis. One of the aims of the experiment was to investigate the neuronal activity related to the movement dynamics during the instructed delay (see below). No time constraints on the reaction time (RT) were imposed during the experiment. In the analysis of the delay activity we excluded anticipated movements (RT500 msec). Only the remaining trials (>89%) were considered for subsequent analysis.
Data analysis: neurons The clustering was performed using either a custom-written software (described in (Li et al., 2001)), or with a semi-manual procedure using a commercially available software (Autocut 3, DataWave Technology). In both cases, waveforms were visually inspected for stability. Only cells with convincingly consistent waveform throughout the session were considered for further analysis. We considered the neuronal activity of single neurons in 4 separate time windows: the center hold time (CH, cue-500 msec to cue); the delay time (DT, go-500 msec to go); the movement time (MT, mo-200 msec to me); the target hold time (TH, rew-500 msec to rew). For each neuron and for each time window, the activity was separately analyzed in the Baseline, Force and Washout conditions. For each condition, we averaged the activity across trials, and obtained a tuning curve. We defined three parameters to characterize the tuning curves. The preferred direction (Pd) was defined as the direction of the vector average of the eight activity vectors. The average firing frequency (Avf) was defined as the scalar average of the neuronal activity across the eight directions. The tuning width (Tw) was defined as the angle over which the firing frequency was higher than half of the maximal neuronal activity among the eight directions (maximum of the tuning curve). The parameters were defined subject to the following pre-conditions. First, only tuning curves with an Avf>1Hz were considered. Second, the Pd and Tw were only defined for tuning curves displaying a significantly unimodal distribution across directions, as stated by the Rayleigh test (p0.035, Rayleigh test).
Field-specific cells, tune-in cells, and tune-out cells As the monkeys adapted to the force field, the activity of the cells modified. In particular, we observed a group of cells that were initially not tuned in the Baseline, became tuned in the Force condition, and lost their tuning again in the Washout. We also observed other cells that were originally tuned in the Baseline, lost their tuning in the Force condition, but regained their tuning in the Washout. These two groups of cells, which appeared dynamic in nature, were named “field-specific” cells. Figure 4a-b illustrates two examples of field-specific cells recorded in PMd (the activity refers to the MT). In total, we found that the group of field-specific cells accounted for 14% of cells in PMd, and for 10% of cells in PMv (MT activity). We found a group of cells whose changes outlasted the exposure to the perturbation. “Tune-in” cells were initially not tuned in the Baseline, and acquired a directional tuning in the Force condition following adaptation to the perturbing force. In the Washout, however, tune-in cells maintained their newly acquired directional tuning. Tune-in cells, which appeared memory in nature, were found in both PMd and PMv. One example of tune-in cell recorded in PMv with a CCK force field is shown in Figure 5a. In total, the group of tune-in cells accounted for 16% of PMd cells, and for 15% of PMv cells.
We also observed a group of “tune-out” cells. Tune-out cells were originally tuned in the Baseline, but lost their tuning in the Force condition, and remained nontuned in the Washout. One example of tune-out cell recorded in PMv with a CK force field is shown in Figure 5b. In total, the group of tune-out cells accounted for 14% of PMd cells, and for 25% of PMv cells.
Delay time: changes of preferred direction in PMd cells We classified cells into the classes kinematic, dynamic, memory I, memory II, and “other,” separately for each of the three time windows (DT, MT, TH), and for each of the three parameters (Preferred direction, Average firing frequency, and Tuning width). With respect to the Preferred direction (Pd), we classified only cells that showed a consistent directional tuning throughout the three behavioral conditions (Baseline, Force and Washout). We observed that during the DT time window the vast majority of PMd cells was either kinematic or dynamic (38% and 41%, respectively). Figure 6a shows an example of kinematic cell recorded in PMd with a CK force field. The Pd of the cells remains essentially unchanged across conditions. The cell shown in Figure 6b was recorded with a CK force field and was classified as dynamic according to its changes of Pd. In the Force condition, the Pd shifts in the CK direction (i.e., the direction of the external force). In the Washout the Pd shifts back to its original direction. In total, 34 PMd cells could be classified in the DT. Of these, 38% were kinematic, 41% were dynamic, 12% were memory I, and 9% were memory II. This PMd population is shown in a scatter plot in Figure 7 (top). The x-axis indicates the shift of Pd in the Force condition compared to Baseline, and the y-axis indicates the shift of Pd in the Washout compared to the Force condition. Positive values correspond to shifts in the direction of the external force. Each symbol represents one cell, and cells are color-coded according to their class. Kinematic cells (black), whose Pd does not change across conditions, lie close to the origin of the axes. Dynamic cells (blue), while Pd shifts in one direction in the Force condition (positive x-axis for shifts in direction of the external force), and in the opposite direction in the Washout, lie one the diagonal through the second and fourth quadrant. Memory I cells (green), whose Pd changes on the Force
condition but not in the Washout, lie on the x-axis. In contrast, memory II cells lie on the y-axis: their Pd does not change in the Force and changes in the Washout. We quantified the shift of Pd in the DT for the entire PMd population. For each cell exhibiting a directional tuning in both the Baseline and the Force condition we computed the shift of Pd. We then averaged the shifts across the entire population. On average, PMd population showed a shift of Pd of 11.0° in the direction of the external force in the Force condition compared to Baseline. This shift reached significance level in the DT (p
