| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
|
|
|
|||||||
|
|||||||||
1Departments of Psychology and 2Neurology, and 3Arizona Research Laboratories and Neural Systems Memory and Aging, University of Arizona, Tucson, Arizona 85721
Submitted 20 January 2004; accepted in final form 6 April 2004
 |
ABSTRACT |
|---|
 TOP ABSTRACT INTRODUCTIONMETHODSRESULTSDISCUSSIONGRANTSACKNOWLEDGMENTSREFERENCES |
|---|
|
INTRODUCTION |
|---|
|
TOPABSTRACTINTRODUCTIONMETHODSRESULTSDISCUSSIONGRANTSACKNOWLEDGMENTSREFERENCES |
|---|
Four studies describing long-term response properties of neurons observed with chronic recordings provide evidence of functionally stable responses in various cortical regions. Schmidt et al. (1976)
implanted electrodes in monkey motor cortex for 
223 days and reported that one neuron, recorded for 29 days, had a stable relation to specific movements. Thompson and Best (1990) reported that the place fields of 10 rat hippocampal neurons were stable for between 6 and 153 days. Nicolelis et al. (1997) presented poststimulus time histogram data showing that 6 rat somatosensory neurons responded similarly in 2 recording sessions separated by 38 days. Finally, Williams et al. (1999) reported that the frequency response properties of one neuron in the guinea pig auditory cortex were stable for over 1 mo.
We decided to examine the issue of functional stability in the macaque dorsolateral prefrontal cortex (dlPFC) for two reasons. First, responses of dlPFC neurons have not been tested for periods longer than a single recording session, leaving open the question of functional stability. Second, we were interested in identifying neural correlates of learning based on recordings performed while monkeys learned correct hand movements to novel images. Both approaches were complementary. Recordings in the highly familiar, standard task provided a "learning-free" comparison condition with which to interpret neuronal responses in the subsequent learning task. In addition, neuronal response stability can be judged in 2 different task paradigms, allowing us to determine whether stably responsive neurons can also respond flexibly when task conditions require learning. Waveform stability measurements were used to identify recording sessions in which highly similar waveforms were present for days, allowing us to discount the likelihood that changes in neuronal response properties could be attributed to signal loss. Comparisons between well-learned and novel conditions are pertinent because neurophysiological recordings show that the responses of single neurons are remarkably consistent within a recording session (Funahashi et al. 1989; Fuster et al. 1973), yet neuropsychological and neurophysiological studies suggest the PFC is important for providing flexibility when dealing with changing task demands (Asaad et al. 1998; Dias et al. 1997; Murray and Wise 2000; Raichle et al. 1994; Stuss and Benson 1986; Wang et al. 2000; Watanabe 1990).
|
METHODS |
|---|
|
TOPABSTRACTINTRODUCTIONMETHODSRESULTSDISCUSSIONGRANTSACKNOWLEDGMENTSREFERENCES |
|---|
All procedures were carried out in accordance with the National Institute of Health Guide for the Care and Use of Laboratory Animals (National Institutes of Health Publications No. 80–23), and the University of Arizona's IACUC. Two rhesus macaques (
20 lb. each) were trained on 2 Go/NoGo visual-discrimination tasks while seated in primate chairs, watching a video monitor (NEC multisync XP17-in. diagonal). Control of stimulus presentation, detection of behavioral responses, and delivery of rewards was achieved using Cortex software developed at the NIMH Laboratory of Neuropsychology. Both tasks required the same behavioral responses, but differed in terms of the number and familiarity of the stimuli. Monkeys learned to use visual images as cues to perform Go responses (bar releases within 500 ms of image presentation) or NoGo responses (bar releases later than 500 ms). Monkeys initiated trials by grasping a response bar, waiting 500 ms for a visual cue while maintaining that grasp, and releasing it at the appropriate time depending on the cue. Stimuli were displayed for as long as the bar was grasped for a maximum of 500 ms. Figure 1 shows a schematic diagram of grasp durations, stimulus durations, and reward delivery within the tasks. Rewards were delivered only after correct NoGo trials; incorrect NoGo trials simply led to the next randomly chosen image. Correct Go trials led to presentation of the next image, whereas incorrect Go trials were followed by a brief time-out waiting period (1,500 ms) and repetition of the trial. Changes in eye position were monitored with an infrared tracker (Microguide, Downers Grove, IL). Horizontal and vertical voltage signals were digitized (ComputerBoards CIO-DAS16/16) and interfaced with Cortex software to identify saccade-related activity. Fixation was not obligatory, but it increased the efficiency of task performance; well-trained monkeys invariably started the trial with fixation at the center of the monitor and made large saccades to targets located in the periphery. This behavior occurs quite naturally, requires little training, and reward is not necessary (Wilson and Goldman-Rakic 1994).
|
), all presented centrally, to test for stimulus selectivity (ÓScalaidhe et al. 1999; Wilson et al. 1993). Ten additional stimuli, used to elicit saccadic eye movements, were horizontal (Go) and vertical (NoGo) black bars presented at the point of fixation and at 4 locations in the periphery (12°) along the horizontal and vertical meridians, one location per trial. The bars were randomly interleaved with presentation of the complex pictures, shapes, and faces. All images subtended ±4° of visual angle from the point of fixation on the monitor. Following this standard task, monkeys performed a learning task (200 trials) that used 2–4 novel cue images requiring the monkey to learn which of the 2 possible behavioral responses (Go or NoGo) was associated with each picture. Novel images were obtained from the Photodisc 6.2 library and consisted of complex scenery images similar to those cueing the Go response in the standard task. Recordings
Recordings were made with independently movable arrays of nichrome electrodes (Wilson et al. 2003) chronically implanted in the left and right dlPFC of monkey 1 and the left dlPFC of monkey 2. Fine-wire electrodes (Gray et al. 1996) were twisted and fused into stereotrode configurations for strength, but recordings were single-ended. Extracellular signals were fed to op-amp "headstages" and a preamplifier, and then to power amplifiers (providing a typical amplification of 10,000) and A/D converters with a 40-kHz sampling rate (Plexon, Dallas, TX). Data-acquisition systems allowed on-line template matching or threshold crossing discrimination of 4 units per channel. Recording files included digitized waveforms, timestamps of action potential occurrence, and timestamps of task events. The digitizing window length for each waveform was user defined within each recording session, ranging from 0.8 to 1.2 ms (32–48 data points per waveform).
Data analysis
Waveform stability.
Interpretation of all results depended on the demonstration of multiple-day action potential waveform stability. To begin, all waveforms from single recording sessions were first sorted into single neuron groups using on-line template matching (Williams et al. 1999) and then were sorted again off-line. Off-line sorting was based primarily on principal-component clustering. When necessary, these groupings were refined by removing individual waveforms (Off-line Sorter version 1; Plexon). Clusters were considered to indicate waveforms originating from a single neuron if 0.5% or fewer of the spikes occurred within 2 ms of any other spike.
Having clustered waveforms within a single day's recording session, these waveforms and clusters were compared across days to determine whether the recordings sampled the same neurons. Comparisons were made by merging consecutive recording files into a single file (PlexUtil; Plexon), recalculating the principal components of the waveforms in the merged file in the Off-line Sorter, and examining the waveform groups for cluster overlap. Single-neuron waveforms were judged stable if the clusters from similar waveforms on both days largely overlapped yet were distinct from any other unit's clusters or noise.
In Fig. 2, waveform clusters from different neurons (1, 2, and 3) are all spatially distinct, whereas clusters from the same neurons on days 1 and 2 show nearly complete overlap. Closely clustered principal-component points represent waveforms with similar shapes and amplitudes. Spatially distinct clusters represent waveform groups that are distinguishably different. The overlap shown in Fig. 2C indicates that waveforms from each neuron recorded on the first day were largely indistinguishable from those recorded on the second day. Multiunit activity was excluded from analysis before waveform similarity assessment, although we did observe stable multiunit clusters. Beyond these measures, our recording stability criteria excluded waveform groups that, when compared, 1) did not overlap, 2) overlapped with different units, or 3) differed by more than 0.2% in terms of the percentage of short-latency (2 ms) interspike intervals on each day.
|
). Figure 3 shows the distribution of between-day differences in peak/trough timing and in the average peak-to-trough amplitude for waveforms judged to be stable. Figure 3A shows that 84/94 (89%) neurons had differences of only 0 or 25 µs in day-to-day, peak-to-trough measurements. These values were calculated by subtracting the time to trough from the time to peak for each neuron on each day and then by taking the absolute value of the difference between day 1 and day 2 measurements. Figure 3B shows that variations in waveforms' total average amplitude between days were greater than timing differences, although these differences were within 15% for 75/94 (80%) neurons. Percentage change was calculated by dividing the absolute value of the day 1 – day 2 difference by the average amplitude of the larger waveform. Therefore all comparisons express change as a percentage decrease from the larger waveform, irrespective of which day's average waveform was actually larger.
|
Functional classification of neuronal responses
After identifying recordings that were stable for 2 or more consecutive daily testing sessions, neuronal responses were classified and quantified. Neuronal responses were analyzed in 3 general steps; 1) response classification based on statistically significant differences between task epochs and/or trial types, 2) quantification of response and/or trial selectivity magnitude, and 3) quantification of neuronal response timing. These general methods were applied to each neuron each recording day, then compared across days to judge each neuron's functional stability. The experimental design allowed us to classify neurons as responsive or unresponsive and selective or nonselective on a daily basis, then to see whether these qualities were maintained across days.
We first determined whether a neuron was responsive by comparing its prestimulus baseline firing rate with its poststimulus firing rate using paired t-test, P < 0.01 (Kralik et al. 2001). Baseline firing rate was estimated by generating prestimulus bin counts for each trial in a 200-ms bin immediately before stimulus onset. Poststimulus firing rates were typically calculated using bin counts for each trial in a time window 100–300 ms after stimulus onset. Neurons were classified as "responsive" if the difference was statistically significant and as "unresponsive" if it was not. Classifications for each neuron were then compared across days. Neurons with significant responses each day were classified as "consistently responsive," those not responding in any task epoch on any day were classified "consistently unresponsive," and those responding one day but not the next were designated "inconsistently responsive" (see Table 1 for results).
|
Calculating response strength and response strength stability
Within a single recording session, each neuron's response strength was quantified with an index (A – B)/(A + B), in which A was the average poststimulus response and B was the baseline firing rate. Next, we calculated the stability of these strength measures by dividing the smaller of 2 days' response strength index values by the larger. Note that the expression does not indicate which day's response was greater. To further analyze these responses and to confirm the within-day results of paired t-test, we used a 2-way ANOVA in which the first factor was the response bin (prestimulus vs. poststimulus) and the second was the recording day. This allowed us to identify significant changes in response strength (an interaction effect, P < 0.005) and significant shifts in overall neuronal excitability (a main effect of recording day P < 0.005). Shifts in overall excitability were observed when firing rates in both analysis bins increased by the similar amounts without affecting the absolute difference between them. Main effects were interpreted only when the interaction effect was not significant.
Identification of neurons with selective responses
Responsive neurons were further analyzed for selectivity by comparing firing rates on different trial types within a single poststimulus bin (typically 100–300 ms after stimulus onset). We used a nested GLM (General Linear Model, SPSS for Windows version 10, P < 0.01) analysis to determine whether neurons responded differently on Go versus NoGo trials, then to determine whether neurons responded differently to specific images. Neurons were defined as Go selective when there was significant differences between Go and NoGo trials and when the firing rates for Go trials were higher; the converse defined NoGo neurons. For neurons with significant effects of stimulus condition, one-way ANOVA (P < 0.01) and Tukey post hoc testing (P < 0.05) was used to identify the specific images associated with significantly different neuronal firing rates (SPSS). Neurons with firing rates varying in relation to image identity were classified as stimulus selective, whereas those differentially responsive during eye movements to the monitor periphery were classified as saccade selective. All neurons with selectivity for stimuli, rewards, or eye movements were designated "Specially" selective or "Special."
Calculation of selectivity strength and selectivity strength stability
Selectivity strength was quantified with an index (A – B)/(A + B), in which A was the neuron's average firing rate during trials evoking the maximal response and B was the neuron's firing rate during trials of the opposite kind. For Go-selective neurons, A was the average poststimulus firing rate on Go trials within a user-specified bin (typically 100–300 ms after stimulus onset), whereas B was the average firing rate on NoGo trials during the same poststimulus period. The converse was true for NoGo-selective neurons. For stimulus-selective neurons, A was the average firing rate after the preferred stimulus, whereas B was the average firing rate after the least-evocative stimulus. For saccade-selective neurons, A was the average firing rate after eye movements in the neuron's preferred direction, whereas B was the average firing rate during eye movements in the opposite direction. Selectivity stability was calculated by dividing the smaller of 2 days' selectivity strength index values by the larger.
As with the across-day statistical testing performed on responsive neurons, we used 2-way ANOVA to determine whether neurons with selective responses remained selective to the same degree across days and to determine whether there were significant shifts in a neuron's overall excitability (factor A = trial type, factor B = recording day; P < 0.005). The trial types used for this analysis were specific to the type of neuron as described above. For example, 2-way analysis of a Go-selective neuron would use the neuron's responses on Go trials from days 1 and 2 as factor A and day number as factor B.
Quantification of neuronal response timing stability
To investigate temporal characteristics of neural responses, we determined neuronal response onset latencies with cumulative sum histograms (Wilson and Rolls 1990; Woodward and Goldsmith 1964). The criterion of a significant change in firing rate was the crossing of the cumulative sum histogram above or below the 99% confidence interval (NeuroExplorer 2.6; Plexon). Neuronal response onset was calculated from the average firing rate across all trials. Comparing these latencies across days for each responsive neuron provided a measure of neuronal response onset timing stability. In addition, we quantified neuronal response timing profiles with perievent time histograms (PETHs). This analysis generated 150 data points for each neuron describing the average firing rate in 10-ms bins from 500 ms before stimulus onset to 1,000 ms after stimulus onset. For this analysis, selective neuronal responses were separated by trial type (Go or NoGo), generating one PETH for each condition. Correlation analysis was used to compare these "timing profiles" across days. For selective neurons, we reported correlation coefficients based on data from the preferred trial type. Resultant Pearson correlation coefficients indicated how well neuronal response peaks, troughs, and other periodic elements occurred at the same moments with respect to the same task events across days. Unresponsive neurons were excluded from these analyses because their firing rates never significantly deviated from baseline rates.
Identification of neurons with learning-related responses
To identify neurons with learning-related activity, linear regression (Microsoft Excel 2002) was used to compare trial number to firing rate in a specified poststimulus bin. Learning effects need not be linear or rate based but analyses focusing on nonlinear effects, transient changes (initial changes followed by a return to initial response levels), and attempts to identify changes in neuronal response onset (time to first spike analysis; spike train onset analysis) were all unfruitful in this data set and will not be described. Neurons with significant time-dependent changes in firing rate were defined as learning-related. Behavioral evidence of learning was provided by regression analysis of trial number and monkeys' hand movement response latency, and by change in the number of errors during the learning task.
|
RESULTS |
|---|
|
TOPABSTRACTINTRODUCTIONMETHODSRESULTSDISCUSSIONGRANTSACKNOWLEDGMENTSREFERENCES |
|---|
Ninety-four neurons [84 (monkey 1) and 10 (monkey 2)] were recorded for 2–4 h daily for a minimum of 2 consecutive testing sessions and a maximum of 9. Chronically implanted electrodes were distributed in a 12 x 12-mm matrix centered 17 mm lateral to the midline and 26 mm anterior to ear-bar zero in the left and right hemispheres of monkey 1 and the left hemisphere of monkey 2. Initial recordings were performed in superficial cortical regions, whereas subsequent recordings sampled neurons at depths of 12 mm from the surface. Twenty-three neurons were recorded for 3 days, 5 were recorded for 4 days, 3 were recorded for 7 days, and 1 was recorded for 9 days. Two of the neurons recorded for 7 testing sessions were actually stable for 9 days because of the intervening weekend.
Neuronal responsiveness classification
For each recording session, neurons were classified as responsive if poststimulus firing rates were significantly different from prestimulus firing rates (paired t-test, P < 0.01, SPSS). Comparing these results across days for each neuron, 66/94 (70%) neurons were responsive in each recording sessions (consistently responsive), 22/94 (23%) showed no significant response in any session (consistently unresponsive), and 6/94 (6%) were responsive some days but not others (inconsistently responsive). For 23 neurons recorded over 3 or more days, 20/23 were consistently responsive across all recording days, and 3/23 were consistently unresponsive across all recording days.
Neuronal selectivity classifications
Of 66 consistently responsive neurons, 55 were selective for specific task parameters on at least one day, whereas 11 were nonselective in that their responses did not significantly vary (GLM, ANOVA, P < 0.01) with respect to either trial type or image number. Neurons were classified as consistently selective if they were selective for the same factors in all recording sessions. Forty-six out of 55 (84%) neurons were consistently selective, whereas 9/55 (16%) were inconsistently selective. Within the group of consistently selective neurons we identified 4 different types of selective responses; 25/46 (54%) were selective for Go trials, 14/46 (30%) for NoGo trials, 2/46 (4%) for individual stimuli, 4/46 (9%) for saccades, and one responded after reward delivery. Table 1 shows the number of neurons with each type of responsiveness and selectivity identified in these experiments.
Figure 4 shows maintained stimulus selectivity, whereas Fig. 5 shows maintained Go, NoGo, and saccade selectivity.
|
|
Figure 5 shows the responses of 3 different neurons, each representative of a different type of stable selectivity. Figure 5 (left) shows a Go-selective neuron that responded more vigorously on Go than NoGo trials both recording days. Figure 5 (middle) displays a NoGo-selective neuron that responded more strongly on NoGo trials both recording days. The saccade-selective neuron (Fig. 5, right), responded most strongly on eye movements to the left of monitor center (12° of visual angle) all 3 recording days. Figures 4 and 5 exemplify stable, selective responses. The variety of stable response types demonstrates that functional stability is not restricted to responses of a specific kind. Last, 15/46 consistently selective neurons were recorded over 3 or more days and all retained their specific selective response properties for all recording days.
Based on the classifications outlined in Table 1, we decided that neurons with stable responsiveness and/or stable selectivity were functionally stable. In contrast, neurons with inconsistent responsiveness or selectivity were considered functionally unstable. Unresponsive neurons did not provide information about function and were considered separately. By these criteria, there were 72 responsive neurons; 57/72 (79%) were considered functionally stable, whereas 15/72 (21%) were considered functionally unstable.
Strength of neuronal response selectivity
In addition to selectivity classifications, we calculated an index of selectivity strength by comparing average firing rates on different trial types. For Go-selective neurons, selectivity strength was calculated with the index (A – B)/(A + B), where A = average firing rate on Go trials and B = average firing rate on NoGo trials. For NoGo neurons, the converse was true. This yields a distribution of selectivity strength values for neurons of each selectivity type. Trial types used to calculate the selectivity strength of "Specially selective" neurons are described in METHODS, but use the same index formula. All day 1 and day 2 selectivity strength index values for all selective neurons are displayed in the scatterplot shown in Fig. 6. For recordings lasting 3 or more days, the day (n) versus day (n + 1) comparisons were considered as additional day 1 – day 2 data points and plotted on the same graph.
|
Having established daily selectivity strength values for each neuron, we calculated the stability of those measures across days. Each between-day comparison produced a stability score for each neuron ranging from 0 to 1, where 1 indicates perfect selectivity strength stability.
Figure 6 shows the distribution of selectivity strength stability scores produced when the smaller of 2 days' index values is divided by the larger, and the correlation between day 1 and day 2 selectivity strength values for all selective neurons. For these neurons, conservation of selectivity strength across days is indicated by the number of neurons with high selectivity strength stability index values, and the high positive correlation (r = 0.86) between day 1 and day 2 selectivity strength index values. Considering Go, NoGo, and "Special" neurons separately, correlations between day 1 and day 2 selectivity strength index values were r = 0.73, 0.91, and 0.97, respectively.
These graphs also provide a representation of the variability across days, showing that some neurons were more stable than others. To determine how much of this variability was statistically significant, we performed a 2-way ANOVA (P < 0.005) comparing selectivity strength across days. Two effects were important: a main effect of day (indicating a consistent shift in firing rates during trials of different types) and an interaction effect (indicating that selectivity strength significantly increased or decreased between days). The main effect of trial type was a redundant calculation, considering that the previously mentioned nested GLM analysis identified neurons with trial type selectivity. Table 2 lists the results of this analysis.
|
Neuronal response strength and multiple-day response strength stability
Previous sections focused on neurons with identifiable "selectivity" and were based on comparisons of firing rates in specific poststimulus bins. However, we also compared prestimulus to poststimulus firing rates with paired t-test, allowing us to initially classify a neuron as responsive or not each day it was recorded. These comparisons best described neurons that had no identifiable selectivity or were unresponsive. Following responsiveness classification, we calculated the strength of responsiveness with an index (A – B)/(A + B), in which A = baseline firing rates and B = poststimulus firing rates. Last, we calculated the stability of these index values by dividing the smaller of 2 days' values by the larger. The resultant proportions ranged from 0 (no stability) to 1 (perfect stability). All day 1 and day 2 response strength index values for all nonselective yet responsive neurons are displayed in the scatterplot shown in Fig. 7. For recordings lasting 3 or more days, the day (n) versus day (n + 1) comparisons were considered as additional day 1 – day 2 data points and plotted on the same graph. Therefore some neurons contribute more than one data point to each graph.
|
As with measures of selectivity, we used 2-way ANOVA (P < 0.005) to find evidence of significant between-day effects and interactions. For consistently unresponsive neurons, 14/25 (56%) possible comparisons resulted in a significant main effect of recording day, indicating that about half of these neurons showed a statistically significant change in baseline firing rates. No other effects were found. For consistently responsive yet nonselective neurons, 7/19 (37%) comparisons showed a main effect of day, whereas a different 5/19 (26%) had a significant interaction effect. As with selectivity, a main effect of day signifies a nonspecific shift in both baseline and poststimulus firing rates, whereas an interaction effect signifies that difference between baseline and poststimulus firing rates was significantly different across days.
Neuronal response instability and selectivity instability
In these experiments, we first determined whether neurons were responsive, then determined whether they were selective for any task parameters. Neurons were defined as functionally unstable if they were either inconsistently responsive or inconsistently selective. Inconsistently responsive neurons (6/94) had statistically significant differences (paired t-test, P < 0.01) between prestimulus and poststimulus firing rates only on one of 2 recording days. Analyzed with 2-way ANOVA, these same neurons all showed a significant (P < 0.005) interaction between response bin (prestimulus vs. poststimulus) and recording day. Half of the unstably responsive neurons responded only on the second recording day, whereas the rest responded only on the first day. None of these was recorded for more than 2 days.
Second, inconsistently selectivity neurons (9/94) either displayed selectivity one day but not the next (6/9) or were selective for the opposite task condition (3/9) on consecutive days. In all, 15/94 (16%) neurons were inconsistently responsive or inconsistently selective. Figure 8 provides examples of inconsistent responsiveness and inconsistent selectivity.
|
Stability of neural response timing
Neuronal response onset latency was determined by using cumulative sum histograms. This analysis determines when a change in neuronal firing rate exceeded a baseline rate, typically a period 0.5 s before stimulus onset. We compared onset latency values across days to determine the stability of neuronal response timing. Of the 66 responsive neurons, 56 neurons exceeded the 99% level criterion on both days. Figure 9 represents these data as scatterplots, day 1 versus day 2. Five responsive neurons with ambiguous onset latencies were excluded from analysis.
|
200 ms). At longer latencies, the variability across days tended to increase. Nevertheless, a correlation between latency values on days 1 and 2 yielded a coefficient of r = 0.95. Cumulative sum histograms (Fig. 9, inset) were calculated for the neuron shown in Fig. 4. In addition to cumulative sum histograms, we calculated the stability of each neuron's entire perievent temporal profile. Perievent time histograms describing 1,500 ms of neuronal firing rates taken in 10-ms bins, beginning 500 ms before and ending 1000 ms after stimulus onset, were generated for each neuron each day, then compared across days with correlation analysis. Correlations between daily PETHs are higher the more firing rates increase and decrease at the same perievent moments across days. Figure 10 shows a histogram representing the number of between-day PETH comparisons resulting in correlations ranging from 0 to 1. Selective neurons have the highest correlations between consecutive daily PETHs, whereas the timing of nonselective neurons is less stable. Most inconsistently responsive/selective neurons as well as consistently unresponsive neurons have low timing stability correlations. The scatterplot in Fig. 10 shows the relationship between day 1 and day 2 comparisons and day 1 to last recording day timing comparisons. The relatively high overall correlation coefficient (r = 0.78) indicates that estimates of timing stability based on consecutive-day comparisons are representative of timing correspondence observed across 3–9 additional days.
|
Having identified and quantified stable neuronal response properties in stable testing conditions, we determined whether these same neurons responded flexibly when the monkeys learned correct responses to novel images. Learning-related responses were defined as significant changes in firing rates correlated with behavioral evidence of the monkeys' learning. We identified 2/94 (2%) neurons that responded consistently across days during the monkeys' performance of the standard task but showed significant (P < 0.01) changes in activity over time only in the learning tasks.
In Fig. 11A, the neuron's poststimulus response decreases gradually with increasing numbers of stimulus presentations. The timing of the release (Go) responses also changes; the latencies decreased with increasing trial number. Initially, the monkeys responded incorrectly by releasing the bar too slowly, but within the first 5 trials, response latencies were fast enough to be counted as correct. Importantly, neuronal responses did not simply reflect the timing of hand movements because the neuron was unresponsive when the monkeys made the grasp, hold, and release movements in a motor control condition (Fig. 11B) without visual cues. These results show that some neurons can respond consistently under rote conditions yet can respond flexibly when monkeys are required to learn new relationships between stimuli and behavioral responses.
|
Previous sections focused on neuronal responses that were significantly modulated during the tasks, but baseline firing rates also provide information about the multiple-day spiking characteristics of single prefrontal neurons. For this assessment, we grouped inconsistently responsive (6/94) neurons with inconsistently selective neurons (9/94), comparing them to all other neurons (79/94). Baseline firing rate stability was described with the index {1 – [( A – B )/(A + B)]}, in which A and B are baseline firing rates from days 1 and 2, respectively. Higher index values indicate greater stability in day-to-day baseline firing rates. The histogram in Fig. 12 shows that stability values for inconsistent neurons (white columns) are more widely distributed and therefore less stable on the whole. Median baseline stability index values were lower for inconsistent neurons (0.56) than for consistent neurons (0.84). This difference is statistically significant (t-test, P < 0.001). Median baseline stability index values were not significantly different (t-test, P = 0.9) for consistently selective neurons (0.85) compared with consistently unresponsive neurons (0.82).
|
|
DISCUSSION |
|---|
|
TOPABSTRACTINTRODUCTIONMETHODSRESULTSDISCUSSIONGRANTSACKNOWLEDGMENTSREFERENCES |
|---|
), rat hippocampus (Thompson and Best 1990), rat somatosensory cortex (Nicolelis et al. 1997), and guinea pig auditory cortex (Williams et al. 1999), neurons throughout the brain may be functionally stable. Additional support for this idea was provided by Shtoyerman et al. (2000) who used optical imaging of approximately 20 mm2 of visual cortex including parts of V1 to show that functional maps of orientation and ocular dominance columns in a macaque monkey remained stable for 1 yr. All such results are important because they begin to describe the fundamental parameters of longer-term neuronal function. For instance, our observations indicated that selectivity for task features (trial type, saccade direction) and stimuli can be maintained in spite of variation in baseline firing rates and in spite of fluctuations in the actual magnitude of that selectivity.
Identification of functionally stable prefrontal neurons agreed well with previous accounts of functionally stable neurons in other cortical regions. However, our results differed somewhat in terms of the recording techniques and duration, and the level of functional analysis. Whereas the recordings in previously mentioned studies typically lasted weeks to months, ours typically lasted 2–3 days. We attribute this to differences in technique and experimental goals. Our implants used independently adjustable electrodes that traveled 12 mm from the cortical surface. We balanced the need for stable recordings with the need to record additional neurons. To accomplish this, we isolated neurons of interest, recorded them for a few days, then advanced the electrodes. Had functional stability been the sole focus of research, it is likely that we could have recorded neurons for longer periods of time.
What is meant by functional stability?
Functional stability referred primarily to the maintenance of neuronal response types over multiple days. It was also described in terms of response strength and timing. Neurons were defined as functionally stable if they had the same type of response and selectivity from day to day. Paired t-test and 2-way ANOVA were used to determine whether neurons had a significant response, whereas a nested GLM, one-way ANOVA, and Tukey post hoc tests were used to determine whether neurons were selective for specific task conditions. These tests showed that the majority of responsive neurons remained responsive across all recording days and that the majority of selective neurons remained selective for the same factors across all recording days. Figure 4, C and D, provides an example of how even highly specific response characteristics were maintained. This neuron responded maximally to the same, single face each day out of all 72 images.
To provide a more precise description of functional stability we quantified response/selectivity strength and the stability of those measures. Following this, 2-way ANOVA identified neurons for which variations in response or selectivity strength were statistically significant. Figures 6 and 7 describe the degree of selectivity and responsiveness stability. Both histograms showed a preponderance of high index values for responsive or selective neurons demonstrating stability. For selective neurons, results summarized in Table 2 allowed us to determine that the variability in selectivity strength for 26% of the neurons was not significant. In contrast, for 41% of selective neurons, there was a significant interaction between selectivity strength and day, demonstrating that selectivity can be present across days, but may be stronger or weaker on any given day. The remaining 33% of selective neurons retained their selectivity strength, but showed a significant generalized change (increase or decrease) in both prestimulation and poststimulation firing rates across days. We equated this with a change in overall neuronal excitability, possibly dependent on the monkeys' arousal or other general factors. We took this to mean that variability in response/selectivity strength is actually important or notable for some neurons but that it did not overshadow their capacity to respond selectively for specific task conditions.
Considering that response type and strength measures were based on average firing rates over hundreds of milliseconds and therefore overlook response timing information, we used cumulative sum histograms to determine each neuron's response onset latency with respect to stimulus onset or response bar releases (hand movement timing). As with the previous measures, the majority of neurons had stable response onset latencies across days. A closer look at neuronal response timing characteristics throughout the entire average trial using correlation analysis of consecutive-day PETHs showed that increased timing variability is evident when many more data points are considered. However, many neurons still displayed notable preservation of precise timing characteristics across days (Fig. 10). Moreover, predictions of timing stability based on day 1 and day 2 PETH correlations were highly correlated with estimated timing stability based on day 1 and last-day PETHs. Combining results of all 3 analyses, functional stability was noted in terms of response and selectivity type, strength, and timing. In addition, the observed stability is likely to be an underestimate of the potential stability because we were only roughly able to control for variability in the monkeys' performance and motivation. The PETH correlation measure was particularly susceptible to this sort of variation because it was the most precise view of the data and because use of PETH depends on the assumption that the experimenter has correctly identified the event to which a neuron responds. In our tasks, Go and NoGo neurons were identifiably related to Go or NoGo trials, but their precise function was not further specified. For these neurons, PETH analysis may not have been ideal. In contrast, PETH correlation analysis of obviously stimulus or saccade locked responses results in higher correlations.
Learning-related responses
Identification of learning-related responses depended on the demonstration of waveform stability and comparison with the standard task. Demonstrable waveform stability provided evidence that changes observed during learning were not likely related to electrode drift. Comparison with the standard task allowed us to eliminate from consideration neurons that showed time-dependent changes in firing rates when learning was not required. Two neurons fit these criteria, responding consistently while the monkeys worked with highly familiar images yet responding flexibly when the monkey learned to make correct responses to novel images.
Functional instability
Unlike functionally stable and consistently unresponsive neurons, a second group of neurons consisting of about 16% of the responsive neurons appeared to be functionally unstable. These cells either responded one day but not the next, or were inconsistently selective for specific task parameters. It is possible that these results simply indicated waveform classification errors, instances in which recordings were thought to be stable when they actually originated from different neurons. It is also possible that the recordings were stable but that the neurons responded to an aspect of the task, daily environment, or the monkeys' behavior that was different between recording sessions. We controlled for differences in performance at the stage of analysis by ignoring results from recording sessions in which the monkeys made too many errors or failed to complete enough trials. However, relatively small differences in performance may still have accounted for some of the observed neuronal response instability. Either possibility could erroneously indicate neuronal response instability.
Further consideration of these issues led to the conclusion that they are specific to interpretation of inconsistent neuronal responses. The most likely alternative explanation of inconsistent neuronal responses was that they may simply have reflected the number of times apparently stable waveforms actually originated from different neurons on consecutive days (i.e., when electrode drift went unnoticed). Such problems were less likely to account for the interpretation of functionally stable responses because more than half of these neurons were recorded simultaneously from the same electrode. For electrode drift to go unnoticed, all signals would have had to appear stable. With more than one neuron recorded on a single channel, it was unlikely that the electrode could have moved into a new group of neurons without any of the waveforms appearing different. In support of this idea, all of the data on functionally unstable neurons originated from recordings in which only one neuron was recorded on any single channel. In addition, the potential problem of unnoticed electrode drift was less likely to account for judgments of functional stability because, in this data set, neighboring neurons rarely had similar response properties. They were sometimes selective for the same task parameters, yet nearly always had clearly distinguishable perievent time histograms. Therefore if the electrodes drifted imperceptibly in terms of waveform characteristics, neuronal responses would have appeared to change rather than remain the same.
Second, if some of the neurons were involved in movement speed or decision-making processes, they might have varied with the monkeys' attention or motivation. Because we had timestamps only for primary task events such as trial start, stimulus onset, response time, and reward delivery, neurons with relations to other factors may have artificially appeared to respond inconsistently. Because the PFC is noted for its diverse anatomical input and complex relations to a host of cognitive abilities (Fuster 1997) it seems unlikely that our task could provide a paradigm with which to interpret all of the neuronal response properties. It would be premature to conclude that a neuron was unstable if the events or factors driving the response were not adequately identified. In general, neurophysiological experiments are used to identify neuronal function, yet the results are often interpretable only if the experiments actually manipulated the variables to which the neurons respond.
In support of this, neurons with inconsistent response or selectivity characteristics between days often had firing patterns that did not align well with any of the available task events (this also reduced the utility of PETH correlation methodology). In contrast, many of the consistently responsive neurons had patterns of activity that did align well with task events. The potential problem of unstable testing or environmental factors does not affect interpretation of functionally stable neuronal responses to the same degree. For these neurons, stable responses were distinguishable in spite of any factors that may have differed from day to day. Experimental or environmental inconsistencies may have explained some of the observed variation in neuronal response strength and timing, but did not change the conclusion that a majority of neurons were functionally stable.
Interestingly, the responses of inconsistently responsive neurons were weaker on average (<2:1 over baseline firing rates) than those of consistently responsive neurons (>3:1 over baseline firing rates). Weaker responses are more difficult to interpret because it is not clear whether small but statistically significant effects indicate a phenomenon that is biologically plausible. In spite of these considerations, it remained possible that some neurons were functionally unstable. It has been suggested that chaotic irregularities in neural responses could prevent neuronal networks from becoming stuck in repetitive firing patterns (Liljenstrom 2003). If this is true, functionally unstable neurons might be involved in such a mechanism.
Unresponsive neurons and baseline firing rates
Our recordings also provided examples of neurons that were consistently unresponsive. Because these neurons had no identifiable function it is not clear whether these should be considered functionally stable because of the consistent lack of response or whether they provided no information about stability. The baseline firing rates of these neurons were nearly as stable as the baseline rates of the consistently responsive/selective neurons. In contrast and as a group, the baseline firing rates of inconsistently responsive/selective neurons were significantly less stable. This could mean that unstably responsive neurons were inherently more variable than those with stable response characteristics or it could simply indicate that the neurons identified as inconsistently responsive were errors based on mistakenly concluding that the same neuron had been recorded for a number of days.
We concluded that the majority of dorsolateral prefrontal neurons were functionally stable in terms of response and selectivity type, strength, and timing. Stability was emphasized, but response variability was also quantified and represented in most figures. We observed 4 different types of stable, selective responses, indicating that functional stability may be a general property of dorsolateral prefrontal neurons. This observation was particularly useful because it shows our results were not restricted to neurons of one functional type. A minority of neurons appeared inconsistently responsive or selective but, because of their generally weak responses and the two interpretation confounds previously mentioned, these results are less convincing. Last, we identified a small number of learning-related neurons, suggesting that some specialized neurons are capable of responding consistently under stable testing conditions and flexibly when learning is required.
|
GRANTS |
|---|
|
TOPABSTRACTINTRODUCTIONMETHODSRESULTSDISCUSSIONGRANTSACKNOWLEDGMENTSREFERENCES |
|---|
|
ACKNOWLEDGMENTS |
|---|
|
TOPABSTRACTINTRODUCTIONMETHODSRESULTSDISCUSSIONGRANTSACKNOWLEDGMENTSREFERENCES |
|---|
|
FOOTNOTES |
|---|
Address for reprint requests and other correspondence: P. A. Greenberg, Department of Psychology, University of Arizona, Psychology Building 68, 1503 East University Blvd., Tucson, AZ 85721 (E-mail: pag{at}u.arizona.edu).
|
REFERENCES |
|---|
|
TOPABSTRACTINTRODUCTIONMETHODSRESULTSDISCUSSIONGRANTSACKNOWLEDGMENTSREFERENCES |
|---|
Dias R, Robbins TW, and Roberts AC. Dissociable forms of inhibitory control within prefrontal cortex with an analog of the Wisconsin Card Sort Test: restriction to novel situations and independence from "on-line" processing. J Neurosci 17: 9285–9297, 1997.
Funahashi S, Bruce CJ, and Goldman-Rakic PS. Mnemonic coding of visual space in the monkey's dorsolateral prefrontal cortex. J Neurophysiol 61: 331–349, 1989.
Fuster JM. Unit activity in prefrontal cortex during delayed-response performance: neuronal correlates of transient memory. J Neurophysiol 36: 61–78, 1973.
Fuster JM. The Prefrontal Cortex: Anatomy, Physiology, and Neuropsychology of the Frontal Lobe (3rd ed.). Philadelphia, PA: Lippincott–Raven, 1997.
Gray CM, Maldonado PE, Wilson MA, and McNaughton BL. Tetrodes markedly improve the reliability and yield of multiple single-unit isolation from multi-unit recordings in cat striate cortex. J Neurosci Methods 63: 43–54, 1996.
Komatsu H and Ideura Y. Relationships between color, shape, and pattern selectivities of neurons in the inferior temporal cortex of the monkey. J Neurophysiol 70: 677–694, 1993.
Kralik JD, Dimitrov DF, Krupa DJ, Katz DB, Cohen D, and Nicolelis MAL. Techniques for long-term multisite neuronal ensemble recordings in behaving animals. Methods 25: 121–150, 2001.[CrossRef][ISI][Medline]
Liljenstrom H. Neural stability and flexibility: a computational approach. Neuropsychopharmacology 28: S64–S73, 2003.
Murray EA, Bussey TJ, and Wise SP. Role of prefrontal cortex in a network for arbitrary visuomotor mapping. Exp Brain Res 133: 114–129, 2000.[CrossRef][ISI][Medline]
Nicolelis MAL, Ghazanafar AA, Faggin BM, Votaw S, and Oliveria LMO. Reconstructing the engram: simultaneous, multisite, many single neuron recordings. Neuron 18: 529–537, 1997.[Cr
