Theta-activity in anterior cingulate cortex predicts task rules and their adjustments following errors

Behavioral Adjustment Following Task Switches.

We recorded in 39 sessions the local field potential (LFP) from one to five neuronal groups (28 and 11 in two monkeys, respectively). Monkeys performed, on average, 74.9% (78.1%) correct on the antisaccade (i.e., prosaccade) task, with lower accuracy in trials immediately following the task change (Fig. S1B; see ref. 7). Quantifying correct performance for successive sets of five trials with correct responses following the rule change revealed that accuracy asymptotes within 10 trials (Fig. S1B). In contrast to accuracy, saccadic reaction times remained constant with pro- and antisaccade latencies averaging to 199 and 228 ms, respectively (Fig. S1B). These behavioral results show that the behavioral adjustment required by the change in SR rule primarily affected preparatory processes required to instantiate and maintain a new SR rule, rather than a readjustment of movement parameters, which would be expressed as variations in saccade latency.

Task Selectivity of Theta-Activity.

Applying the correct SR mapping rule at the time the target stimulus is presented requires advance preparation of neuronal circuitry to selectively represent the relevant task rule. We therefore calculated the evolution of oscillatory activity during the preparatory period from 3 to 30 Hz and quantified task selectivity for each neuronal site. Fig. S1C shows the average evolution of LFP power for an example site for the pro- and antisaccade task (Fig. S1C Upper and Lower, respectively), illustrating prominent theta-frequency activity, which decreases in strength on prosaccade trials and increases during antisaccade trials. We quantified this difference by calculating a task selectivity index (TSI) for each time-frequency data point and derived a statistical map with a cluster randomization approach correcting for multiple comparisons (Materials and Methods). For the example shown in Fig. S1C, task selectivity was significant for a 5- to 10-Hz theta-frequency band well in advance of peripheral stimulus presentation. The peak TSI value of 0.3 corresponds to 85.7% stronger LFP theta-power during the antisaccade task.

Across all sites, the emergence of task selectivity in the theta-band was the most prominent preparatory signature (Fig. 1A), predicting which SR rule was established before the onset of the stimulus. Across all 67 LFP sites, 23.8% (16) showed significant task selectivity during the preparatory time-frequency window in the theta-band, with 13.4% (10.4%) of sites showing enhanced theta-power when preparing the SR rule for the prosaccade (antisaccade) task. The proportion of task selectivity for either task was statistically not different (binomial test, P > 0.05). Sites selective for the anti- and prosaccade task overlapped spatially with a slight trend for antisaccade selective tasks to be more likely more anterior in area 24c. Selecting only those sites with a significant effect in the time-frequency window with the most prevalent task effects (5–10 Hz during the −0.4 to 0 s before stimulus onset) illustrates that task selectivity emerged at approximately −0.4 s before stimulus onset (Fig. 1B), and showed similar time-frequency dynamics for sites preferring pro- and antisaccades (Fig. 1C). This finding also illustrates that task selectivity of preparatory theta-activity is not confounded by a difference of the two tasks in terms of task difficulty.

Frequency Specific Trial-by-Trial Prediction of Task Rules.

To analyze whether frequency-specific oscillatory activity can actually be used by neuronal circuitry to predict the forthcoming interpretation of a stimulus similar to firing rates (see ref. 7), we conducted a receiver operating characteristic (ROC) analysis. Fig. 2A illustrates that an ideal observer could predict which SR mapping will be applied before the stimulus appeared only when relying on theta-activity (5–10 Hz; one-sample t test, P < 0.05), but not when considering lower (2–4 Hz) or higher frequency bands in the low and high β-range (11.5–15.5 Hz and 18.5–22.5 Hz). The average ROC value began to be statistically significantly different from chance (ROC value of 0.5) already at −0.4 s before stimulus onset (one-sample t test, P < 0.05), with no other frequency band reaching statistical significance before stimulus onset (all P > 0.05; see Fig. 2B). Note that spike rates in ACC showed likewise the earliest significant ROC prediction at −0.4 s, but only in those trials immediately following the task rule change. In later trials, ACC spike rates distinguished both tasks at successively later times before stimulus onset (see ref. 7). We therefore analyzed the latency of task selectivity for successive subsets of five correct trials relative to the task rule change by measuring the time when the ROC prediction became significant (P < 0.05) for two successive time windows. Fig. 2C shows that task selectivity of LFP theta-activity remained rather constant across trials, emerging at −0.35 s in trials early after the task rule change, and at −0.4 s thereafter. No other frequency range than theta showed significant task selectivity before stimulus onset in any of the trial subsets. Likewise, the overall strength of the TSI in the −0.4 to 0 s time window before stimulus onset did not decrease after the initial task rule change (Fig. 2D). These findings contrast to the change in firing rate latencies reported in ACC before (7), illustrating that rate effects and LFP theta reflect different signals. Interestingly, significant task selectivity is available in spike rates at −0.4 s even after the fifth trial in the same task, not in ACC, but in other nodes of the cognitive control network (lateral PFC; see ref. 7). Thus, the early emerging task selectivity of LFP theta-activity in the ACC may well reflect information available from input of other nodes of the cognitive control network, but without being necessarily translated into modulated overall spike counts (as described further later).

Theta-Activity and Behavioral Adjustment.

The previous result leaves open whether the observed task-selective theta-activity reflects a genuine functional contribution from ACC with respect to cognitive control or is primarily inherited from other nodes of the network. In a first analysis addressing this issue, we compared task-selective theta-activity in correct trials and error trials, finding that task selective theta-activity was absent on error trials, in which the incorrect SR mapping rule had been applied to the peripheral stimulus (paired t test, P < 0.01; Fig. 3 A and B; note that we excluded from analysis the first trial after the task switch, which was per definition incorrect, but not indicative of a failure to maintain a SR rule). In a second analysis we tested, whether task-selective theta-activity in ACC reflects the adjustment of task rule representations following erroneous SR mapping. Previous results suggest that spike rates in ACC signal the behavioral outcome of a trial and provide the critical information to overcome erroneous representations well before other areas of the cognitive control network (5, 7, 15). We therefore analyzed task-selective theta-activity in correct trials that followed a previous error trial (i.e., E-C) with correct trials following other correct trials (i.e., C-C). Fig. 3C shows that task-selective theta-activity in E-C trials is similar in strength, but emerges earlier compared with C-C trials. To quantify this latency difference, we calculated the average TSI of theta-activity (5–10 Hz) within successive 0.1-s windows during the preparatory period. Fig. 3D reveals that task selectivity rose earlier in E-C trials, becoming statistically significant already within −0.4 to −0.3 s before stimulus onset (one-sample t test, P < 0.05). In contrast, task selectivity in C-C trials became significant later starting between −0.2 to −0.1 s before stimulus onset (Fig. 3D).

Consequences of Task-Selective Theta-Activity for Spike Output.

To this point, task selectivity in the ACC was shown to emerge in a time- and frequency-selective LFP theta-band. It emerged early (at approximately −0.4 s before stimulus onset) even after the fifth trial following the task switch, i.e., where spike counts in ACC do not distinguish anymore between the tasks. These findings suggest that enhanced LFP theta-activity is functionally relevant, but rather than being translated into a higher spike count, it may provide a critical reference to structure spike output in time (29). To test this hypothesis, we analyzed for a subset of 22 most isolated single ACC neurons whether their spike output is phase locked (i.e., phase consistent) to LFP theta-activity and whether spike–LFP phase consistency increases during periods of enhanced LFP theta-activity. The analysis of spike–LFP phase consistency is nontrivial for neurons firing only few spikes (few samples) because common phase locking measures are biased toward higher values the smaller the sample size, rendering the phase locking statistics less reliable and imposing a serious confound when comparing conditions (here, tasks) with different sample sizes. However, even for neurons firing only sparsely in ACC, we observed that spikes often occur at particular phases of the theta-cycle. For illustration, Fig. 4A shows example traces of spikes of one sparsely firing neuron and the LFP (5–10 Hz passband) during three pro- and antisaccade trials. This example neuron is selected because it visualizes (i) that theta-power just before stimulus onset is stronger in only one (antisaccade) task, (ii) that spikes tend to occur around the time of the theta-cycle peak, rather than independent of the theta-cycle, and (iii) that this phase specificity becomes evident in the last 0.4 s before stimulus onset only for the antisaccade trials and not for the prosaccade trials. Measuring spike–LFP phase consistency across the whole sample of neurons for spikes within the 0.4-s window before stimulus onset showed that the spiking output of 31.8% of the neurons (7 of 22) were significantly phase-consistent to the LFP (Fig. 4B; Rayleigh test for circular homogeneity, P < 0.05). Among the phase-consistent neurons, three were significantly phase-locked in both tasks, but for four neurons spike–LFP phase consistency was task-selective (Fig. 4B). Critically, phase consistency tended to be stronger when LFP theta-activity was stronger (Fig. 4C). To test this, we median split for each neuron (n = 22) the theta-power distributions of the −0.4 to 0 s preparatory period from each task and calculated spike–LFP phase consistency separately for the subsets with high and low theta-activity and found that spike–LFP phase consistency was significantly stronger for the antisaccade task when theta-activity was stronger (P < 0.05, paired t test; Fig. 4C). For the prosaccade task the same pattern was observed, but the effect remained nonsignificant (P = 0.3, paired t test). Thus, there is overall evidence that the reported findings pertaining to task-selective LFP theta-activity have consequences for spiking output within ACC.

Experimental Procedures and Paradigm.

We collected data in two macaque monkeys following guidelines of the Canadian Council of Animal Care policy on the use of laboratory animals and the University of Western Ontario Council on Animal Care. Extracellular recordings commenced through a recording chamber with one to five tungsten electrodes on a daily basis, guided by an anatomical MR image to ensure targeting the ACC (area 24c; for details see ref. 7). Data acquisition and filtering were done with a multichannel processor (Plexon), using a headstage with unit gain. The LFP was extracted with a passband filter (0.7–170 Hz), further amplified and digitized at 1 kHz. The powerline artifact was removed from 10-s-long data segments using a discrete Fourier transform filter as described before (36). Behavioral control and stimulus generation were accomplished with the cortex software package, monitoring eye position at 1 kHz with a scleral search coil (David Northmore). Each trial began with a centrally presented, white fixation spot, which was fixated for a variable period of 1.1 to 1.4 s Upon its offset, a peripheral white stimulus appeared (8° to the right or left with equal probability), signaling to the monkey to saccade within 500 ms either toward (i.e., prosaccade) or away from it to the mirror location (i.e., antisaccade) to receive a juice reward. The SR mapping rule (pro- vs. antisaccade) changed every 30 correct trials without overt cue.

Data Analysis.

Analysis was performed with custom Matlab code (Mathworks), using the fieldtrip toolbox (http://www.ru.nl/fcdonders/fieldtrip/). We limited all analysis to the preparatory period from −0.75 to 0.15 s around the peripheral stimulus onset excluding stimulus-onset locked and saccade-related potentials occurring later in the trial. Time-resolved oscillatory activity was calculated, after subtracting the stimulus locked average potential, as LFP power from 3 to 30 Hz based on Hanning tapered Fourier transforms in ±0.334 s time windows calculated every 0.05 s (see ref. 36). To calculate statistical differences between task conditions, we applied a nonparametric cluster-based randomization approach based on the TSI as our test statistics and corrected for multiple comparison as described and validated before (see ref. 66). The TSI is calculated as [(pref task – None-pref task)/(pref task + none-pref task)], resulting in values of ±1 with positive values revealing stronger power for the preferred task. We calculated spike–LFP phases based on the Hanning tapered Fourier transforms of the LFP around ±0.334 of each spike. To measure phase consistency we calculated the average pairwise phase difference between spike-LFP phases, resulting in a completely bias-free measure of phase consistency, which is linearly related to the more commonly used phase locking value. The normalized pairwise phase consistency takes on values between 0 (random phase distribution) and 1 (complete consistency). For details, see ref. 67 and SI Materials and Methods.