The Role of Locus Coeruleus in the Regulation of Cognitive Performance

See allHide authors and affiliations

Science  22 Jan 1999:
Vol. 283, Issue 5401, pp. 549-554
DOI: 10.1126/science.283.5401.549


Noradrenergic locus coeruleus (LC) neurons were recorded in monkeys performing a visual discrimination task, and a computational model was developed addressing the role of the LC brain system in cognitive performance. Changes in spontaneous and stimulus-induced patterns of LC activity correlated closely with fluctuations in behavioral performance. The model explains these fluctuations in terms of changes in electrotonic coupling among LC neurons and predicts improved performance during epochs of high coupling and synchronized LC firing. Cross correlations of simultaneously recorded LC neurons confirmed this prediction, indicating that electrotonic coupling in LC may play an important role in attentional modulation and the regulation of goal-directed versus exploratory behaviors.

Neuromodulators, such as norepinephrine (NE) and dopamine, have long been thought to play a role in regulating nonspecific aspects of behavior, such as motivation and arousal. However, recent evidence indicates that these systems may play a more specific role in task-related cognitive processes. Brainstem dopaminergic neurons respond selectively to stimuli that predict reward (1). Stimulus-specific activity has also been observed in LC neurons. Recent studies found that LC neurons in monkeys performing a visual discrimination task exhibited short-latency stimulus-evoked (phasic) responses to target (CS+) stimuli but not to distractor (CS−) stimuli or other task events (2). The latencies of these LC responses were substantially shorter than, and temporally correlated with, the latencies of corresponding behavioral responses, indicating that LC activity may affect task responding. However, the mechanisms that govern LC activity or its effect on behavior have remained unclear.

LC neurons were recorded in four Cynomolgus monkeys performing a visual discrimination task (2). This task required the monkey to respond to infrequent visual target stimuli but not to frequent distractors (3). In many of our recordings, LC neurons changed levels of tonic discharge several times (Fig. 1A), in association with alterations in task performance. We divided behavioral performance into epochs of “good” and “poor” performance, on the basis of the frequency of false alarm (FA) errors produced [as described previously (2, 4)]. Signal detection sensitivity (d′) was substantially higher in epochs of good compared with those of poor performance, so the difference between these cannot be explained by a simple change in response criterion (4). Furthermore, although mean lever response times (RTs) were not systematically different between the two levels of performance, there was a significant narrowing of the distribution of lever release latencies during the good epochs (4) (Fig. 1B).

Figure 1

Representative data from a typical LC neuron recorded in a monkey during performance of the visual discrimination task. (A) The rate of discharge for an LC cell (top curve) and the number of FAs (bottom curve), both integrated for a sliding window of 20 s. (B) Normalized distributions for behavioral response latencies (lever releases) during “good” epochs (solid line) compared with “poor” epochs (dashed line) averaged for sessions in three monkeys. Similar distributions were obtained for the individual monkeys, and the distributions were consistently more narrow during good epochs. (C to F) Poststimulus time histograms (PSTHs) for LC activity during the visual discrimination task. (C and D) Response for targets. (E and F) Response for distractors. (C and E) “Good” behavioral epochs. (D and F) “Poor” behavioral epochs (FA rate typically > 7%). Stimuli occur at time zero. All histograms are normalized to a standard of 100 trials. Similar results were obtained in another 29 single-cell recordings and 37 multicell recordings.

As shown in Fig. 1A, epochs of poor performance were associated with significantly higher tonic LC activity than were epochs of good performance (3.0 ± 0.3 spikes/s compared with 2.0 ± 0.2 spikes per second P < 0.001; FA frequencies were 7.6 ± 0.9% compared with 1.0 ± 0.2% of trials;P < 0.01; n = 30 cells; pairedt tests). Similar results were obtained in an additional 37 multicell LC recordings. Thus, in addition to our previous finding of a close relation between phasic LC discharge and behavioral responses (2), we also found a close relation between the level of LC tonic activity and behavioral performance. We refer to the lower level of tonic LC activity during epochs of good performance as “intermediate,” to distinguish it from the low (near zero) level typically associated with drowsiness or sleep (2,5).

We also found that sensory-evoked LC responses varied with the level of tonic LC activity and task performance. The phasic responses that LC neurons exhibit selectively for target stimuli in this task occurred almost exclusively during epochs of intermediate tonic LC activity and good task performance (Fig. 1, C to F). For the 30 single-cell recordings described above, response magnitudes to target stimuli during epochs of good performance were significantly greater than during epochs of poor performance (2.7 ± 0.4 compared with 0.8 ± 0.2; P < 0.001; paired t test). Thus, increased tonic LC discharge was associated with decreased responsivity of LC neurons to target stimuli as well as decreased task performance. This three-way association of tonic LC activity, LC phasic responses to target stimuli, and level of task performance was observed consistently across our recordings.

These results suggest that there is a precise relation between LC activity and behavioral performance. To elucidate the mechanisms that might underlie this relation, we developed a computational model of LC function and its effect on performance in this task.

The model is a hybrid, with two primary components: an LC network and a stimulus discrimination (behavioral) network (Fig. 2A). The LC network is relatively fine-grained and designed to simulate physiological mechanisms underlying LC function, whereas the behavioral network is the simplest capable of simulating performance in the visual discrimination task. Although the use of such a hybrid model that combines components at different levels of abstraction may be unusual, this is justified by the correspondence between each component of the model and the level of the phenomena it addresses.

Figure 2

(A) Architecture of the model of task performance. Arrows represent excitatory links and small circles represent inhibition. There is a moderate positive bias on the response unit, which captures the observation that monkeys in this task make many FAs but very few misses (2). (B) Dynamic trajectories of the target, distractor, and response assemblies (as indicated), in response to targets (top) or distractors (bottom). Stimulus presentation is at time zero. Solid lines, target unit; dashed lines, distractor unit; dotted lines, response unit. In response to each stimulus, there is partial activation of both target and distractor decision units due to their overlapping connections with the input. However, because of mutual competition, after about 100 ms, the decision unit corresponding to the activated stimulus typically prevails, and the competing unit is suppressed. When the target unit prevails, the activity of the response unit is driven above threshold, and a response is recorded. FAs occur because of noise in the response unit, which interacts with transient activation of the target decision unit by a distractor stimulus to produce a response. A threshold is set for activation of the response unit (0.6).

The LC network is a population of 250 spiking neurons, each of which is a leaky integrate-and-fire cell (6, 7) that exhibits temporal dynamics similar to those in compartmental models (8). In the model, LC cells interact with each other in two ways. First, lateral inhibition simulates the effect of local NE release (9, 10). Second, a voltage-dependent interaction among LC units simulates the effects of hypothesized electrotonic coupling among LC neurons (11). In addition, each LC cell receives input from the behavioral network (see below), as well as noise that is responsible for a spontaneous firing rate of about 1 spike/s [as observed in vivo (2)].

The behavioral component of the model is a simple connectionist network, consisting of two input units (one for target and one for distractor stimuli), two corresponding decision units, and one response unit (Fig. 2A). Connections between units in different processing layers are excitatory (reflecting information flow), connections within a layer are inhibitory (competition), and the activity of units is subject to small random variations (noise) (12). Each input unit has a strong weight to the corresponding decision unit and a weaker projection to the other decision unit. The target decision unit has a positively weighted connection to the response unit and to the LC network (13). Finally, consistent with previous simulation work, we assume that NE release has the effect of increasing the gain of the activation function for units in the decision and response layers [see below and (14, 15)].

A task trial was simulated by activating the input unit corresponding to the current stimulus, which resulted in the spread of activation to the competing units in the decision layer and then to the response unit and LC. Characteristic dynamic responses of different units in the behavioral network after presentation of each type of stimulus (in the absence of modulation by LC) are displayed in Fig. 2B.

The simulated pattern of LC firing with and without electrotonic coupling, after target and distractor stimuli, is shown in Fig. 3. Target stimuli evoke a transient, synchronized LC response as a result of input from the target decision unit to LC cells. The target-evoked response is terminated by NE-mediated collateral inhibition within the LC. Electrotonic coupling among LC neurons has two main effects. First, coupling causes a stronger response of the LC population to target inputs, as a result of the reinforcement of spike-induced depolarizations in each individual neuron by similar, simultaneous depolarizations in other LC cells within the population. Second, coupling reduces the spontaneous (tonic) firing rate of LC cells by mutually shunting the effect of uncorrelated noise on each cell's membrane potential (16). These simulation results closely resemble the patterns of monkey LC discharge observed during epochs of intermediate (versus high) tonic activity and good (versus poor) behavioral performance (Figs. 1 and 3).

Figure 3

(A to D) PSTHs for the simulated data. (A and B) Response to targets. (C and D) Response to distractors. (A and C) Coupling among LC neurons. (B and D) No coupling among LC neurons. PSTHs are normalized for 100 trials, as for the empirical data (see Fig. 1). (E) Response time distributions for model responses (response unit activations) after targets. Solid line, distribution during simulated coupling among LC neurons; dashed line, distribution during no coupling among LC neurons. The difference of about 150 ms between the latencies of empirical (Fig. 1B) and simulated behavioral responses (E) is consistent with a residual sensory or motor latency.

As noted above, the output of the target decision unit provides input to the LC network, whereas LC activity modulates the gain of units in the decision and response layers of the behavioral network. Unlike in previous models, where the effect of catecholamines on cognitive performance was modeled as a fixed gain parameter throughout a simulation (15), here the value of the gain was determined dynamically by the output of the LC network. Thus, the synchronized, transient responses to target stimuli during epochs of high coupling (Figs. 1, C to F, and 3, A to D) resulted in a temporally modulated process. The effect of LC on the performance of the behavioral network can be seen by comparing the activation of the response unit under conditions of high and low coupling among LC neurons. Increased coupling among LC neurons produced a reduction in FAs (from 12 to 2%), without an increase in misses, and a significant narrowing of the RT distribution, without a change in the mean (Fig. 3E). Thus, a change in coupling among LC neurons in the model reproduces the changes in LC activity, behavioral performance, and the relation between these that is observed empirically.

Changes in electrotonic coupling produce the associated changes in behavior for several reasons. First, increased coupling reduces tonic LC activity, reducing NE release in the behavioral network and thereby lowering the responsivity of those units. For the response unit, this is equivalent to raising its threshold (15), which reduces the number of FAs and anticipatory responses. Ordinarily, raising the response threshold would also increase the number of misses and lengthen mean RT. However, increased coupling enhances evoked LC responses to target stimuli. The enhanced LC response produces a transient reduction in threshold specifically and shortly after target stimuli, which compensates for the overall increase in response threshold and potentiates the processing of target stimuli. This averts an increase in misses or RT (17). This temporal modulation of processing, with maximal gain occurring shortly after a target stimulus, is consistent with an attentional window reported in the cognitive literature (18) and also with recently proposed mechanisms for attentional modulation based on neural synchrony (19). Moreover, this mechanism has the combined effect of eliminating anticipations and of speeding up slow responses, explaining the observed narrowing of the RT distribution during the good behavioral epochs (Figs. 1B and 3E). Thus, a change in a single parameter (an increase in coupling within LC) can account for the reduction in tonic LC activity, the enhanced target-evoked phasic responses, and the association of this pattern of LC activity with a reduction of FAs and a tightening of the RT distribution in behavioral performance.

The model makes the prediction that improved performance is associated with increased electrotonic coupling and therefore should also be associated with greater synchrony in the spontaneous firing of LC neurons (Fig. 4B) (20). We tested this prediction by comparing cross correlograms generated for pairs of simultaneously recorded LC neurons during epochs of good and poor performance. Consistent with our prediction, we found that 18 of 23 pairs of recorded neurons exhibited a central peak in cross correlograms during epochs of good performance that was not present for the same neurons during poor performance (Fig. 4, A and B). Quantitative analyses of correlograms for these 23 pairs of cells indicated that the central peak during good performance was significantly greater than during poor performance.

Figure 4

Cross correlograms for two simultaneously recorded LC neurons during epochs of good (filled histograms) versus poor (contour lines) performance for the data (A) and for coupling versus no coupling, respectively, in the model (B). Note the central peak indicating synchronous activity during good performance (coupling), which is not present during poor performance (no coupling). Average amplitudes of central peaks in the data correlograms were compared as described in (30). Epochs analyzed were at least 500 s in duration. Epochs of target-evoked responses were eliminated by omitting 0.8 s of activity from each cell after target stimuli to avoid any bias produced by the synchrony of firing associated with phasic LC responses that occur primarily during epochs of good performance (coupling); this had no apparent effect on the cross correlograms obtained. Central peaks were significantly greater during good compared with poor epochs (8.08 ± 2.96 compared with 2.64 ± 1.08, respectively;P < 0.05; n = 23 pairs of neurons), whereas the spontaneous rate (baseline) was higher for the poor epochs. Similar results were obtained for pairs of neurons recorded from the same electrode and from two different electrodes.

Our simulation results suggest that electrotonic coupling may be an important mechanism underlying patterns of LC activity and may play a role in regulating behavioral performance. Strong evidence for coupling within the LC of neonatal rats has been reported (21). Although electrotonic coupling appears to decrease postnatally, recent studies indicate that coupling may persist in the LC of the adult rat (22, 23). However, the presence of such coupling in the adult primate has not yet been empirically demonstrated. The model we have developed, together with the data regarding synchronization of LC activity, support this possibility and indicate that modulation of electrotonic coupling may produce potent effects on behavioral performance (24).

One important question concerns the adaptive advantage of the changes in behavior that are produced by changes in LC activity. In our model, intermediate tonic LC activity (due to increased coupling) facilitates a state of selective responding. This state is beneficial in a stable environment such as in our experimental task, where the source of reward is predictable and the behaviors relevant for acquiring it are known and consistent. However, what are the advantages of high tonic LC activity, which is associated with impaired performance in our experimental task? One possible answer is that heightened selectivity may at times be disadvantageous, such as in an uncertain or stressful environment, in which unexpected but imperative stimuli occur (for example, prey suddenly faced with a predator), or when previously reinforced responses lose their reward value (for example, satiety). Such circumstances require reevaluation of the sensory environment and abandonment of current behaviors in the search for more adaptive ones. This ability may also be critical for normal developmental and learning processes, as suggested by recent findings indicating that the best predictor of success in acquiring a new skill is not the speed with which the correct behavior is first discovered but the number of alternatives that are initially explored (25). According to our model, high tonic LC activity (as a result of low coupling) can provide a mechanism for sampling new stimuli and behaviors by reducing attentional selectivity and increasing behavioral responsiveness to unexpected or novel stimuli.

These considerations suggest that a tension exists between optimizing performance in a stable environment and favoring more flexible behavior in a changing or unfamiliar environment or when current rewards lose their value. This is a fundamental trade-off, which has been recognized in computational theories of reinforcement learning that distinguish between states that favor “exploitation” of existing behavioral routines versus “exploration” of new ones (26). The mechanisms responsible for shifting between such states have not been specified. Our model indicates that changes in the mode of LC functioning (produced by alterations of electrotonic coupling) may provide a neural mechanism for mediating such shifts. This hypothesis also helps to integrate previously proposed roles for LC function (27). Future research is needed to directly test this hypothesis (28) and to characterize the neural system or systems providing input to the LC that are responsible for monitoring the current behavioral context and altering coupling among LC neurons when shifts of state are appropriate. It will also be important to determine the relation of the LC-NE neuromodulatory system to others, such as the dopamine system, that are thought to regulate behavior based on expectations about future events (29).

  • * To whom correspondence should be addressed. E-mail: gaj{at}


View Abstract

Navigate This Article