Dynamic coding for cognitive control in prefrontal cortex

Abstract

Cognitive flexibility is fundamental to adaptive intelligent behavior. Prefrontal cortex has long been associated with flexible cognitive function, but the neurophysiological principles that enable prefrontal cells to adapt their response properties according to context-dependent rules remain poorly understood. Here, we use time-resolved population-level neural pattern analyses to explore how context is encoded and maintained in primate prefrontal cortex and used in flexible decision making. We show that an instruction cue triggers a rapid series of state transitions before settling into a stable low-activity state. The postcue state is differentially tuned according to the current task-relevant rule. During decision making, the response to a choice stimulus is characterized by an initial stimulus-specific population response but evolves to different final decision-related states depending on the current rule. These results demonstrate how neural tuning profiles in prefrontal cortex adapt to accommodate changes in behavioral context. Highly flexible tuning could be mediated via short-term synaptic plasticity.

Figure 1

Task, Recordings, and Overall Activity Profile (A) Arbitrary associations between cue and target stimuli were learned prior to the recordings reported here. (B) The cue stimulus at trial onset determined the current target. There followed a series of 0–3 nontargets, followed by the cued target, all in the same location (right or left of fixation, randomly varying across trials). The animal was required to maintain central fixation until offset of the target and was then rewarded for a saccade to the stimulus location. Each nontarget was randomly either a neutral stimulus (a stimulus never used as a target (see A) or a stimulus paired with a different cue and hence serving as a target on other trials (here termed distractor). The presentation order in the task schematic is illustrative only. Actual presentation was randomized but always ended with the target. (C) Schematic diagram of recording sites, illustrated by red and blue symbols for monkeys A and B, respectively. Recording sites for monkey A (right hemisphere) have been transferred to the left. (D) Mean firing rate of the recorded population during each epoch of the task, as a function of time from each stimulus onset. The duration of stimulus presentation is indicated by the gray rectangle. Significantly elevated firing rate relative to 100 ms pretrial baseline is indicated by the thin (p < 0.05) and thick (p < 0.0001) blue line.

Figure 2

Neural Population Dynamics (A) Schematic of two different trajectories through a three-dimensional state space. The distance between two condition-specific states at time t reflects the multidimensional distance in the population response: d(P1_t, P2_t). Within-condition distance between earlier and later states represents change in position as a function of time, i.e., velocity: d(P1_t-n, P1_t+n)/2n. (B) The mean multidimensional distance between the three trial types is shown in blue as a function of time, and the significant periods of above-chance discrimination are indicated by the significance bar along the x axis (p < 0.05, cluster-based correction of multiple comparisons, see Experimental Procedures). For reference, the overall mean activity level (network energy) is shown in gray (right axis) and gray the significance bar indicates above-baseline activity (p < 0.05). (C) Bar plot of the multidimensional distance between trial types calculated for the first delay period in a trial (delay preceding first choice stimulus), a second delay period (present only when first choice stimulus was a nontarget), and a third delay period (present only when first two choice stimuli were nontargets). Error bars represent 95% confidence intervals. (D) Multidimensional distance between trial types is visualized using multidimensional scaling (MDS). Data points are four independent estimates of the population response associated with each trial type, plotted as a function of the first two dimensions at 0 ms, 250 ms, 500 ms, and 750 ms from cue onset. The complete time course of trial type-specific clustering in state space is available as Movie S1. (E) The top panel plots the estimated instantaneous velocity through multidimensional state space for each trial type, as a function of time. The bottom panel shows the equivalent velocity plot for overall energy change.

Figure 3

Pattern Classification (A) The pattern classifier is trained to discriminate trial types based on the population response observed within a 50 ms time window and validated using test data at the same equivalent time window (within-time classification) or at a different time window (cross-temporal classification; see Figure 4). (B) Mean within-time classification index between trial types is shown in blue as a function of time. For reference, the overall population mean firing rate is shown in gray (right axis). The blue bar along the x axis indicates periods of above-chance classification, and gray significance bar indicates above-baseline activity (p < 0.05).

Figure 4

Cross-Temporal Pattern Analysis (A) Cross-temporal pattern classifiers were trained to discriminate trial type using data from the shaded 50 ms time window and tested throughout the cue duration and subsequent delay period. For reference, within-time classification performance is shown in gray to illustrate the upper limit of trial type information at each time point and significant periods of above-chance cross-temporal classification are indicated by the color-coded significance bar below the corresponding trace. The complete time course of these cross-temporal analyses is available online as Movie S2. (B) Cross-temporal classification results from (A) overlaid. (C) Cross-generalization was extended to the target epoch to test for prospective target coding following same conventions as (B), with gray trace reflecting the within-time classification for target type. (D and E) Full cross-temporal classification matrix. Classifiers are trained to discriminate trial type at each 50 ms time window (1 ms increments) during the cue and first delay period, and trial type discrimination is tested throughout the cue (D), first delay, and target period (E).

Figure 5

Trial Type Coding in Patterns Driven by a Fixed Neutral Stimulus (A) Multidimensional distance between the network responses to neutral stimuli presented within each trial type is plotted as a function of time, and the blue bar along the x axis indicates periods of above-chance classification. (B) The multidimensional distance is visualized using the first two dimensions defined by MDS. Each data point reflects an independent sample of the population response to neutral stimuli presented within each trial type (color coding as in Figure 1D). The complete time course of the temporary emergence of condition-specific clustering in state space is available online as Movie S3.

Figure 6

Evolution of Coding during Choice Processing (A) Cross-generalization pattern analysis reveals the time course of stimulus-dependent coding (in gray) and context-dependent coding (in black). Significant periods of above-chance discrimination are indicated by corresponding significance bars along the x axis. (B) State space is represented by the first two dimensions from MDS. Data points correspond to four independent estimates of the population response to each of the choice stimuli presented as a target (filled) or distractor (unfilled). Blue, stimulus serving as target with cue 1; red, stimulus serving as target with cue 2; green, stimulus serving as target with cue 3. The full time course of the coding transformation in state space is available online as Movie S4. (C) Evidence for a “go” or “no-go” decision to choice stimuli is plotted as a function of time after the presentation of each stimulus type within each trial type (separate plots). Color coding for the three stimuli are as in (B). Choice-related evidence is traced in heavy lines for “go” stimuli (i.e., targets) and in a thin line for “no-go” stimuli (i.e., distractors). Evidence for each choice is quantified relative to an independent reference pattern that differentiates go/no-go stimuli at the end of the trial: target minus distractor. Consequently, positive values reflect positive evidence for a “go” decision, whereas negative values reflect positive evidence for a “no-go” decision. Also see Figure S1 in online Supplemental Information.

Figure 7

A Simple Schematic of the Proposed Tuning Mechanism Depending on the context, input matching a particular choice stimulus is routed along a context-dependent trajectory toward an activity state that codes behavioral choice.