Slower prefrontal metastable dynamics during deliberation predicts error trials in a distance discrimination task

Abstract

Cortical activity related to erroneous behavior in discrimination or decision-making tasks is rarely analyzed, yet it can help clarify which computations are essential during a specific task. Here, we use a hidden Markov model (HMM) to perform a trial-by-trial analysis of the ensemble activity of dorsolateral prefrontal cortex (PFdl) neurons of rhesus monkeys performing a distance discrimination task. By segmenting the neural activity into sequences of metastable states, HMM allows us to uncover modulations of the neural dynamics related to internal computations. We find that metastable dynamics slow down during error trials, while state transitions at a pivotal point during the trial take longer in difficult correct trials. Both these phenomena occur during the decision interval, with errors occurring in both easy and difficult trials. Our results provide further support for the emerging role of metastable cortical dynamics in mediating complex cognitive functions and behavior.

Keywords: decision making; errors; hidden Markov model; metastable dynamics; prefrontal cortex; spatial discrimination.

Figure 1.. Experimental paradigm and neural recordings sites

(A) Sequence of task events within a trial. Each trial started when the monkey touched the central switch, leading to the appearance of the central stimulus (reference point), which lasted for 400 or 800 ms (pre-stimulus; pre-S). After this pre-stimulus period, the first stimulus (S1) was presented. S1 was followed by a variable delay (first delay; D1) of 400 or 800 ms, which lasted until the second stimulus (S2) appeared. S2 was followed by a second delay (D2) of 0, 400, or 800 ms. Both S1 and S2 were presented for 1,000 ms and placed either above or below the reference point at 8 to 48 mm (8-mm step) from the reference point. After both stimuli reappeared (placed horizontally), the monkey had to touch the switch below the stimulus that appeared farther from the reference point (the blue circle in the example trial). Correct responses were rewarded with 0.1 mL fluid, while errors were followed by an acoustic feedback. The stimulus feature (blue circle/red square), position (above/below the reference point), distance, and target position (left/right) were pseudo-randomly selected. (B) Penetration sites. Composite from both monkeys, relative to sulcal landmarks. See STAR Methods for details.

Figure 2.. Representative trials of a neural ensemble of simultaneously recorded cells

Each raster plot indicates the spiking activity of each recorded neuron from 400 ms before S2 presentation until the beginning of the following trial. Colored curves represent the posterior state probabilities, and the assigned states are indicated with colored areas. Insets (bottom panels) indicate firing rate vectors associated with each state (same color code as in corresponding top panels). (A) Correct trials. (B) Error trials. Note that similar colors in (A) and (B) do not correspond to similar states, since the HMMs were performed independently on correct and error trials (see STAR Methods). D1, first delay (after S1); S2, second stimulus; D2, second delay; GO, appearance of targets on screen; RT, reaction time.

Figure 3.. Comparison of HMM state sequences for original and shuffled datasets

(A–C) Comparison of HMM state sequences for original and shuffled datasets in one representative session (only trials with S2 in the UP position are indicated; similar plots are obtained in other conditions). Trials were grouped according to the relative distance based on stimulus features (blue circle versus red square), and the separation between the two groups is indicated by the black horizontal line with borders marked by triangles (with red trials above the line). The time in each trial has been warped (stretched or shrunk; see STAR Methods) so as to align the 4 events S2, D2, GO, and RT. (A) HMM model of original data reveals (1) the presence of a coding state for relative distance based on stimulus features (the yellow state appearing between S2 and D2 only in trials above the group separation line, coding for “red square farther”) and (2) reliable state transitions at relevant event times. (B) HMM model of circularly shuffled data indicated in (A). As a consequence, state sequences appear scrambled, and the coding states are lost. (C) HMM model of swap-shuffled data indicated in (A). Compared to (A), sequences are not orderly, despite the presence of fewer states. (D) Boxplots of the difference in BIC score between the fits to the original data and shuffled data across sessions (p < 0.001, Wilcoxon signed-rank test). A smaller score indicates better fit. (E) Optimal number of inferred states across sessions for original data (left), circular-shuffled data (middle), and swap-shuffled data (right). Note that similar color in panels a-c does not imply the same state.

Figure 4.. Examples of sessions with significant coding states

(A) Coding states for relative distance based on stimulus features (blue circle versus red square) during S2 (red box) in correct trials for 2 example sessions. Coding states are the dark green and yellow states in the left panel and the dark green and gray states in the right panel. (B) Coding states for relative distance based on order of presentation during S2 (S2 farther versus S2 closer) in correct trials for 2 example sessions. Coding states are the dark green, orange, and gray states in the left panel and the yellow state in the right panel. In both panels, trials were grouped according to the coded variable (as in Figure 3A) and highlighted by the red box. The same colors in different panels do not imply the same state.

Figure 5.. Comparison of RTs and mean state durations between correct and incorrect trials

The mean values ± SEM are reported. ***p < 0.001, two-sided Mann-Whitney U test. Correct trials are indicated in green; incorrect trials are indicated in red. (A) Reaction times (RTs). (B) Mean state durations. (C) Comparison of mean state durations in the S2-GO versus the GO-END intervals, divided into correct and error trials. The plot indicates the interaction plot of the two-way ANOVA, with factors Trial Type (p <10⁻¹², F(1) = 50.9) and Temporal Window (p = 0, F(1) = 363.6 p(interaction) <10⁻⁵, F(1) = 16.3). (D) Left: distributions of mean state durations across sessions in the S2-GO period in correct and error trials. Right: distributions of state durations in 4 example sessions (green, correct trials; red, error trials). (E) Same as in (D) for the GO-END period.

Figure 6.. Analysis of first transition time after S2 (fTaS2)

(A) Example trial with rasters and HMM segmentation of a neural ensemble of 6 cells recorded simultaneously for one difficult trial, with |S2–S1| = 16. Each raster plot indicates the spiking activity of each recorded neuron from 400 ms before S2 until the end of the stimulus presentation. Same conventions as in Figure 1. The triangle on the horizontal axis marks the fTaS2. Vertical dashed line indicates S2 onset. (B) Same trial as in (A), analyzed with an HMM around the GO signal. A time window of 1,400 ms was used: 400 ms before and 1,000 ms after the GO cue. (C) Same as in (A), for an example of an easy trial with |S2–S1| = 40. (D) Same as in (C), with HMM analysis around the GO signal. (E) fTaS2 versus |S2–S1| plot (mean ± SEM) indicates a significant trend of fTaS2 with trial difficulty (p = 0.006, Jonckheere’s trend test). (F) First transition time (means ± SEM) after the GO signal versus |S2–S1|. No significant trend test, p = 0.06. (G) RTs (means ± SEM) versus trial difficulty, |S2–S1|. No significant trend test, p = 0.10.