Temporal progression along discrete coding states during decision-making in the mouse gustatory cortex

Abstract

The mouse gustatory cortex (GC) is involved in taste-guided decision-making in addition to sensory processing. Rodent GC exhibits metastable neural dynamics during ongoing and stimulus-evoked activity, but how these dynamics evolve in the context of a taste-based decision-making task remains unclear. Here we employ analytical and modeling approaches to i) extract metastable dynamics in ensemble spiking activity recorded from the GC of mice performing a perceptual decision-making task; ii) investigate the computational mechanisms underlying GC metastability in this task; and iii) establish a relationship between GC dynamics and behavioral performance. Our results show that activity in GC during perceptual decision-making is metastable and that this metastability may serve as a substrate for sequentially encoding sensory, abstract cue, and decision information over time. Perturbations of the model's metastable dynamics indicate that boosting inhibition in different coding epochs differentially impacts network performance, explaining a counterintuitive effect of GC optogenetic silencing on mouse behavior.

Fig 1. Decision-making activity in GC is metastable and distinct metastable states encode different task-relevant variables.

A: Perceptual decision-making task schematic [12]. S: sucrose, Q: quinine, M: maltose, and O: sucrose octaacetate. B-D (left): Raster plots and HMM-decoded metastable states for in vivo spiking activity during decision-making behavior. “Taste” indicates the animal’s first lick to the central spout; “Decision” indicates the animal’s first lick to a lateral spout. B-D (right): Classification of hidden states. Quality-coding (B, red), Cue-coding (C, cyan), and Action-coding (D, blue) states. * indicates significant difference by Chi-squared test (p < 0.05, corrected for the number of decoded states in the session). Mouse graphic in A created with BioRender.com.

Fig 2. Coding state onset times exhibit an intuitive temporal sequence.

A: Distribution of onset times of all states classified as either Quality- (red), Cue- (cyan), or Action-coding (blue) for experimental data showing a progression from coding for sensory information to abstract cue information to action information. B: Distribution of onset times for spiking data that are shuffled circularly. C: Distribution of onset times for spiking data that are swapped across time.

Fig 3. Clustered spiking network model for decision-making activity.

A: General structure of the network. Neurons in the network are organized into fourteen excitatory (black) and partner inhibitory (red) clusters and connected in a highly recurrent manner. The zoomed inset indicates the generic structure of all of the connections that exist between any two excitatory clusters and their inhibitory cluster partners (pointed arrowheads are excitatory connections, flat arrowheads are inhibitory connections). B: Details of synaptic connections between taste, cue and action clusters, with inputs simulating the sensory stimulus (left, top) and the gating GO signal (left, bottom). C: Raster plot of metastable network activity in the absence of a stimulus (top) and after stimulation (bottom). Neurons are grouped by clusters (black: excitatory, red: inhibitory). S: sucrose, M: maltose, Q: quinine, O: sucrose octaacetate, CL: cue left, CR: cue right, AL: action left, and AR: action right.

Fig 4. HMM analysis of network-simulated data.

A: Example network activity showing sucrose-triggered response. B: Raster plot with overlaid HMM-decoded hidden states corresponding to the trial shown in A. Arrows indicate onsets of different coding states. C-E (left): Raster plots and HMM-decoded metastable states for simulated spiking. C-E (right): Classification of hidden states. Quality-coding (B, red), Cue-coding (C, cyan), and Action-coding (D, blue) states. * indicates significant difference by Chi-squared test (p < 0.05, corrected for the number of decoded states in the session).

Fig 5. Coding state onset times in the simulated data qualitatively reflect those in experimental data.

A: Distribution of onset times of all states classified as either Quality- (red), Cue- (cyan), or Action-coding (blue) for simulated data showing a progression from coding for sensory information to abstract cue information to action information. B: Distribution of onset times for spiking data that are shuffled circularly. C: Distribution of onset times for spiking data that are swapped across time.

Fig 6. Effect of simulated optogenetic silencing on model performance.

A: Rasters showing representative network activity when silencing is applied during “sampling” (0 to 0.5 s) and during “delay” (0.5 to 3 s). Shaded yellow regions indicate that silencing (100% increase in external baseline current for all inhibitory neurons) is on. B: Stimulus time courses and periods of silencing for 10 networks receiving full stimulus input (top) and their corresponding distributions of task accuracies (bottom). C: Distributions of onset times of coding states for 10 networks with stimulus input with gain of 200% (compare to Fig 5A, wherein the stimulus gain is 60%). D: Stimulus time courses and periods of silencing for 10 networks receiving partial stimulus input (top) and their corresponding distributions of task accuracies (bottom). Partial stimulus input was divided into Head only (left panel) and Tail only (right panel). Grey bars represent means ±1 standard deviation for the corresponding None (baseline) condition; N.S. represents no significant difference with respect to the corresponding None; * represents significant difference with respect to the corresponding None (p < 0.05, Bonferroni-corrected post-hoc for 2-way within-subjects ANOVA with factors Stimulus and Silencing).

Fig 7. Effect of simulated optogenetic silencing on network activity.

A: HMM-decoded states for a network subjected to various simulated optogenetic perturbation conditions. Silencing time courses for each condition are shown at the top of each column; decoded states over all 100 trials for each condition are shown below. Trials are ordered in stimulus blocks from bottom to top (S: sucrose, M: maltose, Q: quinine, O: octaacetate) and are ordered in outcome blocks within each stimulus block from bottom to top (correct, incorrect, omitted; incorrect and omitted trials are indicated by black and grey shading, respectively, on the left-hand side). B: Average state sequence similarities over time. For each silencing condition in A, the corresponding curve represents how similar the state sequence is (on average) to the state sequence obtained under control (i.e., under no silencing). Each raw similarity score (ranging from 0 to 1) was normalized by the score obtained by comparing the control condition with itself (see Materials and methods for details, Eq 5). C: Numbers of trials out of 100 for each condition (W: Weak, S: Strong, B: Beginning, C: Cue onset, M: Middle) that contain correct, incorrect, or no Action-coding states, based on a coding state classification of the states from A. Action-coding states are considered correct or incorrect in each trial based on whether their direction preference (in correct trials) matches that trial’s correct direction. * indicates significant difference vs. control in fraction of trials with correct Action-coding states (Chi-squared test across conditions, followed by pairwise Marascuilo post-hoc with α = 0.05).