Hippocampal and medial prefrontal ensemble spiking represents episodes and rules in similar task spaces

Abstract

Episodic memory requires the hippocampus and prefrontal cortex to guide decisions by representing events in spatial, temporal, and personal contexts. Both brain regions have been described by cognitive theories that represent events in context as locations in maps or memory spaces. We query whether ensemble spiking in these regions described spatial structures as rats performed memory tasks. From each ensemble, we construct a state-space with each point defined by the coordinated spiking of single and pairs of units in 125-ms bins and investigate how state-space locations discriminate task features. Trajectories through state-spaces correspond with behavioral episodes framed by spatial, temporal, and internal contexts. Both hippocampal and prefrontal ensembles distinguish maze locations, task intervals, and goals by distances between state-space locations, consistent with cognitive mapping and relational memory space theories of episodic memory. Prefrontal modulation of hippocampal activity may guide choices by directing memory representations toward appropriate state-space goal locations.

Keywords: CP: Neuroscience; cognitive maps; decision making; episodic memory; frontotemporal interactions; hippocampus; learning; neuronal representations; prefrontal cortex; relational memory space.

Figure 1.. Schematics of tasks, electrophysiology, and analysis pipeline

(A) Spatial memory and cue-approach learning were tested in a plus-shaped maze. In each trial, a rat was placed in either the north or south start arm, chosen pseudo-randomly. In the spatial memory task, the rat discovered by trial and error which of two goal arms contained food. The same spatial goal was rewarded within a block of 10–12 trials. After reaching criterion performance in an initial discrimination (ID) block, the opposite goal was rewarded in a reversal block (R). Each recording session included one ID and three R (R1–R3) blocks. In the cue approach task, the rat learned that the rewarded arm was indicated by a cue light. Rats completed similar numbers of trials in each task. (B) Either CA1 only (16 ensembles, 976 units total) or CA1 (8 ensembles, 488 units total) and mPFC (8 ensembles, 224 units total) were recorded simultaneously as the animals performed the tasks. (C) Single-unit and pairwise-unit activity recorded in 125-ms intervals was described using a state-space analysis of spike cross-correlations (SSASC). The state-spaces were (1) visualized using dimensionality reduction methods (Figure 2; all ensembles from CA1 and mPFC used, n_CA1 = 24, n_mPFC = 8; single exemplar shown), (2) compared using topological data analysis (Figure 3) (all ensembles from CA1 and mPFC used, n_CA1 = 24, n_mPFC = 8), and (3) used to analyze potential CA1 and mPFC interactions (Figure 4) (from simultaneously recorded ensembles, n_CA1 = 8, n_mPFC = 8).

Figure 2.. Simplified CA1 and mPFC state-spaces were organized by task features

State-space dimensionality and simplified models visualize behavioral correlates of ensemble spiking. (A) Latent dimensions. The fewest dimensions needed to account for 90% of state-space variance was assessed by PCA, MIND, and MFTMA. Lines show the mean variance explained by different numbers of dimensions across 24 CA1 and 8 mPFC ensembles. Non-linear methods (MIND, red lines; MFTMA, blue lines) found four to six latent dimensions captured 90% of the variance; PCA (black lines) found approximately seven. Vertical hash marks (|) above the x axes show the mean of all recorded ensembles, asterisks (*) indicate the mean of ensembles grouped by rat; colors indicate the dimension reduction methods. Group variance was within the size of the asterisks. (B‒E) Examples of simplified state-space from a single representative ensemble recorded in the spatial memory task. Multidimensional (MDS) and t-SNE show paths through simplified state-space models. Each dot shows the location of ensemble activity in one 125-ms window in 3D visualizations. (B–D) MDS embeddings. (B) Paths through the simplified state-space corresponded with trial types, i.e., north-to-east, north-to-west, south-to-east, and south-to-west journeys, and (C) trial epochs, i.e., the waiting platform, start arm, choice point, and goal arm. (D and E) Paths through simplified state-spaces corresponded with temporal sequences. (D) MDS embeddings showed time unfolding within trials as filament-like threads expanding from the lower right. (E) t-SNE showed time within trials and trial sequences from the first point at the top of the plot (black) to the last (bright green) of the recording session. Because simplified models by definition reduce the information available in full datasets, all other analyses quantified features of full state-spaces.

Figure 3.. Similar state-spaces were formed by different ensembles within brain regions

(A‒D) Each dot represents the bottleneck distances between two state-spaces, each constructed from a different neuronal ensemble. Boxplots show the null distribution of bottleneck distances between all 1,366,200 pairs of state-spaces constructed from randomized neural activity. Boxplot whiskers show the 5th and 95th percentiles, box borders correspond to the first and third quartiles, and orange lines show the median of the null distribution. Dot color shows the results of permutation tests of bottleneck distance between the recorded state-space and the null distribution (red, p < 0.05; blue, p R 0.05). (A) Different CA1 and (B) mPFC ensembles recorded in the same task formed similar state-spaces. (C) Individual state-spaces tracked wide-ranging combinations of the behavioral similarities and cognitive differences between the spatial memory and cue approach tasks. (D) Simultaneously recorded CA1 and mPFC ensembles formed topologically distinct state-spaces in each task (permutation testing, p < 0.05; spatial memory: 8/8 ensembles; cue approach: 7/8 ensembles), and (E) task-related changes in CA1 and mPFC state-spaces were uncorrelated (8 pairs of ensembles, R² = 0.11, F_1,6 = 0.72, p = 0.43). (F) Summary statistics of bottleneck distances between pairs of state-spaces within tasks (mean ± SEM). The first two sets of columns correspond with (A and B); the last two sets show the mean differences between CA1 and mPFC state-spaces pooled by ensemble (CA1 vs. mPFC) and by rat (pooled by rat).

Figure 4.. State-space topology differed between tasks

(A) Pie charts show distributions of mean state-space differences between tasks compared across all ensembles and pooled by rat. Similar proportions of state-spaces were topologically more different, more similar, and no different than chance between the spatial memory and cue approach tasks (categories correspond to the blue and red dots in Figure 3C. (B) The mean proportion of single units with spatial remapping within ensembles corresponded with categorical differences between state-spaces. (C) Bottleneck distances between state-spaces covaried with spatial (red diamonds) and rate (black squares) remapping. Symbols show means averaged for individual rats.

Figure 5.. mPFC predicted CA1 state-spaces only during the spatial memory task

(A and B) Individual dots represent the (A) Granger value or (B) average reconstruction error for a single ensemble. Black dashed lines show the upper (A) or the lower (B) bound of the 95% confidence interval of Granger values generated from shuffled spike trains. Each dot shows (A) the Granger value and (B) the reconstruction error of predictions between a pair of CA1 vs. mPFC state-spaces compared with the distribution of predictions between pairs of state-spaces from randomized spike trains (permutation tests: red, p < 0.05; blue, p ≥ 0.05). (A) Trial sequence predictions: mPFC consistently predicted CA1 trajectories in the spatial memory task (permutation testing, p < 0.05, mPFC → CA1; spatial memory: 8/8 ensembles; cue approach: 1/8 ensembles). CA1 did not consistently predict mPFC sequences in either task (permutation testing, p < 0.05, CA1 → mPFC; spatial memory: 1/8 ensembles; cue approach: 2/8 ensembles). (B) Trajectory structure predictions. Single mPFC trajectories predicted the location of corresponding CA1 trajectories in simultaneously recorded ensembles only during the spatial memory task (permutation testing, p < 0.05, mPFC → CA1; spatial memory: 8/8 ensembles; cue approach: 3/8 ensembles). CA1 trajectories rarely predicted the location of corresponding mPFC state-spaces in either task (permutation testing, p < 0.05, CA1 → mPFC; spatial memory: 1/8 ensembles; cue approach: 2/8 ensembles).