Anteromedial thalamus gates the selection and stabilization of long-term memories

Abstract

Memories initially formed in hippocampus gradually stabilize to cortex over weeks-to-months for long-term storage. The mechanistic details of this brain re-organization remain poorly understood. We recorded bulk neural activity in circuits that link hippocampus and cortex as mice performed a memory-guided virtual-reality task over weeks. We identified a prominent and sustained neural correlate of memory in anterior thalamus, whose inhibition substantially disrupted memory consolidation. More strikingly, gain amplification enhanced consolidation of otherwise unconsolidated memories. To gain mechanistic insights, we developed a technology for simultaneous cellular-resolution imaging of hippocampus, thalamus, and cortex throughout consolidation. We found that whereas hippocampus equally encodes multiple memories, the anteromedial thalamus preferentially encodes salient memories, and gradually increases correlations with cortex to facilitate tuning and synchronization of cortical ensembles. We thus identify a thalamo-cortical circuit that gates memory consolidation and propose a mechanism suitable for the selection and stabilization of hippocampal memories into longer-term cortical storage.

Keywords: calcium imaging; consolidation; cortex; dynamics; hippocampus; imaging; memory; optogenetics; thalamus; virtual-reality.

Figure 1.. A virtual-reality based contextual memory task spanning months.

(A) Schematic of virtual reality experimental setup. (B) Timeline of behavioral task from preexposure (day 0), through training (days 1–5), and retrieval (probe trials on days 6, 13, 20, 27, 41, and 55). A single retraining session occurs after each retrieval session. Virtual reality linear track with start, cue, and outcome zones, and example behavioral parameters (speed, lick rate) tracked. (C) Raster of individual trials and plot of trial averages (n=5). Anticipatory lick rate (Hz) is the lick rate prior to the outcome zone. Also shown is lick rate 3 seconds into the outcome zone. (D) Left: quantification of average anticipatory lick rate (Hz) 2 seconds prior to outcome zone entry in each context. Right: quantification of discrimination index (see Methods). **p<0.01, Two way ANOVA with repeated measures, data are mean ± s.e.m. (E) Same as (D), but without weekly retrieval sessions (see also Fig. S6), and testing only recent (day 6) and remote (day 34) retrieval (n=4), *p<0.05, Two way ANOVA with repeated measures, data are mean ± s.e.m. ***p<0.001 for R34, Bonferroni corrected for multiple comparisons.

Figure 2.. Learning and recent memory recall require HPC but remote recall becomes ACC dependent.

(A) stGtACR2-based optogenetic inhibition during training (T1-T5), followed by a test of recent memory. Raw lick traces are shown for mCherry (HPC with no opsin control, n=8) and GtACR (HPC with opsin, n=8) in each context during training and recent retrieval sessions, data are mean (solid line) ± s.e.m (shaded area). (B) Quantification of discrimination between reward and aversive lick rates for mice in (A). ***p<0.001 between pre-exposure and recent for mCherry vs p=0.9780 for HPC GtACR, one-way repeated measures ANOVA with Tukey’s multiple comparison test. Individual data points shown, with mean ± s.e.m. (C) GtACR-based optogenetic inhibition during recent memory in animals that were trained without optogenetic inhibition. Raw lick traces are shown for mCherry (HPC with no opsin control, n=8) and GtACR (HPC with opsin, n=8) in each context during training and recent retrieval sessions. (D) Quantification of discrimination between reward and aversive lick rates for mice in (C). **p<0.01 for between pre-exposure and recent for mCherry vs p=0.5812 for HPC GtACR, one-way repeated measures ANOVA with post-hoc Tukey’s multiple comparison test. (E) DREADDS-based chemogenetic inhibition during remote memory in trained animals. Raw lick traces are shown for mCherry (ACC or HPC with no hM4Di, n=8), HPC hM4Di (HPC with hM4Di, N=8), ACC hM4Di (ACC with hM4Di, n=8) in each context during training and recent retrieval sessions. (F) Quantification of discrimination between reward and aversive lick rates for mice in (E). **p<0.01 for mCherry between preexposure and remote, *p<0.05 for HPC hM4Di, p=0.945 for ACC hM4Di, one-way repeated measure ANOVA with post-hoc Tukey’s multiple comparison test.

Figure 3.. Longitudinal photometry recordings with distributed and mixed coding of task-relevant features.

(A) rgAAV-CAG-tdTomato was injected into ACC, revealing retrogradely labeled neurons in anterior thalamus (ANT), basolateral amygdala (BLA), and entorhinal cortex (ENT) among other regions. Each of ANT (via mammillary bodies), BLA, and ENT receive input from the dorsal hippocampus (HPC). Scale: 1mm. (B) Left: Photometry setup. Excitation light is delivered through fiber optic cables placed above the five implanted cannulas, and emitted fluorescence is projected back onto a CMOS camera. The calcium-dependent fluorescence from GCaMP6f is normalized by the calcium-independent fluorescence to correct for movement artifacts and calculate ΔF/F. Bottom and inset: Example traces of ΔF/F from one mouse on training day 1 from all five regions aligned to task and behavioral variables. (C) Robust trial-averaged responses to reinforcement (airpuff or sucrose) on training day 1 (T1) and 5 (T5) (n=7). Data are mean (dark line) and s.e.m. (shaded area). Photometry scale: x/y: 2s/1z. (D) Left: Schematic of the encoding model used to predict bulk neural activity from task and behavioral variables, with example model outputs from each region shown below (see methods). Right: Percentage of variance explained in the model by each variable throughout training (T1-T5) for each region (n=7). Arrows: One-way ANOVA with Tukey’s multiple comparison test between inter-trial interval (start zone) and cue zone for T3 (****p<0.0001), T4 (***p<0.001), and T5 (*p<0.05). Data are mean ± s.e.m.

Figure 4.. A neural correlate of memory in anterior thalamus persists for weeks.

(A) Top: schematic of VR track with cue zone entry in red. Bottom: Mean dF/F in ANT, BLA and ENT aligned to cue zone entry for both contexts (blue vs orange). Bottom row: animal speed aligned to cue zone entry (n=7). Red arrow signifies AM signals quantified in 4B (see Methods). Data are mean (dark line) with s.e.m. (shaded area). Photometry scale: x/y: 2s/1z. Speed scale: x/y: 2s/0.5z. (B) Quantification of mean change in dF/F at 2s vs. 0s in cue zone, assessed separately for each context., *p<0.05, paired t-test. (C) Same as (A), but aligned to outcome zone entry. Photometry scale: x/y: 2s/1z. Lick rate scale: x/y: 2s/5Hz. Red arrow signifies AM signals quantified in 4D (see Methods). (D) Quantification of the mean difference in dF/F between contexts prior to outcome zone entry (anticipatory) compared to cue zone entry (baseline). *p<0.05, paired t-test.

Figure 5.. Inhibiting or enhancing AM->ACC activity drives bi-directional changes to remote memory retrieval

(A) stGtACR2-based optogenetic inhibition during training (T1-T5), followed by a test of recent (R6) and remote (R27) memory. Light was delivered during cue and outcome periods of the trial. (B,C) Injection strategy for targeting anteromedial thalamus (AM) projections to ACC (AM->ACC) in mCherry control (n=13) and stGtACR2 opsin (n=9) cohorts. Raw lick traces are shown for each mouse in each context on recent and remote retrieval sessions, data are mean (solid line) ± s.e.m (shaded area). (D) Quantification of discrimination between reward and aversive lick rates per mouse on preexposure, recent, and remote retrieval sessions;****p<0.0001 for mCherry between Preexposure and Remote, p>0.05 for GtACR, one-way repeated measures ANOVA with post-hoc Tukey’s multiple comparison test. Individual data points shown, with mean ± s.e.m. (E) stGtACR2-based optogenetic inhibition during training, followed by a test of recent and remote memory in a contextual fear conditioning task. (F) Quantification of freezing behavior in training and remote retrieval sessions in mCherry (AM->ACC no opsin control, n=11) and GtACR (AM->ACC with opsin, n=9). **p<0.01 between mCherry and GtACR cohorts on remote, unpaired t-test. (G) Extension of current behavioral task to include two reward contexts (high reward, HR; low reward, LR; aversive, A). (H) SSFO-based enhancement of neural excitability during training (T1-T5), followed by testing of recent and remote memory. (I-K) Injection strategy for targeting projections to ACC in YFP control (YFP; I) and SSFO cohorts for the retrosplenial cortex (RSP) projections (RSP->ACC; J) or AM projections (AM->ACC; K). Raw lick traces are shown for each mouse in each context on recent and remote retrieval sessions. (L) Quantification of discrimination between high reward (HR) and aversive (A) contexts on preexposure, recent, and remote sessions for each cohort; YFP (no opsin control, n=14); RSP (RSP->ACC SSFO excitation, n=7); AM (AM->ACC SSFO excitation, n=7). *p<0.05 for YFP, RSP and AM cohorts between preexposure and remote, one-way repeated measures ANOVA with post-hoc Tukey’s multiple comparison test. (M) Same as in L, but for discrimination between low reward (LR) and aversive (A) contexts. *p<0.05 only for AM cohort between preexposure and remote, one-way repeated measures ANOVA with post-hoc Tukey’s multiple comparison test.

Figure 6.. Longitudinal multi-region imaging reveals neural dynamics that distinguish consolidated memories.

(A) Multi-region cellular resolution imaging setup. Fluorescence emission from the fiber bundle is collected into a common objective and projected onto a CMOS camera (see methods). (B) Custom-built apparatus for stabilizing and aligning the fiber bundle-to-GRIN lens connection. (C) Neural sources are tracked simultaneously in anterior cingulate cortex (ACC), hippocampus (HPC), and anteromedial thalamus (AM) for up to 34 days. Shown are neural sources from one mouse, same field of view tracked across preexposure (Pre), training day 5 (T5) and remote retrieval (R34). Sources captured on all days are highlighted in green. (D) Individual and mean z-scored dF, recorded for the same cells in ACC (top), AM (middle), and HPC (bottom), aligned to cue zone entry, and tracked from pre-exposure through training and remote retrieval. Dashed lines represent the source from 6C of which the dF traces are tracked. Scale: 1 df/f (z-scored) (E) Mean z-scored dF/F for every recorded cell, aligned to cue-zone entry in high reward (HR; left) and low reward (LR; right) in preexposure and training (n=3 mice). Number of neurons for preexposure and training, respectively: ACC: n=426 and n= 334; AM: n= 94 and n=75; HPC: n=461 and n=457. (F) Tuning to high reward vs low reward contexts (see Methods) for every recorded cell in ACC, HPC, and AM across days. ****p<0.0001 for training (AM), R20 (ACC) and R34 (ACC), one sample t-test. (G) Cumulative distribution of the best match correlation of ACC cells paired with AM cells across days. ***p<0.001, Wilcoxin rank-sum between R6 and R20. (H) AM and ACC correlations on R20 predict tuning in ACC on R34. Top left: cue-zone-aligned mean z-scored dF/F of all AM cells on R20 (n=111 cells). Bottom left: Mean z-scored response of ACC cells which form highly correlated pairs with at least one AM cell (n=56 cells across 3 animals, 60% correlated to one AM cell, 20% correlated to two cells, and 14% correlated to three or more cells). Bottom right: mean response of these same tracked ACC R20 cells on R34. For each mean dF/F trace, representative cells selected from one animal are shown to the left. (I) Distribution of tuning for ACC cells correlated to at least one AM cell on the previous recording session compared to the tuning of all remaining cells. ** p<0.01, Wilcoxin rank-sum test. (J) Top: Representative correlation matrices within ACC for either HR or LR context, taken from a single animal. Bottom: Cumulative distribution of best match correlation of ACC cell pairs across recording sessions for HR (left) and LR (right). ***p<0.001, Wilcoxin rank-sum between R6 and R20. (K) Top: Example from one animal of undirected network graphs of ACC cell correlations. Each node represents an individual ACC cell, and connecting edges represent significant correlations in HR (Pairwise Pearson’s r>0.3). Red nodes are the ACC cells most temporally correlated with AM cells (top 5% of Pearson’s r) shown in red. Bottom: Quantification of the proportion of pairwise correlations (number of edges for a given cell normalized by the total number of cells for a given recording secession) in ACC for either red nodes (ACC cell that is highly correlated to at least one AM cell) and blue cells (all remaining ACC cells), compared to chance (in black, see Methods). Shaded area represents the 95% confidence interval around chance distribution. Significance is indicated by bins above this level.

Figure 7.. AM is required to stabilize the tuning and timing of long-term contextual representations in cortex.

(A) Experimental design for optogenetic inhibition and imaging experiment. rgAAV1-hSyn-Cre and AAV1-CKIi-GCaMP6f was injected into anterior cingulate cortex (ACC) and AAV1-SIO-stGtACR2 was injected into anteromedial thalamus (AM). Imaging was performed on preexposure, recent and remote retrieval sessions, whereas inhibition occurred during each training session. (B) Mean z-scored dF for context-tuned neurons, aligned to cue-zone entry, compared between mCherry (n=5) and GtACR (n=7) mice on preexposure, recent (R6) and remote (R27) retrieval. *p<0.05, two-sample t-test. (C) Top: Examples of undirected network graphs of ACC cell correlations in HR from mCherry (left) and GtACR (right) mice during recent and remote retrieval. Each node corresponds to a cell in ACC and each edge is a highly correlated cell pair (Pairwise Pearson’s r>0.3) Bottom: Cumulative distribution of the best-match correlation for every cell in HR from recent to remote retrieval. **p<0.01, Wilcoxin Rank-Sum test. (D) Proposed model for how AM selects and stabilizes consolidated memory representations in cortex. Selection of information about salient cues by AM leads to coordinated responses in ACC at remote timepoints, eventually allowing for recall that is independent of HPC.