Abstract
The hippocampus is crucial for retrieval of contextual memories. The activation of a subpopulation of neurons in the dorsal CA1 (dCA1) of the hippocampus is required for memory retrieval. Given that hippocampal neurons exhibit distinct patterns of response during memory retrieval, the activity patterns of individual neurons or ensembles may be critically involved in memory retrieval. However, this relation has been unclear. To investigate this question, we used an in vivo microendoscope calcium imaging technique to optically record neuronal activity in the dCA1 of male and female mice. We observed that a portion of dCA1 neurons increased their responses to the learned context after contextual fear conditioning (FC), resulting in overall increase in response of neuronal population compared with simple context exposure. Such increased response was specific to the conditioned context as it disappeared in neutral context. The magnitude of increase in neuronal responses by FC was proportional to memory strength during retrieval. The increases in activity preferentially occurred during the putative sharp wave ripple events and were not simply because of animal's movement and immobility. At the ensemble level, synchronous cell activity patterns were associated with memory retrieval. Accordingly, when such patterns were more similar between conditioned and neutral context, animals displayed proportionally more similar level of freezing. Together, these results indicate that increase in responses of individual neurons and synchronous cell activity patterns in the dCA1 neuronal network are critically involved in representing a contextual memory recall.
SIGNIFICANCE STATEMENT Neurons in the dorsal CA1 of the hippocampus are crucial for memory retrieval. By using in vivo calcium imaging methods for recording neuronal activity, we demonstrate that dCA1 neurons increased their responses to the learned context specifically by FC and such changes correlated with memory strength during retrieval. Moreover, distinct synchronous cell activity patterns were formed by FC and involved in representing contextual memory retrieval. These findings reveal dynamic activity features of dCA1 neurons that are involved in contextual memory retrieval.
Introduction
Retrieval is the use of learned information represented in complex brain circuits by sensory or internal cues (Kandel and Schwartz, 1982). It has been considered that the retrieval process involves selection, reactivation, or reconstruction of target representation to express appropriate outcomes (Tulving, 1983; Dudai, 2002). Retrieval is one of the most important concepts in memory studies because it is the only way to prove the existence of memory (Frankland et al., 2019). It is well established that the hippocampus is crucial for the formation and retrieval of contextual memories (Squire, 1986; Goshen et al., 2011; Liu et al., 2012; Denny et al., 2014; Tanaka et al., 2014; Cai et al., 2016). Prior studies in the dorsal hippocampal CA1 of mice have shown that reactivation of a sparse subset of neurons active during learning is required for later memory retrieval (Liu et al., 2012; Tanaka et al., 2014).
In vivo imaging of calcium responses of neurons with miniature microscopes has been widely used in freely behaving animals to link neuronal activity to behavior (Ghosh et al., 2011; Kim and Schnitzer, 2022). By using this optical imaging technique and context-dependent learning tasks in mice, prior studies have reported features of hippocampal neuronal activity patterns observed in the learned context. More neurons become active in the dCA1 during recall (Roy et al., 2017) and neuronal responses to the learned context become more specific in the CA1 and CA3 after learning (Rajasethupathy et al., 2015; Roy et al., 2017; Goto et al., 2021). In addition to the firing rate of individual neurons, it has been also shown that a pattern of correlated pairs between neurons specifically change by learning, resulting in emergence of distinct correlated cell ensembles during recall (Modi et al., 2014; Ghandour et al., 2019; Gonzalez et al., 2019; Jimenez et al., 2020). The correlated ensembles created by FC in the ventral CA1 have been suggested as a mechanism for retrieval of contextual fear memory (Jimenez et al., 2020). Electrophysiological recordings and calcium imaging studies have shown that many place cells from the CA1 region remap their firing locations in response to contextual fear conditioning (O'Keefe and Dostrovsky, 1971; Wilson and McNaughton, 1993; Moita et al., 2004; Wang et al., 2012, 2015; Schuette et al., 2020). Such remapping has been considered as a mechanism for encoding environments of a fearful experience by generating new representations of a single environment. Moreover, it has been shown that reactivation of place cell sequences occurs during fear memory retrieval (Wu et al., 2017).
Despite these findings, however, the activity patterns involved in retrieval of memories in dCA1 remains unclear. It could be partly because of the limitations of optical imaging techniques to detect individual spikes, especially in neurons with low firing rate (Dana et al., 2019). Indeed, the majority of dCA1 neurons have sparse activity dynamics. Over 70% of pyramidal neurons display firing rates <1 Hz (Mizuseki and Buzsaki, 2013). Therefore, to better detect individual spikes from dCA1 neurons, we here used a recently developed calcium indicator jGCaMP7f providing a greater spike detection capability (Dana et al., 2019) and an advanced cell sorting algorithm extended constrained non-negative matrix factorization (CNMF_E), which is suitable for extracting single-cell calcium transients in densely packed neuronal population, such as hippocampus (Zhou et al., 2018; Tran et al., 2020). In this study, by taking advantage of these tools, we attempted to identify activity patterns of dCA1 neurons that are correlated with behavioral performance associated with memory retrieval.
Materials and Methods
Mice
All procedures and protocols were approved by the Korea Advanced Institute of Science and Technology Institutional Animal Care and Use Committee. All experiments were performed in accordance with the guideline of the Korea Advanced Institute of Science and Technology Institutional Animal Care and Use Committee. Mice used in all experiments were 8-week-old C57BL/6J WT (The Jackson Laboratory). Both male and female mice were used. The mice were group-housed (3-5 mice per cage) and maintained in 12 h light/dark cycle at a constant temperature of 22 ± 2°C and 40%-60% humidity. Food and water were available ad libitum throughout the experiment. Mice were group-housed until calcium imaging surgery, after which they were singly housed to prevent damage to the implants.
Virus injection surgery
The mice were anesthetized with pentobarbital (83 mg kg−1 of body weight) by intraperitoneal injection. After fully anesthetized, mice were mounted and fixed on the stereotaxic frame (Stoelting). A small hole was drilled with an electrical driller at the target CA1 pyramidal layer site on the right hemispheres of the brain (AP −2.0 mm, ML −1.5 mm, DV −1.5 mm). In all calcium imaging experiments, AAV1-hSyn-jGCaMP7f-WPRE (Addgene) was used for expressing calcium indicator jGCaMP7f in hippocampal dCA1. The titer of all viruses was 1.3 × 10–13 vg ml. The virus was loaded in a glass needle filled with water and 3.0 μl of mineral oil at the tip. An appropriate volume of virus (0.7 μl) was injected at a rate of 0.1 μl min−1 at the targeted sites. The glass needle was placed at the injection site for an additional 10 min to allow sufficient diffusion of the virus. After the virus injection, the scalp at the surgical site was sutured. Mice were placed on a heating pad for recovery and returned to their home cages.
In vivo calcium imaging surgery and baseplate installation
The general procedure of calcium imaging was conducted as previously described (Lee and Han, 2020). GRIN lens (Inscopix, 1.0 mm in diameter, 4.0 mm in length) implantation surgery was conducted 2 weeks after virus injection surgery. Mice were anesthetized and fixed on the stereotaxic frames as mentioned above. The GRIN lens was implanted with additional craniotomy. The drilled hole used for the virus injection was widened up to the 1 mm diameter. A cylindrical column of tissue from the corpus callosum above alveus to the surface of the neocortex was aspirated gently with saline using a 25-gauge blunt-end needle. The GRIN lens was carefully inserted to 0.2 mm above the virus injection site. Resin cement (Sun Medical, Super Bond C&B) was applied to cover the entire exposed skull; and subsequently, acrylic cement (Lang, Ortho-jet Acrylic Pink) was applied additionally to fix the implanted GRIN lens. The surface of the GRIN lens was protected by epoxy bond until baseplate installation. Dexamethasone (0.2 mg/kg of body weight, Sigma) and carprofen (5 mg/kg of body weight, Sigma) solutions were injected into the peritoneal cavity of the mice. Amoxicillin (0.3 mg/ml in water, Sigma) was delivered by ad libitum water. Mice were placed on a heating pad for recovery and singly housed until experiments.
More than 2 weeks after GRIN lens implantation, magnetic baseplate (Inscopix) installation was conducted to provide a stable and easily manipulatable substructure for a miniature microscope during behavior test. The mice were anesthetized and fixed on the stereotaxic frames as before. The baseplate was attached to the miniature microscope (Inscopix), and the microscope was vertically aligned with the implanted GRIN lens and slowly lowered until blood vessels were clearly observed in microscopic view. Acrylic cement was applied to fix the baseplate on the head of mice. After recovery on a heating pad, mice were returned to their home cages.
Behavior procedures
All mice used for behavior experiments were habituated to the experimenter by handling before imaging. During the 10 d of habituation periods, a dummy microscope (Inscopix) with the same size and weight as an actual microscope was connected to the head of mice. At the start of experiments, the dummy microscope was replaced with an actual microscope. To retain a stable and constant field of view (FOV), the miniature microscope remained attached to the head of mice until the end of all experiments. On day 1, mice were exposed to a training context, which was a metallic square chamber (Coulbourn Instrument) with an electric grid floor (17.8 × 17.8 × 30.5; length × width × height in cm); 70% ethanol was used as background odor. In the context fear conditioning group, mice received an electric foot-shock (2s, 0.6 mA) 5 min after mice entered the chamber. The mice remained in the training context for an additional 28 s and return to their home cages. On day 2, mice were reexposed to the training context for 5 min to test context fear memory retrieval. In the auditory fear conditioning group, mice received an electric foot-shock (2 s, 0.6 mA) after 30 s tone (89 dB, 2.8 kHz), which was presented 5 min after mice entered the chamber. On day 3, mice were exposed to a novel context for context shift, which was a white semicircle acryl chamber (Med Associates) with a flat white acryl floor. The mice were returned to the home cage 5 min after the context exposure. In the auditory fear conditioning group, mice received a conditioned tone for additional 2 min to test auditory fear memory retrieval after 5 min of context exposure and were returned to the home cage. The amount of fear memory retrieval was quantified by measuring the time duration mice spent for freezing. Freezing behavior was manually scored by an experienced experimenter. In manual scoring, the time of immobility of the mice without any movement except breathing was counted as freezing. All behaviors were video recorded with FreezeFrame software (Coulbourn Instrument) and exported in .mov format for further analysis.
In vivo calcium imaging in freely moving mice
Fluorescence change of jGCaMP7f in dCA1 neurons was recorded by a miniature microscope in freely moving mice. Calcium transients were captured with nVista 3.0 acquisition software (Inscopix) at 20 frames/s. LED (475 nm) power was 0.1 mw/mm–2. During context fear conditioning, calcium imaging was designed to synchronize with behavior with the TTL-triggering system. A TTL pulse from FreezeFrame software was received by nVista DAQ boxes to commence the imaging session. On each day of behavior test, the first 1 min of context exposure was used for in vivo calcium imaging. In the auditory fear conditioning group, additional imaging was conducted at the first 1 min of conditioned tone presentation on day 3.
Calcium imaging data processing and cell sorting
Recorded calcium imaging video was processed by using Inscopix Data Processing Software (IDPS). In IDPS, the calcium imaging video of each 3 sessions (Preshock, Retrieval, Novel context exposure) with 12,000 frames was imported. The videos of each recording session were spatially and temporally downsampled by a factor of 2 for the efficient data processing in our system. Spatial filtration was performed by a bandpass Gaussian filter; 0.005 pixel–1 was used for the low cutoff and 0.5 pixel–1 for the high cutoff. Motion correction was performed based on the mean intensity projection image of each recording session. Motion-corrected videos were exported as a tiff image sequence. For tracking neurons across multiple recording sessions, exported tiff image sequences were concatenated and normalized by ImageJ software. The concatenated video was aligned by TurboReg rigid body algorithm, which utilizes a single reference frame (mean intensity image) to shift XY frames. An aligned tiff image sequence was imported to IDPS. Then, extended constrained non-negative matrix factorization for micro-endoscopic data (CNMF_E) was used to identify a single neuron component (Zhou et al., 2018). To determine the average diameter of neuron for CNMF_E analysis, principal component analysis-independent component analysis (PCA/ICA) based cell sorting analysis was performed. All principal components detected in PCA/ICA analysis were verified by human visual scrutiny, and average size of detected neurons was applied to CNMF_E. Deconvolution was performed in IDPS (AR1 model) to remove background noise from raw calcium transient trace obtained from CNMF_E.
Calcium event detection
Calcium event was defined as a transient exceeding a 2 SD amplitude from a 0.5 SD baseline, lasting a minimum duration. The minimum duration was calculated by –ln (1/2)/t_half (t_half is half decay time of jGCaMP7f, which is 270 ms) (Dana et al., 2019). Additional calcium transient events within detected calcium transients were detected using findpeaks function in MATLAB (The MathWorks, MinPeakProminence = 1.5 SD, MinPeakDistance = 1 s) (Jimenez et al., 2020). Event frequency was defined as the number of calcium transient events during each session. The activated neurons were defined as neurons with at least a single event through the all sessions.
Mouse speed analysis
The speed of mice was extracted from the exported behavioral video data using customized MATLAB code. The exported .mov files from FreezeFrame software were imported into MATLAB software as an image sequence. Each image frame was transformed into a grayscale image and binarized with an optimal threshold. In our data, the threshold of each image frame was calculated as 2 SDs below from mean of the image. The centroid of mice was acquired with the MATLAB function regionprops, and the speed of mice was calculated as the distance between centroids of adjacent frames. By converting the pixel to the actual distance, we obtained the speed of mice in cm/s.
Identification of putative sharp wave ripple (pSWR)
Calcium events were used to define pSWRs as any frames in which the number of active cells exceeded the mean by >3 standard deviations (SDs) (Davidson et al., 2009; Schuette et al., 2020). For the comparison on event frequency analysis in Figure 4d, the same number of random frames with pSWR events were selected from the frames that mice showed nonfreeze and low speed (<4 cm/s) as a control because SWRs are known to occur when animals show nonfreeze, low locomotive behaviors in hippocampal CA1 recordings (Foster and Wilson, 2006; Carr et al., 2011; Joo and Frank, 2018; Lisman et al., 2018). We repeated 10,000 times of random samplings and calculated mean event frequency for each cell.
Identification of assembly pattern
Cell assembly patterns were detected using an unsupervised statistical framework based on PCA/ICA as described in previous studies (Lopes-dos-Santos et al., 2013; van de Ven et al., 2016; Grosmark et al., 2021). Calcium event traces of each sorted neuron were normalized by z score transform, followed by the construction of correlation matrix for all neuron pairs. Cell assemblies were identified by calculating the principal component of constructed correlation matrix with eigen-values above a threshold. The threshold was derived from an analytical probability function for a random distribution estimated by the Marchenko-Pastur law. The principal component (called assembly activity pattern) is given by a vector of weights representing a contribution of each neuron to the assembly. The neurons with high weight exceeding 2 SDs above mean were determined as assembly components. By omitting the weight of each neuron in the assembly activity patterns, we defined binarized assembly activity pattern as assembly component vector.
Similarity index
The similarity index of two assembly components (or activity patterns) between two different sessions was calculated by a similarity index equal to the average normalized dot products of all possible vector pairs. For the component similarity index, the assembly pattern vectors were binarized. For the shuffled data, event traces of all active neurons were randomly shuffled with time from the original data. Then, assembly components (or activity patterns) were identified from the shuffled event trace data. The shuffled similarity index is defined as an average normalized dot product of all possible pairs between the original assembly components (or activity patterns) of the reference session and the shuffled assembly components (or activity patterns) of a compared session. Forty shuffled datasets were used for calculating mean shuffled similarity index of each animal.
Histology
After behavior experiments, all mice were killed and their brains were perfused and fixed with 4% PFA solution. After 1 d of post fixation in 4% PFA solution, the brain sliced into 40 μm coronal sections using vibratome (Leica Microsystems). Sections were observed under a fluorescence microscope for histologic verification of virus expression and GRIN lens location. Representative images were obtained using LSM880 confocal microscope (Zeiss). After image acquisition, ZEN 3.1 software (Zeiss) was used for optimizing brightness or contrast.
Experimental design and statistical analysis
GraphPad Prism 9 (GraphPad Software) was used for all statistical analyses except Kruskal–Wallis test in Figures 2c, d and 5c, d. We conducted Kruskal–Wallis test with Bonferroni α correction by using kruskalwallis function in MATLAB. The freezing level data and mean event frequency of Ca2+ transient of each groups were analyzed by one-way or two-way ANOVA followed by Bonferroni's or Tukey's multiple comparisons test, as appropriate. The correlation between event frequency change and freezing difference was analyzed by linear regression. The similarity index comparisons with shuffled data were analyzed by paired t test. Specific statistical test performed are indicated in Results and the figure legends.
Code and software
The MATLAB code that used in this study are available from the corresponding author on reasonable request.
Results
In vivo calcium imaging of dCA1 neurons with jGCaMP7f
We used a contextual fear conditioning paradigm with a head-mounted in vivo calcium imaging system to track the activity of dCA1 neurons of freely behaving mice. We sought to identify features of the activity in individual neurons or neuronal ensembles that are specific to recall of contextual fear conditioning (CFC) memory. For imaging during behaviors, we took advantage of jGCaMP7f, a recently developed genetically encoded calcium indicator providing improved detection of individual spikes (Dana et al., 2019). Mice were injected with AAV1-hSyn-jGCaMP7f virus in the pyramidal layer of the right hemisphere of dCA1 and implanted with a gradient refractory index (GRIN) lens above the virus injection site 2 weeks after the virus surgery (Fig. 1a–c). We reliably imaged ∼290 neurons in the FOV on average (Fig. 1d). Mice in FC group were imaged across 3 consecutive days during training (preshock), retrieval, and exposure to neutral context (Fig. 1e). For stable imaging, a miniature microscope was attached chronically throughout the entire course of experiment. To identify neuronal activity patterns specific to FC memory retrieval, we included context reexposure group (CR) as a control where mice did not receive the electric foot-shock during training (Fig. 1e). As expected, mice only in the FC group, not in the CR group, displayed significantly higher freezing during retrieval compared with training and neutral context, confirming successful FC learning (Fig. 1f; two-way repeated-measures of ANOVA, time × groups interaction, F(2,24) = 6.006, p = 0.0077, Bonferroni's multiple comparisons test between sessions, FC preshock retrieval: p = 0.0013, FC retrieval-neutral: p = 0.0134, FC preshock-neutral: p > 0.9999, CR exposure-reexposure: p > 0.9999, CR exposure-neutral: p > 0.9999, CR reexposure-neutral: p > 0.9999). For the extraction of neurons and calcium activity traces from raw imaging datasets, we used the CNMF-E algorithm (Zhou et al., 2018; Tran et al., 2020), and the calcium event was defined as a transient exceeding a 2 SD amplitude from a baseline (Fig. 1g; see Materials and Methods). Neurons that displayed at least one calcium event were defined as an active neuron. Most of the neurons recorded were active in both preshock and retrieval session (Fig. 1h).
In vivo calcium imaging of dCA1 neurons expressing jGCaMP7f in freely behaving mice. a, A miniature microscope with GRIN lens implanted above dCA1 neurons expressing jGCaMP7f by AAV1-hSyn-jGCaMP7f. b, Representative confocal microscopic picture showing jGCaMP7f expression in dCA1 neurons. Scale bar, 500 μm. c, Magnified confocal microscopic picture of jGCaMP7f expression in dCA1 neurons. Scale bar, 50 μm. d, Maximum intensity projection of recorded calcium imaging data on FOV. Scale bar, 10 μm. e, Experimental procedures for CFC and optical recording across 3 d. f, Only mice in the CFC group, not in the context reexposure group, exhibited significantly higher freezing during retrieval session compared with preshock and neutral context session (CFC group [FC]), n = 9 mice and context reexposure group [CR], n = 7 mice, respectively). g, Heat map and trace plot of calcium transients of 40 representative neurons. h, Mean fraction of active neurons in each recording session among total recorded neurons. Each dot indicates data from an individual mouse in fear conditioning group (n = 9 mice). *p < 0.05. **p < 0.01. Data are mean ± SEM.
Increased responses of dCA1 neuronal population by FC correlate with memory strength during retrieval
We identified neurons that were active during at least one recording session and analyzed a change in calcium event frequency in these neurons across days (preshock, retrieval, and neutral context). Overall, the frequency of calcium events increased during retrieval session compared with preshock and neutral context in the FC group, indicating that the increase in firing rate is a retrieval-specific activity feature in dCA1 (Fig. 2a; one-way repeated-measures of ANOVA, F(2,16), p = 0.0056, Tukey's multiple comparisons test, preshock-retrieval: p = 0.0052, preshock-neutral: p = 0.5791, retrieval-neutral: p = 0.0409). This was specific to FC such that there was no significant change in mean calcium event frequency across sessions in the context reexposure group (Fig. 2b; one-way repeated-measures of ANOVA, F(2,12), p = 0.1110, Tukey's multiple comparisons test, exposure-reexposure: p = 0.9996, exposure-neutral: p = 0.3970, reexposure-neutral: p = 0.0507). In the analysis using all active neurons detected from all animals, we found the same pattern of change in the mean calcium event frequency (Fig. 2c; Kruskal–Wallis test with Bonferroni α correction, preshock-retrieval: H(2) = 673.8, p < 0.0001, preshock-neutral: H(2) = 191.8, p = 0.1279, retrieval-neutral: H(2) = 481.9, p < 0.0001 and Fig. 2d; Kruskal–Wallis test with Bonferroni α correction, exposure-reexposure: H(2) = 47.9, p = 0.9996, exposure-neutral: H(2) = 70.3, p = 0.4658, reexposure-neutral: H(2) = 118.1, p = 0.0507). Consistently, cumulative distribution of event frequency of all neurons in FC group was right shifted (Fig. 2e; two-sample Kolmogorov–Smirnov test, D = 0.1354, p < 0.0001). In the context reexposure group, there was no significant change in cumulative distribution (Fig. 2f; two-sample Kolmogorov–Smirnov test, D = 0.0271, p = 0.5604). Based on these observations, we hypothesized that the changes in neuronal firing rate by FC training may be involved in retrieval of contextual fear memory. To test this idea, we examined whether there is any correlation between the magnitude of changes in mean calcium event frequency versus memory strength during recall and found a significant positive correlation, supporting our hypothesis (Fig. 2g; linear regression, F(1,7) = 6.181, p = 0.0418, R2 = 0.4689). Because the increased responses were specific to the learned context, we next questioned whether there is any correlation between changes in response of dCA1 neuronal population and changes in freezing level from retrieval to neutral context. We found no significant correlation in this case (Fig. 2h; linear regression, F(1,7) = 1.988, p = 0.2014, R2 = 0.2211).
Correlation of responses of dCA1 neuronal population with memory retrieval. a, b, Bar graph represents mean event frequency of Ca2+ transients dCA1 neurons across days in CFC group (n = 9 mice) and context reexposure group (n = 7 mice). c, d, Line graph represents mean event frequency of Ca2+ transients dCA1 neurons across days in CFC group (n = 2717 neurons) and context reexposure group (n = 1695 neurons). e, f, Cumulative distribution of event frequency data from c, d (CFC group, n = 2717 neurons and context reexposure group, n = 1695 neurons, respectively). g, Linear regression; mean Δ event frequency (a) versus Δ freezing from retrieval to preshock session (n = 9 mice). h, Linear regression; mean Δ event frequency (a) versus Δ freezing from retrieval to neutral context session (n = 9 mice). *p < 0.05. **p < 0.01. ****p < 0.0001. Data are mean ± SEM.
A portion of dCA1 neurons specifically increases their responses after FC and such changes correlate with retrieval of contextual fear memory
Neurons in the dCA1 are known to exhibit heterogeneous response patterns (Mizuseki et al., 2011; Grosmark and Buzsaki, 2016; Soltesz and Losonczy, 2018). So, despite the overall increase or decrease in event frequency of neuronal populations, each neuron might exhibit different response patterns across sessions. Indeed, the neurons in both FC and CR group could be sorted into one of four possible types based on their response patterns (Fig. 3a,b). Particularly, we were interested in the type of neurons that showed a specific increase in calcium event frequency during retrieval compared with preshock and neutral context because these neurons (named Type 2 cells) were supposed to be highly associated with memory retrieval based on the data in Figure 2. When a mean calcium event frequency was compared between the same subtypes of neurons from FC and control CR group, only Type 2 cells in FC group exhibited significantly higher mean event frequency during retrieval session compared with control group (Fig. 3c–f; two-way repeated-measures of ANOVA, time × groups interaction, Type 1 F(2,28) = 0.2817, p = 0.7566, Type 2 F(2,28) = 5.197, p = 0.0120, Type 3 F(2,28) = 1.450, p = 2516, Type 4 F(2,28) = 0.6237, p = 0.5432, Bonferroni's multiple comparisons test between groups, Type 1 preshock: p = 0.9080, retrieval: p = 0.3235, neutral: p = 0.1792, Type 2 preshock: p = 0.8681, retrieval: p = 0.0003, neutral: p = 0.9056, Type 3 preshock: p > 0.9999, retrieval: p > 0.9999, neutral: p = 0.3150, Type 4 preshock: p > 0.9999, retrieval: p > 0.9999, neutral: p > 0.9999), suggesting that changes in firing rate of Type 2 cells are critically involved in retrieval of contextual fear memory. Indeed, the degree of learning-induced increase in calcium event frequency of Type 2 cells was positively correlated with Δ freezing level from retrieval to preshock (Fig. 3h; linear regression, F(1,7) = 11.14, p = 0.0125, R2 = 0.6140). This correlation was specifically observed only in Type 2 cells (Fig. 2g–j; linear regression, Type 1 F(1,7) = 1.191, p = 0.3112, R2 = 0.1454, Type 3 F(1,7) = 1.723, p = 0.2297, R2 = 0.1983, Type 4 F(1,7) = 0.005, p = 0.9432, R2 = 0.0008). We also found that mean Δ event frequency of Type 2 cells was positively correlated with Δ freezing from retrieval to neutral context (Fig. 3l; linear regression, F(1,7) = 8.297, p = 0.0236, R2 = 0.5424). Again, this correlation was specifically observed only in Type 2 cells (Fig. 3k–n; linear regression, Type 1 F(1,7) = 0.001, p = 0.9792, R2 = 0.0001, Type 3 F(1,7) = 0.352, p = 0.5719, R2 = 0.0478, Type 4 F(1,7) = 3.262, p = 0.1138, R2 = 0.3179). These correlation data suggest that increased responses of a subset of neurons (Type 2 cells) in the dCA1 are critically involved in the retrieval of contextual fear memory.
Selective increase in responses of a portion of dCA1 neurons by FC and its correlation with memory retrieval. a, Heat map represents Ca2+ event frequency of all dCA1 neurons recorded from either CFC (n = 2717 neurons) or reexposure group (n = 1965 neurons). b, Classification of these neurons into four types according to their response patterns across sessions. Venn diagrams represent the fraction of each cell type in CFC (n = 2717 neurons) and reexposure group (n = 1965 neurons). c–f, Mean event frequency changes of Type 1, 2, 3, 4 cells across sessions (FC, n = 9 mice and CR, n = 7 mice, respectively). g–j, Linear regression; mean Δ event frequency of Type 1, 2, 3, 4 cells versus Δ freezing from retrieval to preshock session (n = 9 mice). k–n, Linear regression; mean Δ event frequency of Type 1, 2, 3, 4 cells versus Δ freezing from retrieval to neutral context session (n = 9 mice). *p < 0.05. ***p < 0.001. Data are mean ± SEM.
We next performed an additional data analysis to examine whether the increases in neuronal activity happened preferentially during freezing in the retrieval session. For this purpose, we divided the entire duration of recording session into freezing and ambulation epoch and calculated an average calcium event frequency for each epoch by using all recorded CA1 cells. We found that the average frequency was significantly higher in the ambulation epoch compared with the freezing epoch (Fig. 4a; Wilcoxon matched-pairs signed rank test, Spearman r = 0.2927, p < 0.0001). By using the Type 2 cells only, we also found the same pattern of result (Fig. 4b; Wilcoxon matched-pairs signed rank test, Spearman r = 0.3280, p < 0.0001). These results indicate that the increases in neuronal activity did not occur preferentially during freezing behavior. Therefore, it is unlikely that the activity increase was simply because of the animals' immobility state. Interestingly, a prior study (Schuette et al., 2020) reported a group of CA1 neurons, referred to as freeze cells, that preferentially fires during freezing behavior. Based on our data, it seems that the subset of CA1 cells we identified here whose increased activity correlated with fear memory retrieval is different from those freeze cells.
The increases in activity of dCA1 neurons preferentially occur during pSWR events but not during freezing. a, b, Bar graph represents the event frequency of Ca2+ transients of all recorded (or Type 2) dCA1 neurons in freezing or ambulation epoch (a, n = 2717 neurons; b, n = 1097 neurons, respectively). c, Representative calcium transient traces of neurons (top), the corresponding mouse speed (middle), and the number of active neurons per frame (bottom). Green shadows represent the frames in which mice exhibited freezing. Red shadows represent the frames in which pSWR events occurred. Red dashed line indicates the 3-SD threshold for pSWR identification. d, Bar graph represents the event frequency of Ca2+ transients of Type 2 cells in pSWR event frames or random frames (n = 1097 neurons). ****p < 0.0001. Data are mean ± SEM.
Numerous reports suggested that reactivation of neural representations of past experiences occurs during hippocampal SWRs and such replay is thought to be a mechanism for memory retrieval (Foster and Wilson, 2006; Davidson et al., 2009; Karlsson and Frank, 2009; Wu et al., 2017). So, we asked whether the increases in neuronal activity happened predominantly during SWR events. Since we did not record local field potentials and therefore cannot accurately identify SWR activity, we conducted additional data analysis by introducing the pSWRs (Fig. 4c; see Materials and Methods), which was defined as frames in which the number of active neurons exceeds the mean by 3 SDs as previously published (Davidson et al., 2009; Schuette et al., 2020). Since Type 2 cells were the cells that increased responses during retrieval test, we focused on the Type 2 cells for the analysis. For the comparison, we selected the same number of random frames with pSWR events from the frames that mice showed nonfreeze and low speed (<4 cm/s) as a control because SWRs are known to occur when animals show nonfreeze, low locomotive behaviors in hippocampal CA1 recordings (Foster and Wilson, 2006; Carr et al., 2011; Joo and Frank, 2018; Lisman et al., 2018). We repeated 10,000 times of random samplings and calculated mean event frequency for each cell. We found that the event frequency of Type 2 cells was much higher in the pSWR frames than in the random frames (Fig. 4d; Wilcoxon matched-pairs signed rank test, Spearman r = 0.1177, p < 0.0001). This result suggests that the activity of Type 2 cells was predominantly recruited in pSWR events during the contextual memory recall, consistent with an idea that the increase in responses of Type 2 cells is a mechanism for retrieval of a contextual memory.
Specificity of increased responses of dCA1 neurons
Because animals displayed freezing behavior specifically during retrieval session, the increased responses of dCA1 neurons during retrieval could be related to the immobility state rather than fear memory recall. To examine this possibility, specificity of dCA1 neuronal responses, we conducted an additional experiment where mice were trained for both contextual and auditory fear conditioning and responses of the same neurons were sequentially tracked across contextual and subsequently auditory fear memory recall (Fig. 5a). Animals displayed significant freezing to the conditioned context and tone compared with baseline levels (Fig. 5b; one-way repeated-measures of ANOVA, F(3,12) = 10.21, p = 0.0013, Tukey's multiple comparisons, preshock-context retrieval: p = 0.0059, preshock-neutral: p = 0.6334, preshock-tone retrieval: p = 0.0028, context retrieval-neutral: p = 0.0486, context retrieval-tone retrieval: p = 0.9715, neutral-tone retrieval p = 0.0228). Notably, freezing levels were similar during context and tone tests (Fig. 5b). As above, we again analyzed changes in mean event frequency from entire recorded neurons from all mice that were active during at least one recording session. Consistent with above results, we again observed a significant increase in responses during contextual memory recall compared with the preshock session. If the increase in responses of dCA1 neurons was related to the immobility because of nonspecific freezing behavior, a similar increase was expected during auditory memory recall considering similar levels of freezing during the two tests. However, event frequency did not significantly change by auditory fear memory recall, supporting the specificity of responses of dCA1 neurons to the retrieval of contextual fear memory (Fig. 5c; Kruskal–Wallis test with Bonferroni α correction, preshock-context retrieval: H(3) = 197.4, p = 0.0158, preshock-neutral: H(3) = 39.6, p = 1.000, preshock-tone retrieval: H(3) = 172.2, p = 0.0522, context retrieval-neutral: H(3) = 157.8, p = 0.0973, context retrieval-tone retrieval: H(3) = 25.2, p = 1.0000, neutral-tone retrieval H(3) = 132.7, p = 0.2598). To further confirm the specificity, we analyzed whether responses of dCA1 neurons correlated with animal's movement and found no significant correlation between mean event frequency and the movement speed of mice measured during preshock session (Fig. 5d; one-way ANOVA, F(4,18) = 0.6978, p = 0.6034). Next, we performed the same analyses by using Type 2 cells. Although these neurons increased their responses during both context and tone fear recall, the increase was much greater during context recall compared with tone recall (Fig. 5e; Kruskal–Wallis test with Bonferroni α correction, preshock-context retrieval: H(3) = 905.6, p < 0.0001, preshock-neutral: H(3) = 157.0, p = 0.0003, preshock-tone retrieval: H(3) = 339.0, p < 0.0001, context retrieval-neutral: H(3) = 748.6, p < 0.0001, context retrieval-tone retrieval: H(3) = 566.6, p < 0.0001, neutral-tone retrieval H(3) = 182.0, p < 0.0001). Moreover, there was no significant correlation between the mean event frequency and the animal's movement speed (Fig. 5f; one-way ANOVA, F(4,18) = 1.002, p = 0.4320). Together, these results support the specificity of dCA1 neuronal responses to the retrieval of contextual fear memory.
Specificity of responses of dCA1 neurons to retrieval of contextual fear memory. a, Experimental procedures for CFC, auditory fear conditioning, and optical recording across 3 d. b, The mice exhibited significantly higher freezing during context and auditory retrieval session compared with preshock session and neutral session, respectively (n = 5 mice). c, Line graph represents mean event frequency of Ca2+ transients of dCA1 neurons across four recording sessions (n = 1661 neurons). d, Comparison of Ca2+ event frequency of dCA1 neurons across speeds (n = 5 mice). e, Line graph represents mean event frequency of Ca2+ transients of Type 2 cells across four recording sessions (n = 581 neurons). f, Comparison of Ca2+ event frequency of Type 2 cells across speeds (n = 5 mice). *p < 0.05. **p < 0.01. ***p < 0.001. ****p < 0.0001. Data are mean ± SEM.
Synchronous cell activity patterns in dCA1 are involved in representing memory retrieval
Correlated activity of neuronal population is thought to be involved in memory retrieval. It has been shown that a population of neurons in the ventral CA1 of hippocampus becomes correlated with shock-encoding neurons during memory retrieval, and the degree of synchronized activity within this population is proportional to memory strength (Jimenez et al., 2020). Despite the importance of neurons in dCA1 for memory retrieval (Tanaka et al., 2014; Roy et al., 2017), however, population activity of neurons correlated with memory retrieval has been unclear. So, we analyzed synchronous cell activity patterns in dCA1 in relation to memory retrieval. To this end, activity correlation between neurons was calculated within the dCA1 neuronal population based on Pearson's correlation coefficient (Fig. 6a,b). Cell assembly activity patterns were determined in each session by extracting the assembly components of calcium events (cell component) and calculating the relative contribution of each cell component to the given cell assembly activity (weight) based on the PCA/ICA method from the calculated correlation coefficient as previously described (Lopes-dos-Santos et al., 2013). In PCA analysis, the principal components whose eigen-value exceeded the threshold estimated by the Marcenko-Pastur law were determined as statistically significant assembly patterns (Fig. 6c). Then, we applied the ICA method to extract assembly patterns given by a vector of the weight of each cell component. The cells with a weight exceeding 2 SDs above the mean were regarded as cell components consisting of a given assembly (Fig. 6d). Higher weight means that the cell component participated more in the assembly activity. To compare the similarity of assemblies, a similarity index was determined by calculating an average value of normalized dot products of pairs of assembly components (or patterns). For the component similarity index, the assembly pattern vectors were binarized to the assembly component vectors (Fig. 6e).
Correlation of similarity of cell assembly activity patterns with memory retrieval. a–e, Schematic descriptions for extracting assembly components and patterns from population activity with PCA/ICA method. a, Example calcium event traces of 20 dCA1 cells in the FOV during retrieval session. b, Pair-wise activity correlation matrix of the corresponding 20 cells. The contribution of each cell component to the assembly activity was determined by Pearson's correlation coefficient. c, Histogram represents the number of principle components and their eigen-values. Red dashed line indicates the eigen-value threshold estimated by Marcenko-Pastur law. d, Schematics of the identified assembly patterns with the cell components, each having its own weight vector. e, Schematic description for the similarity index calculation. f–i, Similarity index of cell assembly component and activity pattern were assessed between preshock versus retrieval and between retrieval versus neutral for FC and CR. A chance level similarity (dashed line) was determined by using shuffled dataset (see Materials and Methods; FC, n = 9 mice and CR, n = 7 mice, respectively). j, k, Linear regression; similarity index of cell assembly component and activity pattern (between preshock vs retrieval) versus Δ freezing from retrieval to preshock session (n = 9 mice). l, m, Linear regression; similarity index of cell assembly component and activity patterns (between retrieval vs neutral context) versus Δ freezing from retrieval to neutral context session (n = 9 mice). *p < 0.05. **p < 0.01. ***p < 0.001. Data are mean ± SEM.
We first assessed the similarity of cell component (similarity of assembly component vector) and second the similarity of cell assembly activity patterns by additionally including the weight value of each cell component (similarity of assembly activity pattern vector). We found that the similarity of cell assemblies between preshock and retrieval was significantly higher than chance level in FC but not in the CR group (Fig. 6f,g; paired t test, Fig. 6f: FC (t, df) = (4.583, 8), p = 0.0018, CR (t, df) = (1.152, 6), p = 0.2933, Fig. 6g: FC (t, df) = (6.771, 8), p = 0.0001, CR (t, df) = (0.533, 6), p = 0.6130), indicating stabilization of cell population activity by FC. The similarity of cell assemblies between the conditioned context (day 2) and neutral context (day 3) was at chance level in both FC and CR group (Fig. 6h,i; paired t test, Fig. 6h: FC (t, df) = (1.693, 8), p = 0.1288, CR (t, df) = (0.784, 6), p = 0.4631, Fig. 6i: FC (t, df) = (1.566, 8), p = 0.1559, CR (t, df) = (0.013, 6), p = 0.9897). These data suggest that distinct synchronous activity patterns were produced in dCA1 neuronal population during retrieval of contextual memory. Because cell assembly patterns were similar to each other above chance level between preshock and retrieval, we assumed that the level of such similarity may be correlated with memory strength during retrieval. However, there was no significant correlation (Fig. 6j,k; linear regression, Fig. 6j: F(1,7) = 0.1713, p = 0.6913, R2 = 0.0239, Fig. 6k: F(1,7) = 0.0862, p = 0.7776, R2 = 0.0122). Alternatively, cell assembly activity patterns in dCA1 may be involved in representing a learned context (context associated with a shock). According to this possibility, when cell assembly patterns are more similar between conditioned and neutral context, animals would display proportionally more similar level of freezing. Consistent with this assumption, we found that cell assembly activity pattern similarity (but not cell component similarity) between conditioned and neutral context was negatively correlated with Δ freezing from retrieval to neutral context (Fig. 6l,m; linear regression, Fig. 6l: F(1,7) = 2.414, p = 0.1642, R2 = 0.2565, Fig. 6m: F(1,7) = 6.236, p = 0.0412, R2 = 0.4712). These results suggest that activity patterns of cell assemblies in dCA1 are involved in representing contextual memory retrieval.
Discussion
By using cell-labeling techniques based on activity-dependent induction of immediate early genes, a subpopulation of neurons whose activation is required for later retrieval has been identified as engram cells in the hippocampus (Liu et al., 2012; Tanaka et al., 2014; Denny et al., 2014). In this study, we identified activity patterns of individual neurons and ensembles in the dCA1 of mouse hippocampus that are involved in representation of retrieval of contextual memory. While neuronal activity has been recorded in the dCA1 of rodent brains, the correlation of dCA1 neuronal activity with memory-related behaviors has been unclear. We speculated that this could be because of the limitations of prior experimental approaches used for detecting individual spikes. The jGCaMP7f is a newly developed calcium indicator that originates from the GCaMP6 family (Dana et al., 2019). Among the GCaMP6 family, the GCaMP6f has faster dynamics in fluorescence change but has smaller fluorescence change than the GCaMP6s. Compared with the GCaMP6, the jGCaMP7f shows 1.5-fold faster dynamics than GCaMP6f and 3-fold larger fluorescence change than GCaMP6s. Accordingly, dimmer structure with fast dynamics in the neuronal network can be more efficiently detected with jGCaMP7f. It has been shown that CNMF_E data processing method is better for isolating calcium signals from neurons and provides higher signal-to-noise ratio of the neuronal signals compared with PCA/ICA (Zhou et al., 2018). Moreover, CNMF_E has an advantage for detecting sparse calcium transients in the densely packed neuronal population, such as cortical areas or hippocampus. Consistent with the improvement in detecting individual spikes, we detected generally a greater number of active neurons (∼290 neurons/FOV) in this study compared with previous reports.
We found that not all but a portion of dCA1 neurons (Type 2 cells) exhibited increased responses to the learned context by FC and the magnitude of such increase correlated with freezing behavior during retrieval, which reflects fear memory strength. Moreover, the increases in activity of Type 2 cells were highly associated with pSWR events. Previous work has identified two neural circuits that independently regulate freezing behavior and stress hormone responses to the conditioned cues, CA1 → dorsal subiculum (dSub) → medial entorhinal cortex layer 5 (EC5) and CA1 → dorsal subiculum (dSub) → mammillary bodies, respectively (Roy et al., 2017). The dSub to EC5 circuit is thought to mediate appropriate freezing behavior during memory retrieval. Given this finding, the Type 2 cells could be the neurons directly projecting to the dSub to EC5 circuit and firing rate of these neurons is somehow related to regulating behavioral output of memory retrieval. Alternatively, the Type 2 cells could be involved in regulating the activation of engrams distributed in other brain regions connected with hippocampus CA1 subfield. Recording with projection-based cell labeling would be necessary to test this idea in the future study. In any cases, considering the correlation with freezing behavior during memory retrieval, it is an interesting idea that increase in firing rate of dCA1 neurons by FC encodes representation of memory strength.
According to our data from the calcium event frequency analysis based on two distinct behavioral states (freezing and ambulation) and experiment with auditory fear conditioning, the increased responses of dCA1 neurons during context fear recall were unlikely simply because of immobility. Nevertheless, because these neurons showed a small but still significant increase in response to conditioned tone, we cannot completely rule out the possibility that some neurons responded to the immobility state. Indeed, the dCA1 neuronal population shows various changes in the firing frequency range according to various behavioral states (McNaughton et al., 1983; Yu et al., 2017), and a small population of dCA1 neurons (∼10%) are known to display a highly selective response to immobility (Yu et al., 2017).
It is an interesting question whether the same dCA1 cells that increased their responses during the retrieval of contextual fear memory also respond to the innately threatening environments or locations in a similar fashion (increase of firing rate). According to the functional dissociation model of hippocampus along dorsoventral axis, dorsal hippocampus more likely contributes to cognitive processes, such as learning & memory, spatial navigation, while ventral hippocampus is involved in modulating emotional regulation (Fanselow and Dong, 2010; Strange et al., 2014). Consistent with this view, evidence from lesion, optogenetic, pharmacological, and in vivo calcium imaging studies suggests that ventral, but not dorsal, hippocampus is critically involved in mood or anxiety-like behaviors (Kjelstrup et al., 2002; Pentkowski et al., 2006; Felix-Ortiz et al., 2013; Jimenez et al., 2018). Based on these studies, we speculate that the same increase in activity may not be seen in environments or locations associated with higher anxiety. Future study is required to test this possibility.
Beyond single-cell activity, we also found that synchronous cell activity patterns in cell assemblies correlate with memory retrieval. If these patterns were more similar between learned context and neutral context, animals should display proportionally similar level of freezing in these two contexts. This relationship suggests that synchronous activity pattern in dCA1 may be a signature of neuronal activity by which hippocampus encodes representation of learned context, which is crucial for specificity of memory. In this sense, it would be interesting to examine how these patterns are established by FC and reactivated by contextual cues. The synchronous activity of neurons represents connectivity between neurons (Cossell et al., 2015). Given the role of synaptic plasticity for modulating connections between neurons, it is likely that synaptic plasticity occurring at a specific set of synapses as a result of learning is critical to establish the synchronous cell activity patterns during memory retrieval (Goto et al., 2021; Jeong et al., 2021). Notably, similarity of cell components within cell assemblies was not significantly correlated with memory retrieval, highlighting the importance of cell activity pattern beyond cell composition for representing memory retrieval.
Our study is limited by the lack of causality demonstration. Thus, although a group of neurons exhibited similar pattern of changes after learning, smaller subsets of neurons or even individual neurons within this group could have a distinct role for encoding or retrieval of memory. The future advancements in techniques that enable precise manipulation of activity of individually targeted neurons in behaving animals combined with simultaneous recording of neuronal activity within networks would address the causal role of hippocampal neurons with distinct activity patterns for encoding and retrieval of memories.
Footnotes
This work was supported by Samsung Science and Technology Foundation Project SSTF-BA1801-10. We thank members of our laboratories for helpful discussions and comments.
The authors declare no competing financial interests.
- Correspondence should be addressed to Jin-Hee Han at han.jinhee{at}kaist.ac.kr