Abstract
Summary The dentate gyrus (DG) is crucial for behaviorally discriminating similar spatial memories, predicting that dentate gyrus place cells change (“remap”) spatial tuning (“place fields”) for memory discrimination. This prediction was never tested, although DG place cells remap across similar environments without memory tasks. We confirm this prior finding, then demonstrate that DG place fields do not remap across spatial tasks that require DG-dependent memory discrimination. Instead of remapping, place-discriminating discharge is observed transiently amongst DG place cells, particularly where memory discrimination is most necessary. The DG network signals memory discrimination by expressing distinctive sub-second network patterns of co-firing amongst principal cells at memory discrimination sites. This is accompanied by increased coupling of discharge from excitatory principal cells and inhibitory interneurons. Instead of remapping, these findings identify that memory discrimination is signaled by sub-second patterns of correlated discharge within the dentate network.
eTOC blurb van Dijk and Fenton report that dentate gyrus place cells signal memory discrimination not by remapping, but by variable sub-second patterns of coordinated place cell network discharge and enhanced discharge coupling between excitatory and inhibitory neurons, at sites of memory discrimination.
Highlights
Dentate gyrus-dependent memory discrimination does not require place cell remapping
Dentate neural correlates of pattern discrimination are transient, lasting seconds
Sub-second dentate network discharge correlations signal memory discrimination
Dentate excitatory-inhibitory coupling is increased at memory discrimination sites
Introduction
Hippocampus is crucial for discriminating between memories (Gilbert et al., 1998; Yassa and Stark, 2011), for navigating space (Maguire et al., 1998; Morris et al., 1982; O’Keefe and Nadel, 1978), and the discharge of hippocampus place cells represents locations (O’Keefe, 1976; Wilson and McNaughton, 1993). Although place cell discharge is unreliable in time (Fenton and Muller, 1998), the discharge reliably localizes to specific regions of space called the cell’s place field (O’Keefe, 1976). These place fields relocate when the environment changes sufficiently, a process known as “spatial remapping” (Muller and Kubie, 1987) or alternatively, they maintain their locations but systematically change firing rates in what is called “rate remapping” (Hayman et al., 2003; Leutgeb et al., 2005b). Both forms of remapping have been associated with memory discrimination (Alme et al., 2014; Colgin et al., 2008; Wills et al., 2005), and the dominant cognitive map theory predicts place fields will remap across conditions requiring distinct place memories (O’Keefe and Nadel, 1978).
Theories of hippocampus memory computation assert the DG is specialized for discriminative functions such that DG outputs are more distinctive than the corresponding inputs, what is called “pattern separation” (Marr, 1971; O'Reilly and McClelland, 1994; Treves and Rolls, 1994). As predicted, compromising DG function impairs difficult memory discriminations in modestly distinct environments (Burghardt et al., 2012; Gilbert et al., 2001; Kheirbek et al., 2013; Lee et al., 2005; McHugh et al., 2007; Nakashiba et al., 2012). In addition, DG principal cells (pDGC) consisting of granule cells and mossy cells remap in response to sufficiently large environmental changes (Danielson et al., 2017; GoodSmith et al., 2017; Leutgeb et al., 2007; Neunuebel and Knierim, 2014; Senzai and Buzsaki, 2017), but the extent to which these observations indicate pattern separation and whether they are relevant to memory discrimination is not established.
Indeed, pDGCs place cells have been characterized but not during an explicit memory discrimination task that requires intact DG function (Danielson et al., 2017; GoodSmith et al., 2017; Leutgeb et al., 2007; McHugh et al., 2007; Neunuebel and Knierim, 2014; Senzai and Buzsaki, 2017). To critically test the prediction that DG remapping mediates memory discrimination we recorded DG neurons while mice on a rotating arena perform DG-dependent memory discriminations in active place avoidance tasks (Burghardt et al., 2012; Kheirbek et al., 2013; Park et al., 2015). The rotation dissociates the environment into two spatial reference frames and reveals that distinct frame-specific patterns of ensemble place representations alternate in the ensemble spike time series (Fenton et al., 1998; Kelemen and Fenton, 2010). We find that DG place cells do not spatially remap across sessions that require DG-dependent memory discrimination, although rate remapping increases on memory trials. When the mice are in the vicinity of the avoided place, DG place cells preferentially discharge in the task-relevant spatial frame, indicating transient and distinctive network states. Instead of remapping, the precise locations where memory discrimination is required were signaled by sub-second co-firing of pDGC place cells that was inconsistent with the overall discharge correlations of the cell pair. This network inconsistency was accompanied by increased sub-second discharge correlations amongst DG place cell-interneuron pairs. Instead of the time-averaged changes in spatial tuning associated with remapping, these findings point to dynamic patterns of correlated DG discharge for memory discrimination.
Results
Confirmation that DG place fields are less stable than CA fields across environments
We began by recording cells across three versions of an environment (standard, 90° cue relocation, and wall removal; Fig. 1A, Fig. S1) to confirm that place cells in DG (45 place cells of 133 cells in 5 mice) are more sensitive to environmental changes than those in Ammon’s horn (20 place cells of 42 cells in 3 mice) (Leutgeb et al., 2007). The majority of pDGC (64%) and CA (75%) place cells had a single place field during the initial recording in the standard environment (test of proportions z = 1.6, p = 0.1). The number of place fields scaled with the 2-fold increase in area from the standard to the wall removal condition (DG standard: 1.42±0.10 fields; removed: 2.70±0.29 fields; CA standard 1.26±0.13 fields; removed 2.0+0.28 fields; region: F1,63.7 = 3.65; p = 0.10; condition: F1,63.3=13.4 p=10−4, interaction: F1,63.3 = 1.26 p = 0.27 post-hoc: removed DG > standard DG and CA; removed CA > standard CA), as previously reported for rat place cells (Fenton et al., 2008; Park et al., 2011). Other discharge properties were also similar for the DG and CA principal cell populations (Table S1). In contrast, pDGC firing rate map stability was less than in CA and the stability of both populations was changed by wall removal but not by the cue relocation (Fig. 1B). Rate remapping in pDGCs and CA cells did not differ across the environment manipulations (region: F1,172=0.16, p=0.68; manipulation: F2,172 = 2.8, p = 0.064; region x manipulation: F2,172 = 0.2, p= 0.81), but the magnitude of the firing rate changes in and out of the primary place field differed across manipulations for CA rates but not for DG rates (Fig. 1C).
Confirming a DG-dependent memory discrimination task
Before testing the prediction that place fields change with memory discrimination, we confirmed that the active place avoidance memory discrimination task depends on DG function (Fig. 2A). Laser illumination of POMC-Halorhodopsin mice that optogenetically silences granule cells in the Cre+ but not Cre− littermates (n’s = 8) showed that DG cells are not essential for learning to avoid the initial location of shock, but are essential for avoiding shock on the conflict trial when the shock is relocated 180° (Fig. 2A,B). Because illumination both directly and indirectly silences DG cells (Senzai and Buzsaki, 2017) this demonstrates the DG is specifically necessary for conflict memory discrimination (Burghardt et al., 2012; Kheirbek et al., 2013).
DG-dependent memory discrimination is not accompanied by dentate firing field remapping
We then used a similar protocol (Fig. 2C) to test if pDGC firing rate maps (n = 42 place cells of 113 cells in 6 wild-type mice) change with memory discrimination (Fig. 2D). Place avoidance learning was normal (Fig. 2E) and measures of spatial firing quality did not differ across the pretraining, initial training, and conflict training trials (Table S2). Importantly, firing rate map stability also did not differ across the “replication manipulation” (pretraining 1 vs. pretraining 2, initial training 1 vs. initial training 2), across the “shock addition” manipulation (pretraining vs. initial training) and across the “shock relocation” manipulation (initial training vs. conflict training; Fig. 2F). These estimates of stability were similar to the estimates across the replication manipulation of the environment (Fig. 1B) for pDGC (z = 0.11, p > 0.9) but were weaker compared to CA cells (z = 2.6, p < 0.01). The magnitude of pDGC-specific firing rate changes (H2,117 = 14.4, p<0.001) and place field-specific rate changes were greater in response to the shock-added and shock-relocated manipulations than across the replication manipulations (Fig. 2G). This pattern of replication < shock addition = shock relocated was also observed in the putative inhibitory neurons (H2,57 = 8.97, p=0.011). These findings demonstrate DG firing rates are sensitive to changes in task contingency, consistent with episodic encoding (Leutgeb et al., 2005b). Note however that these single-cell estimates of neural pattern discrimination were not specific to DG-dependent memory discrimination.
Dentate place cell ensembles transiently and purposefully discriminate places in distinct spatial frames
We next examined these DG data by analyzing the conjoint discharge of cells on subsecond time scales. We were motivated because the assumptions of the preceding session-averaged analyses of single-cell firing (Figs. 1,2) contrast with the strong non-stationarity (Carr and Frank, 2012; Fenton et al., 2010; Fenton and Muller, 1998; Ferbinteanu et al., 2011; Gothard et al., 2001 ; Gothard et al., 1996; Gupta et al., 2010; Huxter et al., 2003; Jackson and Redish, 2007; Redish et al., 2000; Shapiro and Ferbinteanu, 2006; Singer et al., 2010) and ensemble place coding properties (Dupret et al., 2010; Harris et al., 2003; O'Neill et al., 2008; Park et al., 2011; Pastalkova et al., 2008; Pfeiffer and Foster, 2015; Wikenheiser and Redish, 2015; Wilson and McNaughton, 1993) that are reported for hippocampal spiking dynamics. We thus considered a different form of neural pattern discrimination that can be measured on a moment-to-moment basis. The rotation of the arena dissociates the environment into two distinct spatial frames. One is stationary, defined by room-anchored cues and the other is rotating, defined by local arena-anchored cues (Fenton et al., 1998). To solve this task the mouse has to discriminate between room-based spatial information and arena-based spatial information and this is essential in the vicinity of the shock zone if the animal is to successfully avoid the shock that is defined only by room-frame information, and not by arena-frame information. Spatial frame-specific positional discharge was estimated by separately computing momentary positional information (Ipos) in each spatial frame for the place cell ensemble (Kelemen and Fenton, 2010; Olypher et al., 2003), and computing ΔIpos, the difference between room-frame and arena-frame Iposeach 133 ms (Fig. 3A). Like place cell firing rate fluctuations (Fenton et al., 2010), the correlation between ΔIpos and running speed explains little of the variance (r2=0.69%, p < 0.001). Room information dominates arena information more often than vice versa during pretraining but this preference disappears when shock is added or relocated (Fig 3B) even though room information must be used to avoid shock (Bures et al., 1997; Fenton and Bures, 2003). During the ~1 s before mice avoid or enter the initial location of shock (Fig. 3C) ΔIposis positive (i.e. room-preferring) and larger than when the mice are in the corresponding area 180° away in the potentially safest zone (Fig. 3C). The prevalence of room-preferring information followed the shock zone relocation on the conflict trial (Fig. 3C), even though time-averaged firing rate maps did not change (Fig. 2). We observed a similar preference for room-specific discharge in the vicinity of the current shock zone when we examined the locations of frame-specific Ipos values (Fig. 3D), instead of their differences. These findings demonstrate discrimination of neural representations of location in DG discharge; the discrimination is both dynamic and purposeful insofar that it corresponds to the need for discriminative place avoidance memory. Notably, place discrimination into room and arena-defined places was insensitive to optogenetic silencing of DG cells (Fig. 2B).
Variable consistency of discharge coupling within the dentate network, specifically at sites of memory discrimination
Next, we investigated whether the sub-second coordination of DG spike trains changes across the trials requiring memory discrimination. The distributions of spike train correlations (τ) between pairs of pDGC place cells measured at the 133-ms time scale of the Ipos estimates do not differ in the three trial types (H2,381 = 1.57, p = 0.46, Fig. S2.). Nor do they change systematically from one manipulation to another (Fig. 4A; H2,381 = 0.87, p = 0.65, Fig. S2). Standardized firing rates (z) were computed during 5-s intervals that identified passes through firing fields (Fenton et al., 2010; Fenton and Muller, 1998; Jackson and Redish, 2007). The distributions of standardized rates did not differ across the pretraining, training, and conflict sessions (F2,9149 = 0.31; p=0.73). Overdispersion of the standardized rates was not different in the pretraining sessions (var = 5.2) compared to values reported for area CA1 in rats. Overdispersion during pretraining was also indistinguishable from during training (var = 5.74; F4189,3169 = 1.10; p = 0.17), and conflict (var = 5.63; F1791,3169 = 1.08; p = 0.63) sessions (also when the unit of analysis was a single cell with at least 15 passes, F2,178 = 0.44; p=0.64). Similar to time-averaged measures of spatial firing (Fig. 2), single cell estimates of short time-scale spiking dynamics did not differ across the trials (Fenton et al., 2010; Fenton and Muller, 1998; Jackson and Redish, 2007).
We then considered the short time-scale spiking dynamics of place cell pairs. The momentary discharge of place cells with overlapping firing fields can characteristically covary positively so that when the mouse passes through the firing fields the two cells both fire more or less than predicted by their firing fields, or the cell pairs can covary negatively so when one fires excessively the other does not, or they can discharge independently so that the firing variations of one cell do not predict the other cell’s momentary firing (Fenton, 2015; Kelemen and Fenton, 2012). To evaluate the dynamic interactions amongst the network of spiking cells, we then asked whether the momentary firing rate variations of pairs of place cells as they crossed overlapping firing fields (z1 and z2) were consistent with the overall short-time scale discharge coupling between the cell pair (τ1,2; see schematic in Fig. 4B top). Because τ1,2 did not vary across conditions, we computed network consistency (NC = z1 · z2 · τ1,2), which is maximal when the momentary firing rates of two cells with overlapping firing fields covary similar to the overall spike train correlation of the cell pair; network consistency is minimal when the rate covariance is opposite to the spike train correlation. Although overall network consistency was positive and similar across the pretraining, training, and conflict sessions (F2,2523 = 13; p = 0.27), the requirement for memory discrimination predicts greater pattern distinctiveness and thus lower network consistency, particularly in the vicinity of the shock zone when memory discrimination is most required. During pretraining, before shock was ever experienced, network consistency was similar when the mouse was in the vicinity of the future shock zone and the opposite corresponding region of the arena (Fig. 4B bottom). During training trials, network consistency increased opposite the shock zone and network consistency in the same location decreased when the shock was relocated 180° in the conflict trial. While the increased network consistency opposite the initial shock zone was not predicted, the reduced network consistency in this location when shock is relocated is predicted by the greater demand for memory discrimination in the relocated shock zone vicinity, on the hypothesis that distinct network states manifest for distinct memories. Accordingly, momentary discharge amongst cell pairs will be inconsistent with the overall discharge coupling between network cell pairs (Fig. 4B bottom). This view can also explain increased network consistency opposite the initial shock zone if the mice consistently considered it safe and distant from shock. This pattern of change in network consistency can not simply be due to the mice spending more or less time in the shock zones because network consistency was unchanged across the session types in the vicinity of the initial location of shock, whether or not it was a neutral, avoided, of preferred region of the environment. We conclude that unlike session-averaged features of place cell discharge like the location of place fields, the sub-second discharge coordination within the network of pDGCs systematically varies across time and locations and these distinctive network states are associated with increased demand for memory discrimination and may be sensitive to rapid changes with learning experience (Bittner et al., 2015; Cheng and Frank, 2008).
Increased place cell-inhibitory neuron coactivity at sites of memory discrimination
To explore how distinctive network states might transiently arise, we investigated whether the task manipulations alter discharge correlations between place cells and inhibitory neurons. Inhibitory networks may be important for memory discrimination (Buzsaki, 2010; Danielson et al., 2017; Jinde et al., 2013; Park et al., 2015), and network consistency, perhaps in part because high levels of inhibition promote pattern separation in mature granule cells (Marin-Burgin et al., 2012). The distributions of place cell – interneuron discharge correlations did not differ across trials (Fig 4C). In contrast, the correlated firing of the individual cell pairs differed across the three types of manipulation; the changes were higher across the shock-relocation manipulation (Fig. 4C top left) that requires memory discrimination and relies on intact DG function (Fig. 2). This increase in place cell – interneuron co-firing was observed generally, most place cells increased co-firing with at least one interneuron (83% of place cells for the shock-addition (z = 1.15, p = 0.2) and 92% for the shock-change (z = 2.3, p = 0.03) comparisons relative to 75% of place cells for the replication comparison. These findings also hold for 20 ms and 250 ms timescales (Fig. S2).
Because changes in correlated firing may not be homogeneous within a network of neurons (Harris, 2005; Okun et al., 2015; Olypher et al., 2006), the pairs were subclassified for further analysis. During pretraining, cell pairs were significantly positively (81 pairs, 53%), significantly negatively (15 pairs; 10%), or not (56 pairs, 37%) correlated according to statistical criteria. After shock-relocation, the correlations amongst the initially independent place cell-interneuron pairs increased (Fig. 4C top middle) whereas positively correlated pairs did not change, although they decreased with the replication and shock addition manipulations (Fig. 4C top right). Changes were not observed amongst the initially negatively correlated pairs (H2,73 =4.28, p = 0.12). These changes in correlated firing were specific to the vicinity of the shock zone, where memory discrimination is most required. The correlation increases are larger with the shock change when the mouse is close to the currently to-be-avoided zone than when it is close to the previous shock zone location (Fig 4C bottom). These location-specific differences were not detected for the other manipulations. In accord with these changes, during the conflict trial, we observed higher correlations near the current shock zone compared to its prior location (Fig 4C bottom). These differences between the locations were not observed when corresponding random samples of intervals were compared, nor were the location-specific differences observed in the pretraining or training trials. Together, these data indicate that coupling increases between the firing of DG place cells and inhibitory neurons, specifically when the animal needs to perform DG-dependent memory discrimination to avoid shock.
Discussion
We evaluated hypotheses of how the dentate gyrus contributes to memory discrimination (Colgin et al., 2008; Leutgeb et al., 2007; Neunuebel and Knierim, 2014; Wills et al., 2005) by investigating the changes in pDGC place cell firing across different behavioral episodes that included a test demonstrated to require DG-dependent memory discrimination (Fig. 2A; Burghardt et al., 2012; Kheirbek et al., 2013). Contrary to expectations that time-averaged spatial discharge patterns (firing field arrangements) correspond to distinct spatial memories (Leutgeb et al., 2005a; Wills et al., 2005), we did not observe any form of firing field rearrangement (“remapping”) that was specific to the memory discrimination trial (Fig. 2F,G), even though spatial remapping of pDGC responses to spatial cue manipulations was greater than of Ammon’s horn (Fig. 1B; Leutgeb et al., 2007). We also detected that firing rates in and out of place fields varied across cue manipulations (Fig. 1C) and changed task contingencies (Figs. 2G). Thus, the responses to cue manipulations replicate prior reports. In contrast, place fields did not remap with the memory discrimination test. This is in opposition to the idea that distinctive spatial discharge patterns underlie spatial memory discrimination, but resembles how the place cell network responds to displaced objects; by changing which object-coding cells co-fire with which place cells, none of which remap (Muller and Kubie, 1987; Rivard et al., 2004). Instead of memory discrimination triggering remapping, we found that DG-dependent memory discrimination is signaled by network inconsistency: weaker correspondence between momentary co-firing amongst cell pairs with overlapping spatial tuning and overall sub-second discharge correlations (Fig. 4). The memory task-related variations in network consistency and the corresponding modulation of excitatory - and inhibitory cell coupling suggest that memory task-related network representations in DG place cell firing are distinguished by sub-second interactions between excitatory discharge signaling spatial information and local control of that discharge by interneurons (Buzsaki, 2010; Danielson et al., 2017). Because the set of weak pair-wise correlations within a network estimates the overall network state (Schneidman et al., 2006) the present observations demonstrate that the network of cells in the DG adopts distinctive states defined by dynamic and coordinated excitatory-excitatory and excitatory-inhibitory interactions to subserve memory discrimination. Like the requirement for memory discrimination, the coordinated discharge interactions were dynamic, and could be best estimated by analysis of momentary activity rather than by session-averaged measurements. The findings (Fig. 4C-E) indicate that the sub-second co-firing of dentate principal cells and interneurons, transiently increases during moments of difficult memory discrimination, pointing to greater interneuron-driven control of dentate network function during difficult memory discriminations (Fig. S2D). These distinctive memory-associated patterns of network discharge presumably feedforward to CA3 to implement memory discrimination. This network view is based on recordings that do not discriminate between granule and mossy cells in the DG, which may have distinctive extracellular discharge properties (Danielson et al., 2017; GoodSmith et al., 2017; Neunuebel and Knierim, 2012; Senzai and Buzsaki, 2017). Mossy cells reside in the hilus of the DG and receive focal input from local granule cells (Amaral, 1978; Scharfman et al., 1990). In addition to exciting granule cells directly, mossy cells excite DG inhibitory neurons broadly along the dorso-ventral axis that then provide strong global inhibition onto granule cells and may consequently establish the lateral-inhibition component of the functional architecture for a competitive network (Amari, 1977), specialized for discriminating input patterns and network states (Scharfman, 2016; Sloviter and Lomo, 2012). Correspondingly, ablating mossy cells causes hyperexcitability in granule cells and impaired pattern separation (Jinde et al., 2013) and according to a recent computational model loss of mossy cell excitatory drive onto granule cells does not affect pattern separation, while loss of the mossy cell driven inhibition of granule cells impairs pattern separation (Danielson et al., 2017). While the present physiological findings cannot distinguish between differential roles of granule cells and mossy cells in memory discrimination, nonetheless a crucial contribution of these DG cells was demonstrated using optogenetic silencing (Fig. 2B; Kheirbek et al., 2013). We were unable to functionally discriminate between the pDGC recordings by dividing them into distinct classes on the basis of features that may discriminate granule cells and mossy cells, including firing field number, firing rate, spike width and preferred theta phase (GoodSmith et al., 2017; Senzai and Buzsaki, 2017). Nonetheless, the findings point to a network perspective that focuses on the interactions amongst granule, mossy, and inhibitory cells to emphasize the network state of the DG rather than the isolated contribution of single cell classes. Future work will do well to define the integrated roles of these classes by determining if specific cell classes make particular contributions that merit distinctive functional classification. Future work should also evaluate the distinctiveness of inputs in comparison to the pDGC output to critically evaluate the notions of pattern separation that have been assigned to these cells. In summary, substantial evidence from behavioral tests have established an important role of the DG in memory discrimination, which we find is mediated by changes in the network state of sub-second interactions amongst excitatory and inhibitory cells within the DG, rather than by the spatial tuning of principal cells measured across the time scales of minutes, much longer than the time scale of memory discrimination and decision.
Author contributions
MvD and AF designed the experiments. MvD collected and analyzed the data. MvD and AF wrote the paper.
Acknowledgements
Supported by NIH grant R01AG043688. We want to acknowledge Younghun Lim and Zejia Angel Yu for help with experiments. We are grateful to Gyorgy Buzsaki, René Hen, and Helen Scharfman for discussions, guidance and comments on the manuscript.
Footnotes
↵* Lead contact