1. Neuroscience
Download icon

Multiple time-scales of decision-making in the hippocampus and prefrontal cortex

  1. Wenbo Tang
  2. Justin D Shin
  3. Shantanu P Jadhav  Is a corresponding author
  1. Graduate Program in Neuroscience, Brandeis University, United States
  2. Neuroscience Program, Department of Psychology, and Volen National Center for Complex Systems, Brandeis University, United States
Research Article
  • Cited 0
  • Views 1,922
  • Annotations
Cite this article as: eLife 2021;10:e66227 doi: 10.7554/eLife.66227

Abstract

The prefrontal cortex and hippocampus are crucial for memory-guided decision-making. Neural activity in the hippocampus exhibits place-cell sequences at multiple timescales, including slow behavioral sequences (~seconds) and fast theta sequences (~100–200 ms) within theta oscillation cycles. How prefrontal ensembles interact with hippocampal sequences to support decision-making is unclear. Here, we examined simultaneous hippocampal and prefrontal ensemble activity in rats during learning of a spatial working-memory decision task. We found clear theta sequences in prefrontal cortex, nested within its behavioral sequences. In both regions, behavioral sequences maintained representations of current choices during navigation. In contrast, hippocampal theta sequences encoded alternatives for deliberation and were coordinated with prefrontal theta sequences that predicted upcoming choices. During error trials, these representations were preserved to guide ongoing behavior, whereas replay sequences during inter-trial periods were impaired prior to navigation. These results establish cooperative interaction between hippocampal and prefrontal sequences at multiple timescales for memory-guided decision-making.

Introduction

The neural substrates that support decision-making are still not fully understood. The link between decision-making and neural representations at the behavioral timescale has been studied extensively in various cortical and sub-cortical circuits of different species. Early classic work showed that during tasks involving sustained attention or decision-making, neurons in the prefrontal cortex (PFC) and the posterior parietal cortex (PPC) can exhibit persistent activity over seconds throughout retention intervals for maintenance of decision-related information (Fuster, 2015; Goldman-Rakic, 1995; Miller et al., 2018; Sreenivasan and D'Esposito, 2019). In contrast to these low-dimensional representations that require long-lived stable states, decision-related information can also be held in a dynamic population code. At the ensemble level, heterogenous activity patterns comprising sequences of neuronal activation that span entire task periods have emerged as a common coding scheme in many brain regions, including PFC (Baeg et al., 2003; Fujisawa et al., 2008; Ito et al., 2015), hippocampus (Ito et al., 2015; Pastalkova et al., 2008), PPC (Crowe et al., 2010; Harvey et al., 2012), and striatum (Bakhurin et al., 2016; Barnes et al., 2005).

In addition to this behavioral-timescale activity, recent work has raised the possibility that neural dynamics at fast, cognitive timescales that occur transiently during discrete subsets of task periods can also underlie upcoming decisions. In many brain areas, such as prefrontal, parietal, orbitofrontal cortices, and hippocampus, population activity can change in an abrupt, coordinated, and transient manner in support of flexible decisions (Bernacchia et al., 2011; Durstewitz et al., 2010; Johnson and Redish, 2007; Karlsson et al., 2012; Latimer et al., 2015; Lundqvist et al., 2016; Rich and Wallis, 2016). For example, discrete transient bursts of gamma and beta oscillations in PFC have been shown to increase with working-memory load during delays (Lundqvist et al., 2016). In particular, recent studies have identified time-compressed neuronal sequences in the hippocampus as a specific cell-assembly pattern at fast timescales that can support decision-making processes.

Hippocampal theta sequences provide time-compressed ensemble representations of spatial paths within single cycles of theta oscillations (6–12 Hz) during active navigation, which reflect a candidate neural mechanism for planning at decision time (Johnson and Redish, 2007; Kay et al., 2020; Papale et al., 2016; Pezzulo et al., 2019; Zheng et al., 2020; Zielinski et al., 2020). Whether theta sequences exist in PFC during decision-making has yet to be determined. Further, during pauses in exploration, replay sequences are observed in the hippocampus during individual sharp-wave ripples (SWRs; Carr et al., 2011), and are known to interact with PFC to reactivate past and future trajectories on the temporal scale of 100–200 ms during memory-guided decisions (Shin et al., 2019; Tang and Jadhav, 2019). Disruption of fast sequences in the hippocampus, while leaving behavioral-timescale spatial representations intact, can impair navigation decisions (Fernández-Ruiz et al., 2019; Jadhav et al., 2012; Petersen and Buzsáki, 2020; Robbe and Buzsáki, 2009). Thus, fast hippocampal sequences are promising transient activity patterns that can support decision-making at sub-second speed (Buzsáki et al., 2014; Kay et al., 2020; Papale et al., 2016; Pezzulo et al., 2019; Shin et al., 2019). Whether and how these fast-timescale representations are organized together in hippocampal-prefrontal circuits, and how they, especially theta sequences, are linked to behavioral-timescale mechanisms for decision-making is unknown. To address these questions, we examined neuronal ensemble activity simultaneously in the hippocampus and PFC of rats during learning of a spatial working-memory decision task.

Results

Choice-specific neuronal sequences in CA1 and PFC during navigation decisions

We trained nine rats to learn a spatial working-memory task (Figure 1A). In this delayed alternation task, animals had to traverse a spatial delay section (i.e. common ‘center stem’; no enforced delay) of a W-maze on each trial, and the critical memory demand of this task is to distinguish left (L) versus right (R) choices (Figure 1A): when the animals return inward from the side arm to the center reward well, they are required to remember their past choice between two possible locations (L vs. R arm; inbound reference-memory trial, Figure 1A, left); and have to choose the opposite side arm correctly for reward after running outward through the stem when facing the two upcoming options (outbound working-memory trial, Figure 1A, right) (Jadhav et al., 2012; Kim and Frank, 2009). This task is known to require both the hippocampus and PFC for learning (Fernández-Ruiz et al., 2019; Jadhav et al., 2012; Kim and Frank, 2009; Maharjan et al., 2018), and involves memory-guided decision-making (Jadhav et al., 2016; Shin et al., 2019; Yu and Frank, 2015). All subjects learned the task rules over eight training sessions (or epochs; 15–20 mins per sessions) in a single day, and performed with high levels of accuracy at the end of the training (Figure 1B; final performance: 92.5 ± 1.8% for inbound, 80.8 ± 2.8% for outbound, in mean ± SEM).

Figure 1 with 1 supplement see all
Choice-specific sequences in CA1 and PFC during memory-guided navigation decisions.

(A) Diagrams of two possible past (inbound; left) and future (outbound; right) scenarios during a W-maze spatial alternation task. In this task, rats have to remember their past choice between two possible locations (Left vs. Right arms; left), and then choose the opposite arm correctly (right; see Materials and methods). CP: choice point. (B) Behavioral performance of all nine rats that learned the inbound (IN; green) and outbound (OUT; blue) components of the task over eight sessions within a single day. Dashed lines: individual animals. Thick lines with error bars: means with SEMs. Horizontal dashed lines: chance-level of 0.5. Prob.: probability. (C) Simultaneously recorded ensembles of CA1 and PFC cells, forming slow behavioral sequences spanning the entire trajectory (~8 s), fast replay sequences (~260 ms; brown shading), and fast theta sequences (~100–200 ms; blue shadings). Each row represents a cell ordered and color-coded by spatial-map center on the Center-to-Left (C–to–L) trajectory shown top right. Gray lines: actual position. Dashed vertical line: reward well exit. Black, brown, and dark blue lines: broadband, ripple-band, and theta-band filtered LFPs from one CA1 tetrode, respectively. (D–I) Choice-predictive representations of behavioral sequences in CA1 and PFC. (D and E) Four trajectory-selective example cells during (D) outbound and (E) inbound navigation (shadings: SEMs; black arrowheaded line indicates animal’s running direction). (F and G) Example choice-specific behavioral sequences. For each plot pair, the top illustrates a raster of trajectory-selective cell assemblies ordered by spatial-map centers on the preferred trajectory; the bottom shows population decoding of animal’s choice and locations at the behavioral timescale (bin = 120 ms; note that summed probability of each column across two trajectory types is 1). Green and yellow arrowheads indicate the example cells shown in (D) and (E). Color bar: posterior probability. Green lines: actual position. Blue and red arrowheads: the CP. Note that the rasters only show trajectory-selective cells, whereas the population decoding was performed using all cells recorded in a given region. (H and I) Behavioral sequences in (H) CA1 and (I) PFC predicted current choices. Left: decoding accuracy of current choice over locations (n = 9 rats×8 sessions); Right: decoding accuracy of current choice on the center stem across sessions. Note that the decoding performance is significantly better than chance (50%) over locations and sessions (all p’s < 1e-4, rank-sum tests). Error bars: SEMs. OUT: outbound; IN: inbound.

Figure 1—source data 1

Decoding accuracy at the behavioral timescale.

Each row represents a session from one animal.

https://cdn.elifesciences.org/articles/66227/elife-66227-fig1-data1-v2.xlsx

We used continuous and simultaneous recordings from ensembles of dorsal CA1 hippocampal and PFC neurons as rats learned this task (Figure 1C and Figure 1—figure supplement 1; mean ± SEM = 43.9 ± 7.6 CA1 place cells, 29.8 ± 5.6 PFC cells per session). As a result of spatially specific firing, sequences of CA1 and PFC cells successively activated on the timescale of seconds as the rat ran through a trajectory (i.e. behavioral sequences), as previously reported by several groups (Frank et al., 2000; Fujisawa et al., 2008; Ito et al., 2015; Kinsky et al., 2020; Shin et al., 2019; Stout and Griffin, 2020; Wood et al., 2000). Further, these sequences were reactivated on the timescale of hundreds of milliseconds during SWRs at the reward well (i.e., replay sequences; Figure 1C), confirming our previous findings (Shin et al., 2019). Notably, within each cycle of theta oscillations during navigation, which corresponds to expression of fast CA1 theta sequences (Dragoi and Buzsáki, 2006; Foster and Wilson, 2007; Gupta et al., 2012; Skaggs et al., 1996), we found that PFC cells were organized into sequences as well and could occur concurrently with the hippocampal sequences (Figure 1C). Given these observations, we further quantified these sequences and investigated their roles in decision making.

First, we characterized how CA1 and PFC neurons encode choices on the behavioral timescale. We found that at the single-cell level, many CA1 and PFC cells exhibited strong preferential firing during navigation on L- versus R- side trajectory (trajectory-selective cells; mean ± SEM = 38.4 ± 1.2% in CA1, 23.5 ± 1.3% in PFC for inbound, 35.0 ± 0.7% in CA1, 20.9 ± 0.6% in PFC for outbound; Figure 1D and E), consistent with prior reports (Frank et al., 2000; Fujisawa et al., 2008; Ito et al., 2015; Kinsky et al., 2020; Shin et al., 2019; Stout and Griffin, 2020; Wood et al., 2000). These trajectory-selective cells, when ordered by the peak firing on the preferred trajectory, form unique sequences for each choice type spanning the entire trial length at the behavioral timescale, including the common center stem prior to the choice point (CP) and the side arms after the CP (Figure 1F and G).

To directly assess the cell-assembly representation of the animals’ choices, we used a memoryless Bayesian decoding algorithm (see Materials and methods; decoding bin = 120 ms) (Shin et al., 2019). We found that behavioral sequences in CA1 and PFC consistently predicted the animals’ current location and choice well above chance level across all positions of a trial (Fujisawa et al., 2008; Ito et al., 2015; Kinsky et al., 2020; Shin et al., 2019; Stout and Griffin, 2020; Wood et al., 2000), even on the common center stem (Figure 1H and I and Figure 1—figure supplement 1). Notably, the decoding accuracy on the center stem significantly increased in CA1 (p=0.0003 and <1e-4 for outbound and inbound, respectively), and was stable in PFC across sessions (p’s > 0.05 for outbound and inbound, Kruskal-Wallis test with Dunn’s post hoc; Figure 1H and I, right), in agreement with our previous findings of an increased proportion of CA1 ‘splitter cells’ (center-stem trajectory-selective cells) and a stable proportion of PFC ‘splitter cells’ over learning (Shin et al., 2019). In addition, the larger proportions of inbound splitter cells (Shin et al., 2019) also contributed to better decoding for inbound choices compared to that for outbound (Figure 1H and I). Nonetheless, behavioral sequences of CA1 and PFC cells represented unique choice types throughout the course of learning.

Therefore, these results suggest that choice information progresses through heterogeneous neuronal sequences in CA1 and PFC at the behavioral timescale as rats run along each trajectory. Importantly, these choice representations also provide distinguishable templates for the Bayesian decoder to identify and determine the content of fast theta sequences.

Fast theta sequences in CA1 and PFC

Theta oscillations are prominent in CA1 during navigation, and CA1 cell assemblies are organized into theta sequences within single oscillation cycles. While it is not known whether ordered sequences of PFC cells occur during hippocampal theta oscillations, previous studies have shown that PFC cells phase-lock and phase-precess (i.e. spikes of a cell occur at progressively earlier theta phases as an animal move through its spatial field) to hippocampal theta oscillations (Jadhav et al., 2016; Jones and Wilson, 2005a; Siapas et al., 2005; Sigurdsson et al., 2010); further, theta-frequency synchrony and coherent spatial coding between the hippocampus and PFC is prominent during memory-guided navigation (Benchenane et al., 2010; Gordon, 2011; Hasz and Redish, 2020; Jones and Wilson, 2005b; Sigurdsson et al., 2010; Zielinski et al., 2019).

In order to statistically identify theta sequences in CA1 and PFC, the firing pattern within each candidate hippocampal theta cycle (≥5 cells active in a given brain region) during active running were analyzed using the Bayesian decoding approach (decoding bin = 20 ms), and the sequential structure of the Bayesian reconstructed positions was evaluated by shuffling procedures (see Materials and methods). Using this method, clear theta sequences were found in CA1 during inbound and outbound navigation across all sessions (Figure 2A–C), Intriguingly, significant theta sequences were also detected in PFC (Figure 2D–F). The prevalence of theta sequences in PFC was similar to that in CA1 (Figure 2G), although as expected, higher trajectory scores (suggesting more reliable timing of sequences) were observed in CA1 compared to PFC (Figure 2H). Furthermore, the trajectory scores and slopes (sequence speed) in both regions increased over learning (Figure 2H and I). Finally, consistent with previous studies in CA1 (Gupta et al., 2012; Wikenheiser and Redish, 2015; Zheng et al., 2016), the majority of CA1 (~70%) and PFC (~60%) theta sequences successively represented past, present, and future locations within each theta cycle (i.e. forward sequences; Figure 2J and Figure 2—figure supplement 1A–C), at a velocity approximately 4–15 times faster than an animal’s true running speed (Figure 2I). These properties of CA1 and PFC theta sequences were replicated with a 10 cell threshold and different shuffling procedures (Figure 2—figure supplement 1), suggesting that the observation of theta sequences in CA1 and PFC was not a trivial result of our decoding methodology.

Figure 2 with 2 supplements see all
Theta sequences in CA1 and PFC over learning.

(A and B) Four examples of theta sequences in CA1 for (A) outbound and (B) inbound trajectories. Left: spikes ordered and color coded by spatial-map center on the decoded trajectory (see Right) over a theta cycle. Broadband (black) and theta-band filtered (green) LFPs from CA1 reference tetrode shown below. Middle: corresponding linearized spatial firing rate maps (blue colormap for L-side trajectory, red colormap for R-side trajectory). Right: Bayesian decoding with p-values based on shuffled data denoted (see Materials and methods). Yellow lines: the best linear fit on the decoded trajectory. Cyan lines and arrowheads: actual position. Note that summed probability of each column across two trajectory types is 1. (C) Example CA1 theta sequences across eight sessions (or epochs, E1 to E8). Each column of plots represents a theta-sequence event. Sequence score (r) on the decoded trajectory denoted. (D and E) Four examples of theta sequences in PFC for (D) outbound and (E) inbound trajectories. Data are presented as in (A) and (B). LFPs are from CA1 reference tetrode. (F) Example PFC theta sequences across eight sessions. Data are presented as in (C). (G) Percent of significant theta sequences out of all candidate sequences didn’t change significantly over sessions (Session 1 vs. 8: p’s > 0.99 for CA1 and PFC, Friedman tests with Dunn’s post hoc). Data are presented as mean and SEM. Black line: PFC; Orange line: CA1. (H) Theta sequence scores (r) improved over sessions (****p’s < 1e-4 for CA1 and PFC, Kruskal-Wallis tests with Dunn’s post hoc). Data are presented as median and SEM. (I) Theta sequence slopes increased over sessions (*p=0.0427 for CA1, and ****p<1e-4 for PFC, Kruskal-Wallis tests with Dunn’s post hoc). Dashed gray lines: mean ± SEMs of animals’ running speed (small error bars may not be discernable). Data are presented as median and SEM. (J) Percent of forward theta sequences didn’t change significantly over sessions (p=0.56 and 0.41 for CA1 and PFC, Friedman tests). Data are presented as mean and SEM.

Figure 2—source data 1

Percent of significant theta sequences, sequence score, sequence slope, and percent of forward theta sequences.

https://cdn.elifesciences.org/articles/66227/elife-66227-fig2-data1-v2.xlsx

To test whether theta phase precession can account for the occurrence of theta sequences, we performed a phase-jitter analysis (Foster and Wilson, 2007), in which we randomly chose a phase of a spike from the distribution of possible phases of that cell in the position bin and shift the spike time accordingly for each candidate event. We found that sequence scores of actual events were significantly greater than those of shuffles in both CA1 and PFC (Figure 2—figure supplement 2), and thus theta phase precession does not account for the full extent of observed theta sequences in CA1 and PFC.

Look-ahead of theta sequences during outbound versus inbound navigation

Prior work has shown that hippocampal forward theta sequences encode paths ahead of the animal, potentially providing a ‘look-ahead’ prediction of upcoming locations (Dragoi and Buzsáki, 2006; Foster and Wilson, 2007; Gupta et al., 2012; Lisman and Redish, 2009; Maurer and McNaughton, 2007; Skaggs et al., 1996; Wikenheiser and Redish, 2015). To investigate how these theta sequences related to animals’ memory state and upcoming behavior, we examined the Bayesian reconstructed positions of forward-shifted candidate theta sequences in different theta-phase bins during reference-memory-guided inbound versus working-memory-guided outbound navigation. We found that the positions behind the actual location of the animal were decoded with higher probability during inbound than outbound navigation in both CA1 and PFC, (Figure 3A–D), implying that inbound sequences started farther behind and outbound sequences sweep father ahead of the current animal position. To further confirm this result, we directly compared the start and end positions of each significant theta sequence, and indeed, forward sequences in both CA1 and PFC started farther behind the actual position of the animal during inbound navigation, whereas they ended farther ahead of the animal during outbound navigation (Figure 3E and F). A possibility that could bias this look-ahead difference is that the choice point may represent a salient behavioral state, acting as an attractor state ‘pulling’ representations toward it. This hypothesis would predict a stronger effect when an animal approaches the choice point compared to the reward well. However, we obtained similar results in these two situations (Figure 3—figure supplement 1), indicating the difference in theta-sequence look-ahead is not a simple result of the attractor behavioral states of the choice point or reward well.

Figure 3 with 1 supplement see all
Look-ahead of CA1 and PFC theta sequences differs during outbound versus inbound navigation.

(A and B) Theta sequences representing past, current, and future locations on outbound (OUT) and inbound (IN) trajectories in (A) CA1 and (B) PFC. Each plot shows the averaged Bayesian reconstruction of all forward candidate theta sequences (sequence score r > 0), replicated over two theta cycles for visualization, relative to current position (y = 0; y > 0: ahead, or future location; y < 0: behind, or past location). Color bar: posterior probability. Arrowheaded line: animal’s running direction. (C and D) Distance index in (C) CA1 and (D) PFC across eight sessions (E1–E8), compared the posterior probabilities on future versus past locations (<0, biased to past; >0, biased to future). 1st half of theta phases (light circles): -π to 0; 2nd half of theta phases (dark circles): 0 to π. Error bars: 95% CIs. (E and F) Distributions for start and end of reconstructed trajectories relative to actual position (x = 0) of all significant forward theta sequences in (E) CA1 and (F) PFC. Top: Histograms show the distributions across all eight sessions (dashed vertical lines: median values). Bottom: Averaged trajectory start (light circles) and end (dark circles) positions in individual sessions. ****p<1e-4, **p=0.007, Kolmogorov-Smirnov test.

Figure 3—source data 1

Start and end of reconstructed trajectories of all significant theta sequences.

https://cdn.elifesciences.org/articles/66227/elife-66227-fig3-data1-v2.xlsx

We then investigated potential neuronal mechanisms underlying this shift in ahead-sequence length. On a single-cell level, previous theoretical and experimental evidence has suggested that the initial tail of asymmetric spatial fields allows cells with fields ahead of an animal’s position to fire during earlier theta cycles, which results in ‘look-ahead’ of theta sequences (Burgess and O'Keefe, 2011; Mehta et al., 2002; Mehta et al., 2000; Skaggs et al., 1996; Wikenheiser and Redish, 2015). We therefore examined the relationship between firing field asymmetry and the shift in ahead-sequence length during outbound and inbound navigation (Figure 4A–D). We found that while the asymmetry developed with experience, as reported previously in CA1 (Figure 4A–CMehta et al., 2000), working-memory-guided outbound navigation was associated with fields with a more extended initial tail compared to inbound travel for both CA1 and PFC (Figure 4A–C). This effect was also stronger for trajectory-selective than non-selective cells in both regions (Figure 4D).

Comparisons of spatial-field asymmetry and theta phase precession during outbound versus inbound navigation in CA1 and PFC.

(A and B) Spatial-field asymmetry during Session 1 (top) and Sessions 7–8 (bottom) in (A) CA1 and (B) PFC. Blue: outbound fields (OUT); Green: inbound fields (IN). Left: Averaged firing rate relative to field center (x = 0) across all cells in the given sessions (error bars: SEMs). Right: Distributions of spatial-field asymmetry index (colored vertical lines: mean values). See also single-field examples in (E) and (F). (C) Spatial-field asymmetry index across sessions (****p<1e-4, ***p<0.001, **p<0.01, *p<0.05, n.s. p>0.05, signed-rank tests compared to 0). Lines are derived from polynomial fits. (D) Trajectory-selective cells exhibit highly asymmetric fields on the preferred (Pref) trajectory compared to the non-preferred trajectory (Non-pref). Non-select: non-selective cells. P-values for each condition derived from signed-rank tests compared to 0; p-values across conditions derived from rank-sum tests (****p<1e-4, ***p<0.001, **p<0.01, *p<0.05, n.s. p>0.05). (E and F) Single-cell examples of theta phase precession in (E) CA1 and (F) PFC for outbound (blue) and inbound (green) trajectories. For each example, linearized firing fields are shown on the top (trajectory type denoted above); spike theta phases against positions (i.e. phase precession) within individual spatial fields (indicated by an orange bar on the firing-field plot) are shown on the bottom (phases are plotted twice for better visibility; red lines represent linear-circular regression lines; linear-circular correlation coefficient r and its p-value denoted). AI: spatial-field asymmetry index. (G) Distributions of phase precession slopes for outbound and inbound fields. Top: outbound; bottom: inbound. Vertical lines: median values. (H) Phase precession slopes were similar during outbound versus inbound navigation in CA1 and PFC (n.s., p’s > 0.99, Kruskal-Wallis test with Dunn’s post hoc), and biased toward negative values (****p<1e-4, **p=0.0012, signed-rank tests compared to 0). This bias is stronger in CA1 than PFC (p’s < 1e-4, Kruskal-Wallis test with Dunn’s post hoc). Only fields with significant phase precession (see Materials and methods) are shown (mean ± SEM = 46.0 ± 2.8% in CA1, 11.2 ± 1.7% in PFC for inbound, and 48.8 ± 3.3% in CA1, 10.3 ± 1.2% in PFC for outbound). Error bars: SEMs.

Further, a prominent model suggests that phase precession of individual neurons may give rise to the look-ahead forward sweep of theta sequences in CA1 (Dragoi and Buzsáki, 2006; Skaggs et al., 1996). We therefore examined if the shift in ahead-sequence length was a consequence of change in phase precession speed during inbound versus outbound navigation. Consistent with previous studies (Jones and Wilson, 2005a; Skaggs et al., 1996), we found CA1 and PFC cells phase-precessing to hippocampal theta oscillations (Figure 4E and F). However, similar slopes of theta phase precession were observed for inbound and outbound navigation in both regions (Figure 4G and H), which therefore cannot simply explain the difference in inbound and outbound look-ahead of theta sequences.

Together, these results suggest that beyond the pure sensory features of the environment, memory demands influenced the look-ahead properties of theta sequences in both CA1 and PFC, and the increased look-ahead distance during working-memory-guided outbound navigation allows the animal to represent future locations earlier in the trajectory, which can aid in decision-making.

Theta sequences support vicarious memory recall

How do CA1 and PFC theta sequences relate to the animals’ upcoming choices when multiple options are available, and do they encode current goal throughout navigation similar to the behavioral sequences (Figure 1)? Prior work has reported that hippocampal population activity at the theta timescale can represent alternatives, which potentially supports deliberation (Johnson and Redish, 2007; Kay et al., 2020), and such a population code is linked to single-cell cycle skipping, in which cells fire on alternate theta cycles (Kay et al., 2020). Therefore, we first examined single-cell firing at the theta timescale in CA1 and PFC. Consistent with previous studies (Dragoi and Buzsáki, 2006; Kay et al., 2020), we observed normal theta rhythmic (i.e. non-skipping; firing on adjacent cycles) and cycle skipping (i.e. firing on every other cycle) cells in CA1 (Figure 5A and B). Intriguingly, we found that a large proportion of theta-modulated cells in PFC also fired in regular alternation during hippocampal theta oscillations (i.e. cycle skipping; Inbound: 53.4%, 95 out of 178; outbound: 49.7%, 91 out of 183; Figure 5C and D).

Theta cycle skipping in CA1 and PFC.

(A) Spike rasters of example cells during CA1 theta oscillations. Top: a non-skipping (firing on adjacent cycles) CA1 cell; Middle: a cycle-skipping (firing on every other cycles) CA1 cell; Bottom: a cycle-skipping PFC cell. Black and green lines: broadband and theta-filtered LFPs from CA1 reference tetrode, respectively. (B) Auto-correlograms (ACGs) of three example single cells in CA1 (left) and PFC (right). Each plot is of data from a single type of maze travel (outbound or inbound; travel type denoted). For each plot, cycle skipping index (CSI) is denoted on the upper left corner (CSI < 0: firing on adjacent cycles; CSI > 0: cycle skipping), and cell number with maze travel type (IN or OUT) matched to (A) is denoted on the upper right corner. Red line: low-passed (1–10 Hz) ACG to measure CSI (see Materials and methods). Note that cells on the bottom two rows exhibit cycle skipping, with CSI > 0. (C) ACGs of all theta-modulated cells in CA1 (left) and PFC (right) ordered by their CSIs (high to low from top to bottom). Red arrowheads indicate division between cells with CSI > 0 (above) vs. <0 (below). Each row represents a single cell with one type of maze pass (outbound or inbound). Only cells with theta-modulated ACGs are shown (see Materials and methods). Note that the proportion of theta-cycle skipping cells in CA1 is consistent with that reported in previous studies (Kay et al., 2020). (D) CSI of theta-modulated cells didn’t differ significantly on outbound versus inbound trajectories for each region (n.s., p’s > 0.99 for CA1 and PFC), but was larger in PFC than CA1 (****p<1e-4, Kruskal-Wallis tests with Dunn’s post hoc). Data are presented as mean and SEMs.

Given the single-cell property of cycle-skipping identified above in both regions, we next examined how populations of CA1 and PFC cells encoded choices during theta sequences. Indeed, we found that CA1 theta sequences can encode alternatives (Figure 6A), as previously reported (Johnson and Redish, 2007; Kay et al., 2020). However, when we examined the representations of choices by theta sequences along each trajectory, we found these representations were distinct in CA1 and PFC during decision-making periods prior to the choice point (CP). Before the CP (corresponding to periods on the center stem for outbound navigation; Figure 6B), CA1 theta sequences serially encoded both alternative trajectories (Figure 6C and G and Figure 6—figure supplements 12; Video 1). In contrast, PFC theta sequences preferentially encoded the animal’s current choice (Figure 6E and G and Figure 6—figure supplements 12). After the choice was made (i.e., after CP; Figure 6B), as expected, both CA1 and PFC theta sequences preferentially encoded the animal’s current choice as the animal’s ran down the track toward reward (Figure 6D, F and G and Figure 6—figure supplements 12). Note that after CP, CA1 theta sequences representing the alternative choice only constituted a minority of the total sequences, reminiscent of the previous findings of a small proportion of hippocampal theta sequences representing the alternative running direction on a linear track (Feng et al., 2015; Wang et al., 2020), and encoding locations far away from an animal’s current position (Gupta et al., 2012; Wikenheiser and Redish, 2015).

Figure 6 with 2 supplements see all
Theta-sequence representations of behavioral choices in CA1 and PFC.

(A) Single-event examples of CA1 theta sequences encoding alternative. Two detailed example sequences are shown in (Ai) and (Aii) (presented as in Figure 2A). Three more examples with decoding plots only are shown in (Aiii) (presented as in Figure 2C). (B) Diagram showing two task segments of a behavioral trial (or trajectory): before choice point (gray shading), and after choice point (green shadings). (C–F) Four decoding examples for before and after choice point during outbound navigation in (C and D) CA1 and (E and F) PFC. Left: Animal’s behavior. Green line: the trajectory pass shown on the middle; Blue/red arrowheaded line: currently taken trajectory. Green circles: locations where theta-sequence events occurred (numbered corresponding to middle). Middle: Decoding plots. Data are presented as in Figure 1F and G (bin = 120 ms), except that whenever a theta sequence was detected, the decoding was performed on the theta timescale (bin = 30 ms) and color-coded by trajectory type for clarity (red or blue: R-side or L-side trajectory; bars above show decoded identity and timing of each event). Note that at both timescales, summed probability of each column across two trajectory types is 1. Yellow shading: example event with detailed view shown on the right. Prob.: probability. (G) Percent of theta sequences representing actual or alternative choice before and after choice point (CP) in CA1 (left) and PFC (right) (****p<1e-4, n.s. p>0.05, session-by-session rank-sum paired tests). Error bars: SEMs. (H) Trial-by-trial theta-sequence prediction of choice (****p<0.0001, n.s., p>0.05, trial-label permutation tests). Red horizontal lines: chance levels (i.e. 95% CIs of shuffled data) calculated by permutation tests. CP: choice point. (I) Theta-sequence prediction persists over sessions. Early: Sessions 1–3; Middle: Sessions 4–5; Late: Sessions 6–8. Data are presented as in (H). Only correct trials are shown in (H) and (I). CP: choice point. (J–L) Coherent CA1-PFC theta sequences biased to actual choice. (J and K) Two examples of coherent CA1-PFC theta sequences. (L) Percent of coordinated CA1-PFC theta sequences coherently representing actual vs. alternative choices (for each condition from left to right: p=0.0312, 0.94, 0.0312, 0.48, signed-rank test compared to 50%; for comparisons between two conditions from left to right, p=0.0312 and 0.0469, rank-sum tests). OUT: outbound; IN: inbound. Alter.: alternative choice. Error bars: SEMs.

Figure 6—source data 1

Choice representation of theta sequences, and number of coordinated theta sequences for each session.

https://cdn.elifesciences.org/articles/66227/elife-66227-fig6-data1-v2.xlsx
Video 1
Slow-motion video of behavioral-timescale and theta-sequence decoding in CA1 when the rat is running a Center-to-Right outbound trajectory.

The video plays 7.5 times slower than real time. To better visualize fast theta sequences, when a significant theta sequence is detected, the video plays 15 times slower. Audio represents spiking of all example units shown on the raster (top left; each spike was correspondingly sped up 7.5 times around spike detection for better perception). Bottom left: Decoding plot. For each theta sequence, reconstruction of only the decoded trajectory was shown for clarity. Right: Behavioral video. Green circle: true position. White circle with a pair of arrowheads: estimated position decoded at the behavioral timescale (arrowhead colors indicate trajectory type, blue for L-side trajectory, red for R-side trajectory; solid arrowheads: decoded trajectory type. hollow ones: the alternative). Note that the raster only shows cells that participated in theta sequences for better visualization, whereas the decoding was performed using all place cells recorded.

The choice representations of theta sequences were robust across sessions (Figure 6G), enabling trial-by-trial prediction of decisions, in which upcoming choice was decoded by PFC theta sequences well above chance before the CP, whereas CA1 theta sequences encoded actual and alternative available paths equivalently before the CP (Figure 6H and I). Similar results were found for inbound trials (Figure 6—figure supplements 12; Video 2). Note that these results cannot be accounted for by similar spatial-map templates for L versus R choices on the center stem, because spatial-map activity can decode L versus R choices well above chance within the center stem in both CA1 and PFC, and this decoding accuracy was in fact higher for CA1 than PFC (Figure 1H and I). Furthermore, similar effects were found after controlling for different shuffling procedures (Figure 6—figure supplement 2A) and examining the last theta sequence on the center stem (Figure 6—figure supplement 2D).

Video 2
Slow-motion video of behavioral-timescale and theta-sequence decoding in CA1 when the rat is running a Right-to-Center inbound trajectory.

The video is presented as in Video 1.

Next, we asked if PFC theta sequences encoded choices that were coherent with CA1 sequences within single theta cycles. We detected PFC theta sequences simultaneously with CA1 theta sequences for a subset of theta cycles (Figure 6J and K; mean ± SEM = 10.53 ± 1.42%, and 11.17 ± 1.52% out of significant sequences for outbound and inbound, respectively; p=0.78 comparing outbound vs. inbound proportions, rank-sum test). Among these synchronous sequence events, when CA1 theta sequences represented the actual choice, PFC sequences were also significantly biased to the actual choice, whereas this coherent CA1-PFC representation was not observed when CA1 theta sequences represented the alternative choice (Figure 6L). Note that this result cannot be simply accounted for by the fact that the majority of PFC theta sequences depict the actual choice, assuming independence of CA1 and PFC theta sequences (see Materials and methods).

Taken together, these results suggest that CA1-PFC theta sequences occurred in tandem with, but distinct from, behavioral sequences for choice representations, and that CA1-PFC theta sequences underlie a novel mechanism that supports vicarious memory recall on a fast timescale for deliberative decision-making.

CA1 and PFC sequences during incorrect trials

While we found a clear relationship between CA1 and PFC sequences at both behavioral and theta timescales for upcoming decisions, it remained unclear if these contributed to correct versus erroneous decisions. We therefore analyzed neural activity during correct versus incorrect trials, with incorrect trials corresponding to erroneous outbound navigation to the same side arm as the past inbound visit (Figure 7A). We found that sequential firing that occurred at the behavioral timescale during incorrect trials was similar to that during correct trials (Figure 7B), and the decoding accuracy for the chosen side was comparable for correct and incorrect trials (Figure 7C; inbound trials for the incorrect condition were considered as the one right before an incorrect outbound trial). Furthermore, CA1 and PFC theta-sequence prediction of upcoming choice was also similar for correct and incorrect trials (Figure 7D).

Figure 7 with 2 supplements see all
Choice representations of replay sequences, but not behavioral and theta sequences, were altered in CA1 and PFC during incorrect trials.

(A) Illustration of a set of correct (top) and incorrect (bottom) trials. For incorrect trial, the actual choice is the unrewarded side. (B and C) Behavioral sequences encoded current choice during incorrect trials. (B) Rasters and population decoding during an incorrect outbound trial. Data are taken from the same session and animal and presented as in Figure 1F. (C) Choice decoding accuracy during incorrect trials is not significantly different from that during correct trials in CA1 (n.s., p>0.99 for outbound,=0.67 for inbound) and PFC (n.s., p>0.99 for outbound,=0.11 for inbound; Friedman tests with Dunn’s post hoc). Corr.: correct trials; Incorr.: incorrect trials. Inbound trials for the incorrect condition were taken from the one right before an incorrect outbound trial (i.e. ‘Past’ trial of the diagram shown in A). (D) Choice representations of theta sequences were similar during incorrect and correct (see Figure 6H) trials. Black bars are for PFC theta sequences, orange bars are for CA1. CP: choice point. (E) Example forward CA1-PFC replay sequences representing actual future choice (see also Figure 1C for this event with example cells, and ripples from a different tetrode). (Ei) Ordered raster plot during a SWR event (black line: ripple-band filtered LFPs from one CA1 tetrode). (Eii) Corresponding spatial firing rate maps. (Eiii) Actual (immediate future) trajectory (orange circle: current position when replay sequences occurred). (Eiv) Reactivation strength (trajectory schematics on the bottom). Blue horizontal lines: 95% CIs computed from shuffled data. Red bar: the decoded trajectory. See Figure 7—figure supplement 2C for an example of reverse CA1-PFC replay sequences. (F) CA1-PFC replay strength predicts correct and incorrect responses. Left: Prediction using replay strength of CA1-PFC forward events. Right: Prediction using replay strength of CA1-PFC reverse events. ROC curves were computed for the SVM classifiers (p-value from trial-label shuffling denoted; see Materials and methods). Shadings: SDs.

Figure 7—source data 1

Decoding accuracy at the behavioral timescale for correct versus incorrect trials.

https://cdn.elifesciences.org/articles/66227/elife-66227-fig7-data1-v2.xlsx

Correct versus incorrect trials did not differ in running speed and theta power (Figure 7—figure supplement 1A–B). CA1-PFC theta coherence and the strength of single-cell phase-locking to theta oscillations during navigation, which have been proposed to support spatial working memory (Benchenane et al., 2010; Gordon, 2011; Jones and Wilson, 2005b; Sigurdsson et al., 2010), were also similar between correct versus incorrect trials (Figure 7—figure supplement 1C–L). Therefore, both theta-linked phenomena at the two timescales likely represented maintenance mechanisms for working memory and decision-making, making it plausible that incorrect destinations were chosen prior to embarking on trajectories from the center well.

We therefore examined replay sequences during SWRs in the inter-trial periods prior to trajectory onset (Figure 1C). Previously, we have reported that CA1 replay sequences, similar to its theta sequences reported here, underlie deliberation between actual and alternative choices, whereas CA1-PFC reactivation represents actual choice for current trials (Shin et al., 2019). Here, we confirmed these observations (Figure 7E and Figure 7—figure supplement 2; Video 3). Importantly, using CA1-PFC reactivation strength for actual versus alternative choices preceding the correct and incorrect trials (see Materials and methods), we could predict correct and incorrect responses significantly better than chance (Figure 7F), indicating impaired CA1-PFC reactivation prior to incorrect outbound navigation. These results suggest that CA1-PFC replay sequences during awake SWRs prime initial navigation decisions, which are further maintained by theta-sequence and trajectory-selective mechanisms during retention periods on a trial-by-trial basis, underlying successful performance in the working memory task.

Video 3
Video of CA1 replay sequences representing possible future choices.

The video is displayed in real time, but two times slower during immobility at the reward well to better visualize fast replay sequences. Top left: Raster of example CA1 cells ordered and color coded by place field center on the actual future trajectory (C-to-L). Bottom left: Raster of the same CA1 population shown on top left, but ordered and color coded by place field center on the alternative future trajectory (C-to-R). Right: Behavioral video. Green circle: true position. Large arrowhead: estimated location at the behavioral timescale. Due to the long immobility period at the reward well (9.58 s), only 1 s around each replay event detected was shown. When a replay sequence is detected, the decoded trajectory is represented by an arrowheaded line (colored according to the trajectory type, blue for L-side trajectory, red for R-side trajectory).

Discussion

In this work, we discovered theta sequences, theta cycle skipping and theta-sequence prediction of behavioral choices in PFC. These prefrontal phenomena follow a succession of results on hippocampal sequences, but our study points to different yet complementary roles of prefrontal and hippocampal sequences at multiple timescales. By dissecting fast cognitive-timescale sequences from slow behavioral-timescale sequences in a spatial working-memory and navigation task, these findings thus provide a unified framework that integrates hippocampal and prefrontal mechanisms of multi-timescale cell-assembly dynamics for memory-guided decision-making.

First, during delay periods of the spatial navigation task on the center stem, choice information was maintained by behavioral-timescale sequences in CA1 and PFC on correct as well as incorrect trials. These sequences are contextually modulated by current journeys, and can enable choice-related information processing on a behavioral timescale for planning actions (Harvey et al., 2012; Ito et al., 2015). In addition, we determined the existence of compressed-timescale theta sequences in PFC, similar to previously described CA1 theta sequences (Dragoi and Buzsáki, 2006; Foster and Wilson, 2007; Kay et al., 2020; Skaggs et al., 1996). These transient theta sequences in CA1 and PFC were nested within behavioral-timescale sequences during navigation. However, in contrast to trajectory-dependent firing sequences at the behavioral-timescale, theta sequences retain representations of choice options to support vicarious memory recall and deliberative decision-making, and function in a complementary manner with fast replay sequences involved in decision priming prior to onset of navigation, highlighting novel roles of these compressed sequences in guiding ongoing choice behavior. Notably, the mechanism of using internally generated sequences to simulate future scenarios has been used as a key feature that improves performance of model-based learning computations (Daw and Dayan, 2014; Mattar and Daw, 2018; Pezzulo et al., 2019; Pezzulo et al., 2014).

At the level of single theta cycles, we found that prior to the decision-making point in the navigation task, CA1 theta sequences report both actual and alternative choices, and are unable to predict chosen destination till after the decision has been made. On the other hand, we found robust PFC theta sequences that maintain prediction of upcoming choice prior to the decision point. Thus, theta sequences underlie a cognitive timescale mechanism that also maintains choice information, with the key distinction that this fast timescale mechanism can support vicarious memory recall of different choices, which was not seen at the behavioral timescale. These findings build on previous results that demonstrated theta-timescale activity patterns in the hippocampus representing future locations by recruiting cells encoding positions after the choice point (i.e. non-splitters; Johnson and Redish, 2007; Kay et al., 2020; Papale et al., 2016), and show that trajectory-specific coding of splitter cells at the behavioral timescale is preserved in fast theta sequences for representing future choices prior to the choice point as well. This process may be important in the event that animals have to change decisions or adapt to change in contingencies (Buzsáki et al., 2014; Rich and Wallis, 2016; Yu and Frank, 2015). In agreement with this idea, a recent study has shown that theta timescale mechanisms in CA1 can not only represent possible future paths, but also possible directions of motion on a moment-to-moment basis (Kay et al., 2020). Representation of past locations within theta cycles has also been recently reported (Wang et al., 2020). Complementing these results, previous studies have shown that theta oscillation cycles comprise cognitive computation units, corresponding to segregation of cell assemblies that represent different spatial experiences (Brandon et al., 2013; Geisler et al., 2007; Gupta et al., 2012; Jezek et al., 2011) and alternatives (Johnson and Redish, 2007; Kay et al., 2020; Papale et al., 2016). The results shown here establish that the representation of alternatives in the hippocampus interacts with prefrontal theta sequences in a content-specific manner, which can be used to guide actual choices.

The exact mechanism underlying these interactions, however, remains unclear. It has been shown that cortical neurons are sensitive to temporally organized inputs from the hippocampus (Branco et al., 2010; Siapas et al., 2005; Sigurdsson et al., 2010), and pharmacological disruption of prefrontal activity results in impaired theta sequences in the hippocampus (Schmidt et al., 2019), as well as impaired performance in the W-track task (Maharjan et al., 2018). Behavioral-state-dependent prefrontal-hippocampal interactions during theta oscillations can potentially also be mediated through connections with entorhinal cortex (Fernández-Ruiz et al., 2017), nucleus reuniens (Ito et al., 2015), or other regions (Eichenbaum, 2017) that contribute to theta generation and memory retrieval. The specific roles of these circuits will need to be further elucidated in future investigations.

Notably, theta-associated mechanisms at both behavioral and theta timescales during active navigation supported retention during working memory periods, but did not predict errors in decisions. Rather, we found that compressed SWR replay sequences during inter-trial periods prior to the onset of trajectories prime the decision without apparent external cue triggers, which is maintained from the onset of the trajectory via behavioral and theta sequences. Previous studies have shown that SWR-reactivation in the hippocampus is less coordinated prior to incorrect versus correct trials (Shin et al., 2019; Singer et al., 2013), and disrupting awake SWRs leads to increase in errors in the spatial working memory task (Fernández-Ruiz et al., 2019; Jadhav et al., 2012). In addition, we have previously shown a relationship between reverse and forward CA1-PFC replay with past and future trajectories, respectively (Shin et al., 2019). Consistent with all these prior results, our current findings provide definitive evidence that coherent CA1-PFC replay of future trajectories prior to navigation onset primes the chosen destination in this memory-guided decision-making task, and errors in CA1-PFC replay predict incorrect decisions.

It is important to note that the expression of behavioral- and compressed-timescale sequences representing different trajectories are inextricably linked. Choice-specific representations of both theta and replay sequences depend on choice encoding through trajectory preferred firing of behavioral sequences. Furthermore, trajectory-selective neurons showed an extended tail of their spatial fields, which is inherently a behavioral-timescale characteristic, and likely contributes to trajectory-modulated look-ahead of theta sequences. Finally, there is evidence that degradation of theta sequences results in impaired sequential activation during sleep replay in the hippocampus (Drieu et al., 2018). Thus, it is the interactions among these multi-timescale activity patterns that potentially govern decision-making. The network mechanisms that enable expression of sequences at distinct timescales in multiple circuits remain a key question for future investigation.

Overall, our results provide a critical extension to classic models, which emphasize behavioral-timescale activity patterns typically spanning entire retention intervals, by establishing a role of discrete, fast timescale ensemble activity patterns in decision-making processes. Such a mechanism is broadly supported by recent findings of rapid shifts in activity patterns during decision-making (Bernacchia et al., 2011; Durstewitz et al., 2010; Karlsson et al., 2012; Latimer et al., 2015; Rich and Wallis, 2016; Sadacca et al., 2016), including discrete gamma oscillatory bursts in PFC underlying working memory (Lundqvist et al., 2016; Miller et al., 2018), which occur at a similar timescale to compressed theta and replay sequences. Together, these results suggest the possibility of transient LFP oscillations as informative signatures of fast evolving cell assemblies that bear on decision-making processes, and the cooperative behavioral- and cognitive-timescale mechanisms described here may reflect a general organizing principle of neural dynamics underlying decision-making.

Materials and methods

Key resources table
Reagent type
(species) or resource
DesignationSource or referenceIdentifiersAdditional information
Strain, strain background (Long Evans rats; male)Long EvansCharles RiverCat#: Crl:LE 006
RRID: RGD_2308852
Chemical compound, drugCresyl VioletAcros OrganicsCat#: AC229630050
Chemical compound, drugFormaldehydeFisherCat#: 50-00-0,67561,7732-18-5
Chemical compound, drugIsofluranePatterson VeterinaryCat#: 07-806-3204
Chemical compound, drugKetaminePatterson VeterinaryCat#: 07-803-6637
Chemical compound, drugXylazinePatterson VeterinaryCat#: 07-808-1947
Chemical compound, drugAtropinePatterson VeterinaryCat#: 07-869-6061
Chemical compound, drugBupivacainePatterson VeterinaryCat#: 07-890-4881
Chemical compound, drugBeuthanasia-DPatterson VeterinaryCat#: 07-807-3963
Software, algorithmMATLAB 2017aMathworks, MARRID: SCR_001622, V2017a
Software, algorithmTrodesSpikeGadgetshttps://spikegadgets.com/trodes/, V1.9
Software, algorithmMatclustMattias P. Karlssonhttps://www.mathworks.com/matlabcentral/fileexchange/39663-matclust, V1.7
Software, algorithmLibsvmChang and Lin, 2011RRID:SCR_010243
https://www.csie.ntu.edu.tw/~cjlin/libsvm/, V3.12
Software, algorithmChronuxPartha MitraRRID:SCR_00554
http://chronux.org/, V2.12
Software, algorithmmeasure_phaseprec ToolboxKempter et al., 2012; Sanders et al., 2019https://github.com/HoniSanders/measure_phaseprec
Software, algorithmPrism 8GraphPad SoftwareRRID: SCR_002798, V8.0

Subjects

Nine adult male Long-Evans rats (450–550 g, 4–6 months) were used in this study. All procedures were approved by the Institutional Animal Care and Use Committee at the Brandeis University and conformed to US National Institutes of Health guidelines. Data from six subjects have been reported in an earlier study (Shin et al., 2019).

Animal pre-training

Request a detailed protocol

Animals were habituated to daily handling for several weeks before training. After habituation, animals were food deprived to 85–90% of their ad libitum weight, and pre-trained to run on a linear track (~1 m long) for rewards (sweetened evaporated milk), and habituated to an high-walled, opaque sleep box (~30 × 30 cm) as described previously (Jadhav et al., 2012; Jadhav et al., 2016; Tang et al., 2017). After the pre-training, animals were surgically implanted with a multi-tetrode drive.

Surgical implantation

Request a detailed protocol

Surgical implantation procedures were as previously described (Jadhav et al., 2012; Jadhav et al., 2016; Shin et al., 2019; Tang et al., 2017). Eight animals were implanted with a multi-tetrode drive containing 32 independently moveable tetrodes targeting right dorsal hippocampal region CA1 (−3.6 mm AP and 2.2 mm ML) and right PFC (+3.0 mm AP and 0.7 mm ML) (16 tetrodes in CA1 and 16 in PFC for four animals; 13 in CA1 and 19 in PFC for three animals; 15 in CA1 and 17 in PFC for one animal). One animal was implanted with a multi-tetrode drive containing 64 independently moveable tetrodes targeting the bilateral CA1 of dorsal hippocampus (−3.6 mm AP and ±2.2 mm ML; Figure 1—figure supplement 1A, left) and PFC (+3.0 mm AP and ±0.7 mm ML; Figure 1—figure supplement 1B, left) (30 tetrodes in CA1 and 34 tetrodes in PFC). On the days following surgery, hippocampal tetrodes were gradually advanced to the desired depths with characteristic EEG patterns (sharp wave polarity, theta modulation) and neural firing patterns as previously described (Jadhav et al., 2012; Jadhav et al., 2016; Shin et al., 2019; Tang et al., 2017). One tetrode in corpus callosum served as hippocampal reference (CA1 REF), and another tetrode in overlying cortical regions with no spiking signal served as prefrontal reference (PFC REF). The reference tetrodes reported voltage relative to a ground (GND) screw installed in skull overlying cerebellum. Electrodes were not moved at least 4 hr before and during the recording day.

Behavioral task

Request a detailed protocol

Following recovery from surgical implantation (~7–8 days), animals were food-deprived and again pre-trained on a linear track for at least 2 days before the W-track sessions started. During the recording day, animals were introduced to the novel W-track (Figure 1A;~80 × 80 cm with ~7 cm wide tracks) for the first time, and learned the task rules over eight behavioral sessions (or epochs, denoted as E1-E8; Figure 1B). Each behavioral session lasted 15–20 min and was interleaved with 20–30 min rest sessions in the sleep box (total recording duration ≅ 6 hr within a single day) (Shin et al., 2019). On the W-maze, animals were rewarded for performing a hippocampus- (Fernández-Ruiz et al., 2019; Jadhav et al., 2012; Kim and Frank, 2009) and prefrontal-dependent (Maharjan et al., 2018) continuous alternation task: returning to the center well after visits to either side well (left or right well; inbound trajectories), and choosing the opposite side well from the previously visited side well when starting from the center well (outbound trajectories). Rewards were automatically delivered in the reward wells (left well: L; right well: R; center well: C) triggered by crossing of an infrared beam by the animal’s nose. Therefore, animals performed four types of trajectories during correct behavioral sequences in this task: center-to-left (C-to-L), left-to-center (L-to-C), center-to-right (C-to-R), and right-to-center (R-to-C). Among these trajectory types, C-to-L and C-to-R are outbound trajectories, while L-to-C and R-to-C are inbound trajectories. When animals were on the center stem, the two inbound trajectories thus represented possible past paths (one actual, and one alternative; Figure 1A, left), and the two outbound trajectories represented possible future paths (Figure 1A, right). The learning curves were estimated using a state-space model (Figure 1BJadhav et al., 2012; Shin et al., 2019; Smith et al., 2004).

Behavioral analysis

Request a detailed protocol

Locomotor periods, or theta states, were defined as periods with running speed >5 cm/s, whereas immobility was defined as periods with speed ≤4 cm/s. The animal’s arrival and departure at a reward well was detected by an infrared beam triggered at the well. The well entry was further refined as the first time point when the speed fell below 4 cm/s before the arrival trigger, whereas the well exit was defined as the first time point when the speed rose above 4 cm/s after the departure trigger (Shin et al., 2019). The time spent at a reward well (i.e. immobility period at well) was defined as the period between well entry and exit. Only SWRs occurring during immobility periods at reward wells were analyzed in this study (see also SWR detection). The center stem of the W-maze was defined as the set of linear positions (see Spatial firing rate maps and linearization) between the center well and the center junction (i.e., choice point, CP). For a given behavioral trajectory, the before-CP period was defined as the time spent at the center stem, and the after-CP period was defined as the time spent at locations between 10 cm away from the center stem and the side well (Figure 6 and Figure 6—figure supplements 12). Therefore, for outbound trajectories, before-CP periods began when animals exited the center well and ended when animals reached the choice point (Figure 6), and for inbound trajectories, before-CP periods began when animals entered the choice point from the side arm and ended when animals entered the center well (Figure 6—figure supplements 12).

Neural recordings

Request a detailed protocol

Data were collected using a SpikeGadgets data acquisition system (SpikeGadgets LLC) (Shin et al., 2019; Tang et al., 2017). Spike data were sampled at 30 kHz and bandpass filtered between 600 Hz and 6 kHz. LFPs were sampled at 1.5 kHz and bandpass filtered between 0.5 Hz and 400 Hz. The animal’s position and running speed were recorded with an overhead color CCD camera (30 fps) and tracked by color LEDs affixed to the headstage. Single units were identified by manual clustering based on peak and trough amplitude, principal components, and spike width using custom software (MatClust, M. P. Karlsson) as previously described (Jadhav et al., 2016; Shin et al., 2019; Tang et al., 2017). Only well isolated neurons with stable spiking waveforms were included (Shin et al., 2019).

Histology

Request a detailed protocol

Following the conclusion of the experiments, micro-lesions were made through each electrode tip to mark recording locations (Shin et al., 2019). After 12–24 hr, animals were euthanized (Beuthanasia) and intracardially perfused with 4% formaldehyde using approved procedures. Brains were fixed for 24 hr, cryoprotected (30% sucrose in 4% formaldehyde), and stored at 4°C. The recording sites were determined from post hoc Nissl-stained coronal brain sections based on The Rat Brain in Stereotaxic Coordinates (Paxinos and Watson, 2004Figure 1—figure supplement 1A–B).

Cell inclusion

Request a detailed protocol

Units included in analyses fired at least 100 spikes in a given session. Putative interneurons were identified and excluded based on spike width and firing rate criterion as previously described (Jadhav et al., 2012; Jadhav et al., 2016). Peak rate for each unit was defined as the maximum rate across all spatial bins in the linearized spatial map (see Spatial firing rate maps and linearization). A peak rate ≥3 Hz was required for a cell to be considered as a place cell.

Spatial firing rate maps and linearization

Request a detailed protocol

Spatial firing rate maps (or rate maps) were calculated only during locomotor periods (> 5 cm/s; all SWR times excluded) at positions with sufficient occupancy (> 20 ms). To construct the 1D linearized spatial firing rate maps on different trajectory types, animal’s linear positions were first estimated by projecting its actual 2D positions onto pre-defined idealized paths along the track, and further classified as belonging to one of the four trajectory types (Shin et al., 2019; Tang et al., 2017). The linearized spatial firing rate maps were then calculated using spike counts and occupancies in 2-cm bins of the linearized positions and smoothened with a Gaussian curve (4-cm SD). We found all linearized positions along each trajectory type were sufficiently covered by the spatial firing rate maps of CA1, as well as PFC, populations (Shin et al., 2019).

Trajectory selective index

Request a detailed protocol

To measure the trajectory selectivity of single cells, a trajectory selectivity index (SI) was calculated by comparing the mean firing rates on the Left- (or L-) vs. Right- (or R-) side trajectories for outbound (C-to-L vs. C-to-R) and inbound (L-to-C vs. R-to-C), respectively:

SI=FRLFRRFRL+FRR,

where FRL is the mean firing rate on the L-side trajectory, and FRR is for the R-side trajectory. Only cells that had a peak firing rate ≥3 Hz detected on either the L- or R-side trajectory were considered, and the rate maps in different sessions were analyzed separately. A cell with |SI| > 0.4 in CA1, or |SI| > 0.2 in PFC was classified as trajectory-selective cells (|SI| = 0.384 ± 0.004 and 0.154 ± 0.003 for all CA1 and PFC cells, respectively; mean ± SEM) (Kay et al., 2020). The trajectory type (L vs. R) with highest firing rate was designated as the cell’s preferred (Pref) trajectory, and the other type designated as the non-preferred (Non-pref) trajectory.

SWR detection

Request a detailed protocol

Sharp-wave ripples (SWRs) were detected during immobility periods (≤ 4 cm/s) as described previously (Jadhav et al., 2012; Jadhav et al., 2016; Shin et al., 2019; Tang et al., 2017). In brief, LFPs from CA1 tetrodes relative to the CA1 reference tetrode were filtered into the ripple band (150-250 Hz), and the envelope of the ripple-filtered LFPs was determined using a Hilbert transform. SWRs were initially detected as contiguous periods when the envelope stayed above 3 SD of the mean on at least one tetrode, and further refined as times around the initially detected events during which the envelope exceeded the mean. For replay and reactivation analysis, only SWRs with a duration ≥ 50 ms were included as in previous studies (Pfeiffer and Foster, 2013; Shin et al., 2019).

Theta phases and theta cycles

Request a detailed protocol

Peaks and troughs of theta oscillations, as well as theta phases, were identified on the band-passed (6-12 Hz) LFPs from the CA1 reference tetrode (CA1 REF) (Jadhav et al., 2016; Lubenov and Siapas, 2009). To precisely define a theta cycle for theta sequence detection, theta phase locking of each cell in CA1 was calculated across locomotor periods (> 5 cm/s) in each session using the methods developed in previous reports (Jadhav et al., 2016; Siapas et al., 2005). A phase histogram was then calculated by averaging across all phase-locked CA1 cells (Rayleigh tests at p < 0.05) in each session, and the phase with minimum cell firing was used to separate theta cycles in the given session (approximately valley-to-valley of hippocampal REF theta, or peak-to-peak of hippocampal fissure theta) (Gupta et al., 2012).

Theta phase precession

Request a detailed protocol

Theta phase precession was examined in linearized spatial firing rate maps with a peak rate ≥ 3 Hz, and multiple fields of a single cell were analyzed separately (e.g. Figure 4E, top right). For each firing peak of linearized spatial firing rate maps detected (using MATLAB findpeaks function with a 20-cm minimal peak distance), a spatial field was defined as contiguous positions with rate > 10% of the peak rate, and at least 8 cm large (Figure 4E and FSchmidt et al., 2009). For spikes within each spatial field, phase precession was computed using a circular-linear fit as previously described (cl_corr function in the measure_phaseprec toolbox; https://github.com/HoniSanders/measure_phaseprec) (Kempter et al., 2012; Sanders et al., 2019). The slope, correlation coefficient (r), and its p-value from the circular-linear regression were reported (Figure 4E-H).

Theta power and coherence

Request a detailed protocol

Power spectra and coherograms were computed from the LFPs referenced to GND using multitaper estimation methods from the Chronux toolbox (http://chronux.org; version 2.12) (Shin et al., 2019). We obtained the SD and mean for each frequency across a given session, and normalized the power of that frequency as a z-score (Figure 1—figure supplement 1B). Coherence between a pair of CA1 and PFC tetrodes was calculated during locomotor periods (> 5 cm/s; locations within 15 cm of the reward well were excluded to prevent contamination from SWR activity). Coherograms averaged over all available CA1-PFC tetrode pairs with simultaneously recorded LFPs were shown in Figure 1—figure supplement 1C-D. Theta power and coherence were measured as the mean power and coherence between 6-12 Hz, respectively.

Spatial field asymmetry

Request a detailed protocol

For cells showing significant phase precession (circular-linear regression at p < 0.05), we further analyzed their spatial field asymmetry (Figure 4A-D). Only fields with the highest peak rate of a single cell for each trajectory type, and at least 20 cm large, were used. The spatial fields were then binned into 10% field length relative to field center (Souza and Tort, 2017), and the asymmetry index (AI) was calculated as:

AI=ARALAR+AL,

where AR denoted the area under the firing rate profile to the right of the field center (i.e. x > 0 in Figure 4A and B), while AL represented the same to the left of the field center (i.e. x < 0 in Figure 4A and B). Therefore, a negative AI corresponds to a spatial field with an extended initial tail.

Theta cycle skipping

Request a detailed protocol

To quantify theta cycle skipping in single cells (Figure 5), we measured a cycle skipping index (CSI) on their auto-correlograms (ACGs). Data on different trajectory types were analyzed separately, and thus a single cell could contribute to more than one ACG (Figure 5B–D). For each ACG, data was restricted to locomotor periods (>5 cm/s) that lasted at least 1.5 s, and with at least total 100 spikes (Kay et al., 2020). Each ACG was first estimated as a histogram of nonzero lags across the interval ±400 ms (bin = 10 ms; denoted as ACG_raw) (Brandon et al., 2013; Kay et al., 2020), and was further corrected for the triangular shape caused by finite duration data (Kay et al., 2020; Mizuseki et al., 2009):

ACG(t)=ACG_raw(t)1|t|T,

where t is the time lag (-T < t < T), and T is the total duration of the spike train used to compute the ACG. The corrected ACG was then smoothed (Gaussian kernel, SD = 20 ms) and peak-normalized. To detect the theta-modulated peaks of ACGs, power spectra for ACGs were generated using FFT, and the relative theta power of an ACG was calculated by dividing its power in the theta band (6-10 Hz) by its total power in the 1-50 Hz range (Deshmukh et al., 2010). An ACG with relative theta power > 0.15 was considered theta-modulated. For all theta-modulated ACGs, the ACGs were band-pass filtered between 1 and 10 Hz (Deshmukh et al., 2010), and the amplitudes of the first and second theta peaks on the filtered ACG were then determined by finding a first peak (p1) near t = 0 in the 90-200 ms window, and the second peak (p2) near t = 0 in the 200-400 ms window, as described previously (Kay et al., 2020). The CSI was then determined as:

CSI=p2p1max(p1,p2),

The CSI ranges between −1 and 1, and higher values indicate more theta cycle skipping.

Sequence analysis

Request a detailed protocol

Sequence analysis here focused on three different ensemble sequences: behavioral sequences, theta sequences, and replay sequences (Figure 1C). To evaluate neural representations at the ensemble level, Bayesian decoding was implemented as previously described (Davidson et al., 2009; Shin et al., 2019; Tang et al., 2017; Zhang et al., 1998): a memoryless Bayesian decoder was built for different trajectory types (for outbound, C-to-L vs. C-to-R; for inbound, L-to-C vs. R-to-C) to estimate the probability of animals’ position given the observed spikes (Bayesian reconstruction; or posterior probability matrix):

P(X,Tr|spikes)=P(spikes|X,Tr)P(X,Tr)P(spikes),

where X is the set of all linear positions on the track for different trajectory types (i.e., Tr; Tr ∈{L, R}, where L represents the L-side trajectory, R represents the R-side trajectory), and we assumed a uniform prior probability over X and Tr. Assuming that all N cells active in a sequence fired independently and followed a Poisson process:

P(spikes|X,Tr)=i=1NP(spikesi|X,Tr)=i=1N(τfi(X,Tr))spikesieτfi(X,Tr)spikesi!,

where τ is the duration of the time window (see below), fi(X,Tr) is the expected firing rate of the i-th cell as a function of sampled location X and trajectory type Tr, and spikesi is the number of spikes of the i-th cell in a given time window. Therefore, the posterior probability matrix can be derived as follows:

P(X,Tr|spikes)=C(i=1Nfi(X,Tr)spikesi)eτi=1Nfi(X,Tr),

where C is a normalization constant such that k=12j=1DP(xj,trk|spikes)=1 (xj is the j-th position bin, D is the total length of the track, and trk is the k-th trajectory type; k = 1 or 2, representing L- or R-side trajectory, respectively).

Specifically, for behavioral sequences, the Bayesian decoder was used to decode animal’s current location (x) and choice (tr) (Figure 1F–I and Figure 1—figure supplement 1C–H) as in previous studies (Shin et al., 2019). Data was restricted to locomotor periods (>5 cm/s; locations within 15 cm of the reward well were excluded for decoding to prevent contamination from SWR activity), and binned into 120 ms bins (i.e. τ = 120 ms; moving window with 60 ms overlap). For each time bin, the location and choice (i.e. trajectory type) with maximum decoded probability was compared to the actual position and choice of the animal in that bin (Figure 1F–I and Figure 1—figure supplement 1C–H). Decoding error of positions in this bin was determined as the linear distance between estimated position and actual position (Figure 1—figure supplement 1C–H), and the accuracy of animal’s choices decoded was reported (Figure 1H and I).

For theta sequences, we first defined candidate events as theta cycles with at least five cells active in a given brain region (CA1 or PFC). Only theta cycles with running speed >10 cm/s, and a duration ranging from 100 to 200 ms were used (Feng et al., 2015). A time window of 20 ms (i.e. τ = 20 ms; moving window with 10 ms overlap) was used to examine theta sequence structure at a fast, compressed timescale.

To calculate the distance index of each candidate event (Figure 3C and D), decoded probabilities ± 60 cm around the animal’s current location, and ±1⁄2 cycle around the 0 phase of the theta cycle (i.e. -π to 0 as 1st half of theta phases, and 0 to π as 2nd half of theta phases), were divided equally into four quadrants (Feng et al., 2015). The distance index of the 1st half of theta phases is thus measured by comparing the probabilities in the quadrants ahead (quadrant III, future) and behind (quadrant II, past) as (III – II)/ (III + II). Similarly, the distance index of the 2nd half of theta phases is thus measured as (IV – I)/ (IV + I).

To identify sequential structure within a theta cycle, two measures were adapted from previous theta-sequence studies (Drieu et al., 2018; Farooq and Dragoi, 2019; Feng et al., 2015; Zheng et al., 2016). In the first method, a weighted correlation (Farooq and Dragoi, 2019; Feng et al., 2015), r(x,t|Pmat), was calculated for the posterior probability matrix of each trajectory type (Pmat, D × T, D is the total number of spatial bins, and T is the total number of temporal bins). The weighted means were computed across locations (x) and time (t) as:

EX(x|Pmat)=i=1Tj=1DPmatijxji=1Tj=1DPmatij,
ET(t|Pmat)=j=1Di=1TPmatijtij=1Di=1TPmatij,

and the weighted covariance, covar(x,t|Pmat), was computed as:

covar(x,t|Pmat)=i=1Tj=1DPmatij(xjEX(x|Pmat))(tiET(t|Pmat))i=1Tj=1DPmatij,

where ti is the i-th temporal bin, and xj is the j-th spatial bin of the posterior probability matrix (Pmat). The weighted correlation was then calculated as:

r(x,t|Pmat)=covar(x,t|Pmat)covar(t,t|Pmat)covar(x,x|Pmat),

and the weighted correlation was reported as the sequence score (r).

In the second method, we measured whether the decoded positions in successive temporal bins of the posterior probability matrix were tightly arranged along an oblique line as previously reported (Davidson et al., 2009; Drieu et al., 2018; Feng et al., 2015). Briefly, the best-fit line of a theta sequence (e.g. yellow lines of the decoding plots in Figure 2A–F) was determined by a fitted line that yielded maximum posterior probability in an 8 cm vicinity (d). For a given candidate line with a slope v and an intercept ρ, the average likelihood R that the decoded position is located within a distance d of that line is:

R(v,ρ)=1nk=0n1P(|posvkΔt|d),

where k is the temporal bin of the posterior probability matrix, and ∆t is the moving step of the decoding window (i.e. 10 ms). To determine the best-fit line for each theta sequence, we densely sampled the parameter space of v and ρ (v > 1 m/s to exclude stationary events) to find the value that maximized R (Rmax, i.e. goodness-of-fit).

In order to assess the significance of theta sequences, we circularly shifted the space-bins of the posterior probability matrix (n = 1000 times) as described previously (Drieu et al., 2018; Farooq and Dragoi, 2019; Zheng et al., 2016), and calculated the weighted correlation and the goodness-of-fit from the shuffled data. A sequence was considered significant if it met two criteria: first, its sequence score (i.e. weighted correlation) exceeded the 97.5th percentile or was below the 2.5th percentile (for reverse sequences) of the shuffled distributions; and its goodness-of-fit (Rmax) was higher than the 95th percentile of their shuffles. We considered the significant trajectory type as the decoded trajectory, and if more than one trajectory type were significant, the trajectory with the highest sequence score was considered as the decoded trajectory. For plotting purposes only, a moving window (30 ms, advanced in steps of 5 ms) was used for displaying theta sequences (Figures 2 and 6 and Figure 2—figure supplement 1 and Figure 6—figure supplement 1).

For synchronous CA1 and PFC theta sequences, we examined if their representations were coherent or independent (Figure 6J-L). If the representations of CA1 and PFC theta sequences (denoted as SeqCA1 and SeqPFC, respectively) are stochastically independent, then p(SeqPFC = actual| SeqCA1 = actual) = p(SeqPFC = actual| SeqCA1 = alternative) = p(SeqPFC = actual) = 1 - p(SeqPFC = alternative). Given that p(SeqPFC = actual) and p(SeqPFC = alternative) are significantly different from the chance level (50%; p’s < 1e-4, signed-rank tests compared to 50%), stochastic independence would predict that the distributions of sequences coherently representing actual and alternative (i.e., p(SeqPFC = actual| SeqCA1 = actual) and p(SeqPFC = alternative | SeqCA1 = alternative)) both significantly differ from the chance level. However, if they are dependent, different distributions for actual and alternative should be observed (Figure 6L).

The detection of replay sequences has been described previously (Shin et al., 2019). Briefly, candidate replay events were defined as the SWR events during which ≥ 5 place cells fired. Each candidate event was then divided into 10 ms non-overlapping bins (i.e. τ = 10 ms), and decoded based on the Bayes’ rule described above. The assessment of significance for replay events was implemented by a Monte Carlo shuffle, in which the R-squared from linear regression on the temporal bins versus the locations of the posterior probability matrix was compared to the R-squared derived from the shuffled data (i.e. time shuffle, circularly shuffling temporal bins of the posterior probability matrix). A candidate event with an R-squared that exceeded the 95th percentile of their shuffles (i.e. p<0.05) was considered as a replay event.

For SWR events with significant CA1 replay sequences detected, we further measured CA1-PFC reactivation during these events (Figure 7—figure supplement 2) as described previously (Shin et al., 2019). We only analyzed the events where ≥ 5 place cells and ≥5 PFC cells fired. Briefly, for an event with N CA1 and M PFC cells active, a (N × M) synchronization matrix during RUN (CRUN) was calculated with each element (Ci,j) representing the Pearson correlation coefficient of the linearized spatial firing rate maps on a certain trajectory type (see Spatial firing rate maps and linearization) of the i-th CA1 cell and the j-th PFC cell. To measure the population synchronization pattern during the SWR, the spike trains during the SWR were divided into 10 ms bins and z transformed (Peyrache et al., 2009; Shin et al., 2019). The (N × M) synchronization matrix during the SWR (CSWR) was then calculated with each element (Ci,j) representing the correlation of the spike trains of a CA1-PFC cell pair. Finally, the reactivation strength of this event for each trajectory type was measured as the correlation coefficient (R) between the population matrices, CRUN and CSWR.

Theta-sequence prediction of behavioral choices

Request a detailed protocol

For theta-sequence prediction of behavioral choices (Figure 6H and I and Figure 6—figure supplement 2), trial-by-trial classification analysis was performed using support vector machines (SVMs) through the libsvm library (version 3.12) (Chang and Lin, 2011). For each region (CA1 or PFC), two independent SVMs were trained for before-CP and after-CP periods. For each trial, the numbers of theta sequences representing the actual versus the alternative choice during a given period (before CP, or after CP) were used as a feature (n = 2) to predict the current choice (k = 2, L or R). All classifiers were C-SVMs with a radial basis function (Gaussian) kernel and trained on correct trials. Hyperparameter (C and γ; regularization weight and radial basis function width, respectively) selection was performed using a random search method with leave-one-out cross-validation to prevent overfitting. The selected hyperparameters were then used to report the leave-one-out cross-validation accuracy. The percentage of correctly inferred trials was computed across all training/test trial combinations to give prediction accuracy. The significance of this prediction was determined by comparing to the distribution of shuffled data by randomly shuffling the trial labels (L or R), and this shuffled dataset was used to train a classifier in the same way as the actual dataset. A prediction accuracy based on the actual dataset that was higher than the 95th percentile of its shuffles (p < 0.05) was considered as significant. Only trials with at least one theta sequence for actual and alternative choices were used for prediction.

CA1-PFC reactivation prediction of correct and incorrect responses

Request a detailed protocol

For prediction of correct and incorrect responses with CA1-PFC reactivation (Figure 7F), SVMs were used similar to the theta-sequence prediction analysis (see above). Two independent SVMs were trained on forward and reverse replay events. For each trial, the averaged CA1-PFC reactivation strength for the actual versus the alternative trajectory across all reactivation events during immobility at the reward well prior to the trial was used as a feature (n = 2; Figure 7—figure supplement 2A-B) to predict correct versus incorrect responses (k = 2, correct or incorrect, regardless of which side arm that the animals choose Singer et al., 2013). The significance of this prediction was determined by comparing it to the distribution of shuffled data by randomly shuffling the trial labels (correct or incorrect), and this shuffled dataset was used to train a classifier in the same way as the actual dataset. Given the unbalanced nature of the dataset (a lot more correct trials than incorrect trials), we resampled the incorrect trials (with replacement) to match the correct trials and used ROC analysis to measure the predictive power of the classifiers. The area under each ROC curve (AUC) was computed, and an AUC based on the actual dataset that was higher than the 95th percentile of its shuffles (p < 0.05) was considered as significant. Only trials with at least one reactivation event of the given type (forward or reverse) were used for prediction.

Statistical analysis

Request a detailed protocol

Data analysis was performed using custom routines in MATLAB (MathWorks) and GraphPad Prism 8 (GraphPad Software). We used nonparametric and two-tailed tests for statistical comparisons throughout the paper, unless otherwise noted. We used Kruskal-Wallis or Friedman test for multiple comparisons, with post hoc analysis performed using a Dunn’s test. p<0.05 was considered the cutoff for statistical significance. Unless otherwise noted, values and error bars in the text denote means ± SEMs. No statistical methods were used to pre-determine sample sizes, but our sample sizes are similar to those generally employed in the field.

References

  1. Conference
    1. Buzsáki G
    2. Peyrache A
    3. Kubie J
    (2014)
    Emergence of cognition from action
    Cold Spring Harbor Symposia on Quantitative Biology. pp. 41–50.
    1. Chang C-C
    2. Lin C-J
    (2011) Libsvm
    ACM Transactions on Intelligent Systems and Technology 2:1–27.
    https://doi.org/10.1145/1961189.1961199
    1. Daw ND
    2. Dayan P
    (2014) The algorithmic anatomy of model-based evaluation
    Philosophical Transactions of the Royal Society B: Biological Sciences 369:20130478.
    https://doi.org/10.1098/rstb.2013.0478
  2. Book
    1. Fuster J
    (2015)
    The Prefrontal Cortex
    Academic Press.
    1. Lisman J
    2. Redish AD
    (2009) Prediction, sequences and the Hippocampus
    Philosophical Transactions of the Royal Society B: Biological Sciences 364:1193–1201.
    https://doi.org/10.1098/rstb.2008.0316
  3. Book
    1. Paxinos G
    2. Watson C
    (2004)
    The Rat Brain in Stereotaxic Coordinates
    Academic Press.

Decision letter

  1. Laura L Colgin
    Senior and Reviewing Editor; University of Texas at Austin, United States

In the interests of transparency, eLife publishes the most substantive revision requests and the accompanying author responses.

Acceptance summary:

This is an exciting and important manuscript that will be of broad interest to readers in the field of learning and memory, place cells, sharp wave-ripples, and theta rhythms, as well as those interested in the functional significance of hippocampal-medial prefrontal cortex networks. This paper involves recordings of large neuronal ensembles in the hippocampus and medial prefrontal cortex in animals performing a hippocampal-dependent spatial working memory task. The paper provides a number of descriptions, at different timescales and with different analyses, of how neural activity in the rat hippocampus and prefrontal cortex relates to behavioral performance of a memory task.

Decision letter after peer review:

Thank you for submitting your article "Multiple time-scales of decision making in the hippocampus and prefrontal cortex" for consideration by eLife. Your article has been reviewed by three peer reviewers, one of whom is a member of our Board of Reviewing Editors, and the evaluation has been overseen by Laura Colgin as the Senior and Reviewing Editor. The reviewers have opted to remain anonymous.

The reviewers have discussed their reviews with one another, and the Reviewing Editor has drafted this to help you prepare a revised submission.

Essential Revisions:

Reviewers agreed that this is a strong submission with beautiful data and figures, as well as analyses that are appropriate and solid. However, they felt that there was a lack of clarity regarding the paper's major conclusions and how some of the analyses led to those conclusions. Reviewers are confident that the authors will be able to address these concerns. No new data are required to support the conclusions.

Reviewer #1 (Recommendations for the authors (required)):

1) To avoid confusion, the authors should perhaps reconsider using the phrase "population spikes" here. This phrase often has a different meaning in neurophysiology from what is meant here.

2) To improve readability, the authors may want to consider not using the needless, non-standard abbreviation CP for choice point.

3) In the figures, perhaps a different color than blue could be used for the left arm to improve clarity and readability. The color blue indicates outbound in some figures/panels and left arm trajectory in other figures/panels.

4) I am confused by the broken y-axis in Figure 1B. Why is the axis disrupted with 0.4 written twice?

5) The distance index in Figure 3C-D is not described in the Materials and methods section. The description in the figure legend is confusing. The description states that it compares posterior probabilities for future versus past locations, making it sound as though this means future and past relative to the present location of the animal. However, the figure legend also refers to theta phases, making it sound as though this measure was calculated based on theta phases of spikes. How exactly was this measure computed?

6) The claim "Note the robust difference of CA1-PFC reactivation of taken versus not-taken path during correct trials compared to that during incorrect trials" is not well supported because no interaction effect is reported (i.e., interaction showing that CA1-PFC reactivation strength for same vs. alternative paths was significantly different for correct vs. incorrect trials). It seems as if only multiple paired t-tests were used here. Also, the data for correct and incorrect trials are plotted on different y-axis scales for panels A and B in Figure 7—figure supplement 2, making comparisons of effects difficult.

Reviewer #2 (Recommendations for the authors (required)):

In this manuscript, Tang and colleagues extend previous work on medial prefrontal-hippocampal interactions in several important ways. In rats performing a well-validated spatial working memory task, they demonstrate the existence of theta sequences in PFC ensembles that mirror the organizational properties of theta sequences in hippocampus (albeit with less reliability than the canonical spatial representations in hippocampus). They show further that PFC theta sequences are coordinated with theta sequence representations in hippocampus; specifically, when hippocampus represents the animal's impending choice destination in a theta cycle, there is an increased likelihood that the simultaneous PFC representation shows the same representation. Finally, the authors show that the content of these fast timescale "cognitive" representations in theta cycles is dissociable from slower "behavioral timescale" representations. Specifically, at the behavioral time scale so called "splitter cells" provide sufficient information to decode impending choice from either hippocampus or PFC. In contrast, fast time scale representations diverged between the two structures before the animal committed to one choice option, with PFC representing the animals upcoming choice and hippocampus representing both possible choice options. This dissociation, however, was dynamically modulated, because after animals committed to a decision, both structure represented the chosen outcome.

I think this manuscript is a potentially important contribution to our understanding of PFC-hippocampal dynamics. First, the identification of clear theta sequences in PFC is, on its own, an important result that is complementary to a recent report showing reactivation of spatially-tuned PFC cells that is highly reminiscent of reactivation of place cells in hippocampus (Kaefer...Csicsvari, 2020). In addition, however, the authors probe the coordination of prefrontal and hippocampal theta sequences, and find that there is an apparent relationship that is dynamically modulated by the behavioral demands of the decision making task. Together, I think the body of work reported here is an important contribution to our understanding of how these structures work together.

The analysis is described clearly and with excellent detail, and appears rigorous and appropriate to the authors' questions. This is an incredibly information-rich manuscript, but it is well organized and well explained throughout. Overall, I think this is a strong contribution, and I have only very minor suggestions to improve it.

At several points the authors mention that results in the present manuscript replicated their previous report (Shin et al 2019). This is important, but is somewhat tempered by the fact that 6/9 rats were reanalyzed from that 2019 data set.

Reviewer #3 (Recommendations for the authors (required)):

This is an unusual paper in that it covers a large amount of ground, with 4-5 main claims that each individually have been (or could be) the subject of entire papers. In its attempt to integrate each of these findings into an overall theory, the paper reads a bit like a hybrid between a review and an empirical paper. One the one hand, this is a strength because it aspires to much-needed synthesis; on the other hand, I found it difficult to identify which statements are the primary empirical claims advanced in the paper, and which statements are interpretations. Moreover, for those statements that looked like main claims, I am not convinced that as presented, there is sufficient support. Thus, overall, I think the paper requires (a) more explicit statements about what the key primary claims are, and (b) additional analyses to better support those claims and/or better separation of result and interpretation.

To expand on the above: as far as I can tell, the main novel claims seem to be:

1) the prefrontal cortex (PFC) expresses theta sequences

2) theta sequences in both the hippocampus (HC) and PFC are trajectory-specific ("splitter sequences")

3) PFC, but not HC, theta sequence content predicts future choice

4) joint HC-PFC activity during replay predicts error vs. correct trials

Based on these findings, the authors conclude that interaction at multiple timescales between HC and PFC supports memory-guided decision-making.

If adequately supported, each of these claims would be a useful addition to the field, with the dissociation between HC and PFC theta sequences of particular interest. However, as currently presented, the claims seem out of step with the underlying evidence; in each case, this overall issue could be addressed by bolstering the analyses and/or by making the claims more precise or toned down.

Specifically:

Claim 1: What counts as a "theta sequence" requires clarification. Do the authors consider the ensemble pattern generated from independently phase precessing cells a theta sequence? If yes, then the claim that mPFC has theta sequences follows directly from previous single cell results (Jones and Wilson, 2005 PLoS Biology; Zielinski et al., 2019 J Neurosci) and they should acknowledge this. If no, they need provide a clear definition and associated analysis criterion of what constitutes a theta sequence, perhaps using a peer prediction analysis along the lines of Harris et al. 2003 Nature; Foster and Wilson, 2007 Hippocampus; Chadwick et al., 2015 eLife. Judging by the authors' comments, it seems they are aware of this issue and don't in fact think that their results can be explained by independently phase precessing cells, but if this one of their central claims then a clear definition and associated analysis should be provided up front as well as throughout the relevant sections.

Claim 2: The authors need to be explicit about what, if any, claim they are making about theta sequences being trajectory-specific. If the claim is that HC theta sequences represent alternative futures before making a choice, i.e. recruiting cells representing positions after the choice point, this has been shown previously (Johnson and Redish, 2007 J Neurosci; Kay et al., 2020 Cell). If, on the other hand, the claim is that based on tuning curves prior to the choice point only, i.e. HC and/or PFC theta sequences are trajectory-specific ("splitter sequences") this would be novel, but requires analyses to determine whether the observed sequences are more split than would be expected from a resampled distribution.

Claim 3: This is potentially the most interesting finding, but as currently presented I found it confusing: isn't there an inconsistency between on the one hand, the claim that behavioral-timescale activity supports choice prediction based on pre-choice point (CP) activity in HC (Figure 1H), and on the other, the result that HC pre-CP theta sequences cannot predict choice (Figure 6H)? Is this because the decoding followed by binary classification (as L or R) of theta sequences ignores firing rate differences that could be used to decode future choice? Or is this because the theta sequence analysis throws out some part of the data that is included in the behavioral-timescale analysis? The concern is that if it is the former, a different classifier that uses firing rates would lead to a quite different result; if it is the latter, it seems the current result could be quite sensitive to arbitrary decisions of what data to include. The authors should discuss, and perhaps perform additional analyses, to explain the apparent inconsistency (see minor comments below for suggestions).

Claim 4: By itself, the current analyses of the HC/PFC interaction during SWR periods does not seem enough to conclude that there is something important about the interaction unless the authors also test how predictable performance is from HC and PFC content alone (perhaps with a GLM that could determine how much the prediction is improved by adding their measure of HC-PFC coordination). Also, how does this relate to the Shin et al. Neuron 2019 paper showing that CA1-PFC coordination could distinguish between replay of chosen and non-chosen path: is this just a restatement of that same result?

Regarding the overall interpretation that HC and PFC interact during memory guided decision making: wouldn't that require that HC and PFC theta sequences should be related? In fact, because PFC but not HC theta sequences predict upcoming choice, would that not be equally or more consistent with these structures not interacting? The analysis in Figure 6L could potentially speak to this point, but without additional analysis it is not clear to me if the reported result isn't just a consequence of PFC but not CA1 sequences being predictive of upcoming choice. Without a more direct demonstration of theta sequence coordination and a relationship to choice, the evidence for interaction seems to be mostly based on the SWR analysis, which as pointed out above needs to be more thorough in order to support an interaction.

– Different criteria for selecting splitter cells are used in HC and PFC, why not use the same criteria for both?

– Goal vs trajectory representation: "maintained representations of current goals"; "modulated by journeys and goals" - the authors should avoid making claims about goal representation: unless the authors analyze error trials, goal representation is confounded with memory of the previous trial. In addition, splitter cells seem to be better described as representing trajectories rather than goals (Grieves et al., eLife 2016).

– The replay results are given relatively high importance in the discussion and feature in the abstract, but are only presented in a supplementary figure; its relative prominence across sections should be made more consistent.

– "In this delayed alternation task, animals had to traverse a delay section ..." please clarify if an actual delay was enforced (i.e. keeping the animal confined to the central stem for some specified time) or if the task was self-paced. This is important because adding delays to continuous alternation tasks is known to make the tasks hippocampus-dependent (e.g. Ainge et al. Hippocampus 2007).

– Figure 1D: if the goal is to show a splitter cell, the authors should show an example where the signal splits in the stem, not around the choice point (where we know behavior is likely to differ)

– "approximately 50% of all theta sequences in both CA1 and PFC met these additional stringent criteria": please explain why the Figure 2—figure supplement 1 - panel D shows a different number (around 10%)?

– The authors should cite and discuss Stout and Griffin 2020 Frontiers Behavioral Neuroscience and Kinsky et al., 2020 Nature Communications as some of their findings are similar to those reported here.

– "both theta-associated mechanisms at the two timescales" - isn't there just one timescale for theta-associated mechanisms, i.e. theta sequences?

– "we have previously shown a relationship between forward and reverse CA1-PFC replay with past and future trajectories respectively (Shin et al., 2019)." It seems that in the cited paper forward replay was associated with future trajectories while reverse replay was associated with past ones, while the text implies the reverse - shouldn't it be "with future and past trajectories respectively"?

– "resampled the incorrect trials with replacement to match the correct trials" - shouldn't it be the reverse (downsampling correct trials?)

– "spatial fields" or "place fields" are generally used to refer to specific regions of high firing of a place cell (O'Keefe, 1976). Instead it seems that what the authors mean here is "rate maps", i.e., representations of firing rate as a function of space including all activity of the cell. Indeed, no criteria for place field detection are indicated in this paragraph. The authors could replace "spatial fields" by rate maps when appropriate and only use spatial fields or place fields when considering the actual place fields to avoid any confusion.

https://doi.org/10.7554/eLife.66227.sa1

Author response

Essential Revisions:

Reviewers agreed that this is a strong submission with beautiful data and figures, as well as analyses that are appropriate and solid. However, they felt that there was a lack of clarity regarding the paper's major conclusions and how some of the analyses led to those conclusions. Reviewers are confident that the authors will be able to address these concerns. No new data are required to support the conclusions. Please see the separate review below for details.

We thank the reviewers for their constructive comments and feedback, which have substantially improved the quality of our manuscript. The reviewers overall found our study important and results significant. Since the manuscript covers a lot of ground, the reviewers had suggestions to clarify the primary claims, some results and interpretations through additional explanation or analyses, which we have fully addressed in the revision.

Reviewer #1 (Recommendations for the authors (required)):

1) To avoid confusion, the authors should perhaps reconsider using the phrase "population spikes" here. This phrase often has a different meaning in neurophysiology from what is meant here.

We thank the reviewer for this helpful suggestion. We have changed the phrase “population spikes” to “the firing pattern”.

2) To improve readability, the authors may want to consider not using the needless, non-standard abbreviation CP for choice point.

We appreciate the reviewer’s concern. We adopted the abbreviation CP for choice point from Singer et al., 2013. To avoid confusion, we have now used the full term, choice point, whenever possible (e.g., Figures 1 and 6, Figure 6—figure supplement 1, and Figure 7—figure supplement 1), except for a few figure panels, where the text becomes too crowded and hinders visualization. In these instances, we redefine the abbreviation for clarity.

3) In the figures, perhaps a different color than blue could be used for the left arm to improve clarity and readability. The color blue indicates outbound in some figures/panels and left arm trajectory in other figures/panels.

We thank the reviewer for this thoughtful comment. We apologize that the colors for left choice and outbound trajectories were difficult to discern, and have now used a different color to indicate left choice in all related figures (e.g., Figures 1 and 6).

4) I am confused by the broken y-axis in Figure 1B. Why is the axis disrupted with 0.4 written twice?

We apologize for not being clear on this. The top and bottom segments of the y-axis are segregated at 0.4 to emphasize the data in the relevant range, and 0.4 was repeated erroneously. For clarity, we have updated this representation.

5) The distance index in Figure 3C-D is not described in the Materials and methods section. The description in the figure legend is confusing. The description states that it compares posterior probabilities for future versus past locations, making it sound as though this means future and past relative to the present location of the animal. However, the figure legend also refers to theta phases, making it sound as though this measure was calculated based on theta phases of spikes. How exactly was this measure computed?

We apologize for not having provided enough detail about the distance index. To calculate the distance index of each candidate event, decoded probabilities ±60 cm around the animal’s location, and ±1⁄2 cycle around the 0 phase of the theta cycle (i.e., -π to 0 as 1st half of theta phases, and 0 to π as 2nd half of theta phases), were divided equally into four quadrants. The distance index of the 1st half of theta phases is thus measured by comparing the probabilities in the quadrants ahead (quadrant III, future) and behind (quadrant II, past) as (III – II)/ (III + II). Similarly, the distance index of the 2nd half of theta phases is thus measured as (IV – I)/ (IV + I) (Author response image 1).

Author response image 1

We have now added these details into the Materials and methods section:“To calculate the distance index of each candidate event (Figure 3C and D), decoded probabilities ±60 cm around the animal’s current location, and ±1⁄2 cycle around the 0 phase of the theta cycle (i.e., -π to 0 as 1st half of theta phases, and 0 to π as 2nd half of theta phases), were divided equally into four quadrants (Feng et al., 2015). The distance index of the 1st half of theta phases is thus measured by comparing the probabilities in the quadrants ahead (quadrant III, future) and behind (quadrant II, past) as (III – II)/ (III + II). Similarly, the distance index of the 2nd half of theta phases is thus measured as (IV – I)/ (IV + I).”

6) The claim "Note the robust difference of CA1-PFC reactivation of taken versus not-taken path during correct trials compared to that during incorrect trials" is not well supported because no interaction effect is reported (i.e., interaction showing that CA1-PFC reactivation strength for same vs. alternative paths was significantly different for correct vs. incorrect trials). It seems as if only multiple paired t-tests were used here. Also, the data for correct and incorrect trials are plotted on different y-axis scales for panels A and B in Figure 7—figure supplement 2, making comparisons of effects difficult.

We thank the reviewer for raising this point. As per the suggestion, we have now removed the claim “Note the robust difference of CA1-PFC reactivation of taken versus not-taken path during correct trials compared to that during incorrect trials”. We have also changed the y-axis scale for panel A as the same for panel B in Figure 7—figure supplement 2.

The claim of “robust difference” for correct trial compared to incorrect trials referred to lower and significant p-values observed for correct trials compared to incorrect trials (correct trials: p = 0.0016 and 0.0037 for forward and reverse events, respectively; incorrect trials: p = 0.56 and 0.024 for forward and reverse events, respectively; Figure 7—figure supplement 2A-B). These p-values were derived from paired tests on a replay event-by-event basis. The rationale for the paired test here is that the strength of a replay event is affected by the number of active neurons and different aspects of selectivity of these neurons during the event. Since replay events during correct versus incorrect trials are two different sample sets, multiple paired tests and interaction effects between correct versus incorrect trials were thus not used.

To assess the trial-by-trial difference of replay strengths for correct and incorrect conditions, which may not be detectable and/or suitable using multiple paired tests, as the reviewer also noted, we therefore used SVM classifiers to examine whether the replay strengths have predictive power for correct and incorrect responses on a trial-by-trial basis. Indeed, significant differences between correct and incorrect trials are seen using this method, and we now report this result in the main figure (Figure 7F).

Reviewer #2 (Recommendations for the authors (required)):

In this manuscript, Tang and colleagues extend previous work on medial prefrontal-hippocampal interactions in several important ways. In rats performing a well-validated spatial working memory task, they demonstrate the existence of theta sequences in PFC ensembles that mirror the organizational properties of theta sequences in hippocampus (albeit with less reliability than the canonical spatial representations in hippocampus). They show further that PFC theta sequences are coordinated with theta sequence representations in hippocampus; specifically, when hippocampus represents the animal's impending choice destination in a theta cycle, there is an increased likelihood that the simultaneous PFC representation shows the same representation. Finally, the authors show that the content of these fast timescale "cognitive" representations in theta cycles is dissociable from slower "behavioral timescale" representations. Specifically, at the behavioral time scale so called "splitter cells" provide sufficient information to decode impending choice from either hippocampus or PFC. In contrast, fast time scale representations diverged between the two structures before the animal committed to one choice option, with PFC representing the animals upcoming choice and hippocampus representing both possible choice options. This dissociation, however, was dynamically modulated, because after animals committed to a decision, both structure represented the chosen outcome.

I think this manuscript is a potentially important contribution to our understanding of PFC-hippocampal dynamics. First, the identification of clear theta sequences in PFC is, on its own, an important result that is complementary to a recent report showing reactivation of spatially-tuned PFC cells that is highly reminiscent of reactivation of place cells in hippocampus (Kaefer...Csicsvari, 2020). In addition, however, the authors probe the coordination of prefrontal and hippocampal theta sequences, and find that there is an apparent relationship that is dynamically modulated by the behavioral demands of the decision making task. Together, I think the body of work reported here is an important contribution to our understanding of how these structures work together.

The analysis is described clearly and with excellent detail, and appears rigorous and appropriate to the authors' questions. This is an incredibly information-rich manuscript, but it is well organized and well explained throughout. Overall, I think this is a strong contribution, and I have only very minor suggestions to improve it.

At several points the authors mention that results in the present manuscript replicated their previous report (Shin et al 2019). This is important, but is somewhat tempered by the fact that 6/9 rats were reanalyzed from that 2019 data set.

We agree with the reviewer and have reworded ‘replicate’ to state ‘we have previously shown/ reported’, or that the results were consistent with our previous reports. This was especially relevant for the CA1-PFC reactivation results originally reported in Figure 7—figure supplement 2. Here, we would also like to note that the trial-by-trial prediction of correct and incorrect responses based on replay events (originally reported in Figure 7—figure supplement 2E-F) was a new result which was not achieved in Shin et al. 2019, and therefore not just a simple replication of our previous finding. Given the novelty and importance of the replay prediction result, especially in the context of lack of error prediction by theta sequences, we have now elevated this result to the main figure (Figures 7E and 7F).

Reviewer #3 (Recommendations for the authors (required)):

This is an unusual paper in that it covers a large amount of ground, with 4-5 main claims that each individually have been (or could be) the subject of entire papers. In its attempt to integrate each of these findings into an overall theory, the paper reads a bit like a hybrid between a review and an empirical paper. One the one hand, this is a strength because it aspires to much-needed synthesis; on the other hand, I found it difficult to identify which statements are the primary empirical claims advanced in the paper, and which statements are interpretations. Moreover, for those statements that looked like main claims, I am not convinced that as presented, there is sufficient support. Thus, overall, I think the paper requires (a) more explicit statements about what the key primary claims are, and (b) additional analyses to better support those claims and/or better separation of result and interpretation.

To expand on the above: as far as I can tell, the main novel claims seem to be:

1) the prefrontal cortex (PFC) expresses theta sequences

2) theta sequences in both the hippocampus (HC) and PFC are trajectory-specific ("splitter sequences")

3) PFC, but not HC, theta sequence content predicts future choice

4) joint HC-PFC activity during replay predicts error vs. correct trials

Based on these findings, the authors conclude that interaction at multiple timescales between HC and PFC supports memory-guided decision-making.

If adequately supported, each of these claims would be a useful addition to the field, with the dissociation between HC and PFC theta sequences of particular interest. However, as currently presented, the claims seem out of step with the underlying evidence; in each case, this overall issue could be addressed by bolstering the analyses and/or by making the claims more precise or toned down.

We thank the reviewer for noting the importance of our work and providing many constructive suggestions to improve our manuscript.

Regarding primary new claims, the findings of prefrontal theta sequences, theta cycle skipping and theta sequence prediction are all novel, to the best of our knowledge. Further, we would like to argue that it is a significant strength of the manuscript that by presenting these findings in context of hippocampal phenomena, we are able to integrate both hippocampal and prefrontal results for behavioral sequences, theta sequences and replay sequences in a unified framework for memory-guided decision making. While we are aware of the extensive work on hippocampal sequences which is cited in the manuscript, there have been only very few studies investigating the relationship between these different sequences, even just in the hippocampus (e.g., Papale et al., 2016; Drieu et al., 2018), which is a major goal in the field (Pezzulo et al., Ann. N.Y. Acad. Sci., 2017). In the current study, we therefore present the new prefrontal results in the context of hippocampal sequences, and also build upon our previous replay results (Shin et al., 2019) to contrast the roles of the three types of sequences in memory-guided decision making. We hope the reviewer will share our vision and appreciate the strength of these findings.

We have now provided explicit statements and an updated focus on these primary advances of our current study in the Discussion section:

“In this work, we discovered theta sequences, theta cycle skipping and theta-sequence prediction of behavioral choices in PFC. These prefrontal phenomena follow a succession of results on hippocampal sequences, but our study points to different yet complementary roles of prefrontal and hippocampal sequences at multiple timescales. By dissecting fast cognitive-timescale sequences from slow behavioral-timescale sequences in a spatial working-memory and navigation task, these findings thus provide a unified framework that integrates hippocampal and prefrontal mechanisms of multi-timescale cell-assembly dynamics for memory-guided decision making.”

We have also followed the reviewer’s suggestions and provided additional evidence with new analyses and explanations to support our main novel conclusions. In overview, (1) we have now provided evidence that theta sequences in CA1 and PFC observed here, as ensemble-level phenomena, cannot be fully accounted for by single-neuronal phase precession, and thus the observation of the PFC theta sequences is a genuine and novel finding of the present work; (2) we have further clarified that our finding of “trajectory-specific” CA1 and PFC theta sequences is an important extension of the previously observed choice coding of theta sequences in the hippocampus; (3) we have demonstrated that the difference between behavioral- and theta-sequence coding of choices in CA1 and PFC is not a result of the methodological concerns suggested; (4) we have made changes in the presentation of the replay results, and added more explicit interpretation of the theta-sequence coordination result to enhance clarity. We believe that these changes, prompted by the reviewer’s suggestions, make the Results clearer and more accessible.

Specifically:

Claim 1: What counts as a "theta sequence" requires clarification. Do the authors consider the ensemble pattern generated from independently phase precessing cells a theta sequence? If yes, then the claim that mPFC has theta sequences follows directly from previous single cell results (Jones and Wilson, 2005 PLoS Biology; Zielinski et al., 2019 J Neurosci) and they should acknowledge this. If no, they need provide a clear definition and associated analysis criterion of what constitutes a theta sequence, perhaps using a peer prediction analysis along the lines of Harris et al. 2003 Nature; Foster and Wilson 2007 Hippocampus; Chadwick et al. 2015 eLife. Judging by the authors' comments, it seems they are aware of this issue and don't in fact think that their results can be explained by independently phase precessing cells, but if this one of their central claims then a clear definition and associated analysis should be provided up front as well as throughout the relevant sections.

We thank the reviewer for this suggestion, and we agree that it is an important point, and investigating the relationship between theta sequences and phase precession could further strengthen our results. As the reviewer suggested, we have now performed the phase-jitter analysis from Foster and Wilson Hippocampus 2007 (Author response image 2), and as expected, theta phase precession by itself does not account for the full extent of observed theta sequences in CA1 and PFC.

Author response image 2
Cumulative distributions of theta sequence scores (or correlations, r) for actual and phase-jittered events.

(A) Distributions for events during outbound navigation (****p = 3.72e-96, and 1.28e-39 for CA1 and PFC, respectively, KS tests). (B) Distributions for events during inbound navigation (****p = 2.39e-132, and 4.29e-59 for CA1 and PFC, respectively, KS tests). Gray lines: phase-jittered events for a given brain region. Orange line: original CA1 events. Black lines: original PFC events.

As the reviewer noted, determining the relationship between theta sequences and phase precession in the hippocampus has been a topic of active experimental and computational work. This relationship, however, hasn’t been studied before in PFC and is thus of interest. To address this issue, we used the phase-jitter analysis from Foster and Wilson Hippocampus 2007. In brief, for any given cell at any given position, there is a range of theta phases at which the cell will fire. Therefore, within each candidate theta sequence event, for each spike, we randomly chose a phase from the distribution of possible phases, and shifted the spike time accordingly. The correlation between cell and time for each event was thus altered by the phase jittering (Foster and Wilson, Hippocampus, 2007). “If this (phase-position) relationship alone determines the occurrence of theta sequences, it should be possible to choose randomly from any of the phases available for a given cell and given position, and then observe theta sequences to the same degree” (Foster and Wilson, Hippocampus, 2007).

The response figure (Author response image 2) shows the distribution of correlations across 100 phase-jitter shuffles of all events (gray lines), compared with the distribution of correlations of original events (black and orange lines for PFC and CA1, respectively). Although the difference is more apparent for CA1 than PFC, correlations for actual events were significantly greater than those for shuffles in both regions (p(CA1) = 3.72e-96, and p(PFC) = 1.28e-39 for outbound; p(CA1) = 2.39e-132, and p(PFC) = 4.29e-59 for inbound; KS tests).

We have now included this result in a new figure supplement (Figure 2—figure supplement 2), and added the following description into the Results section:

“To test whether theta phase precession can account for the occurrence of theta sequences, we performed a phase-jitter analysis (Foster and Wilson, 2007), in which we randomly chose a phase of each spike from the distribution of possible phases of that cell in the position bin and shift the spike time accordingly for each candidate event. We found that sequence scores of actual events were significantly greater than those of shuffles in both CA1 and PFC (Figure 2—figure supplement 2), and thus theta phase precession does not account for the full extent of observed theta sequences in CA1 and PFC.”

Claim 2: The authors need to be explicit about what, if any, claim they are making about theta sequences being trajectory-specific. If the claim is that HC theta sequences represent alternative futures before making a choice, i.e. recruiting cells representing positions after the choice point, this has been shown previously (Johnson and Redish, 2007 J Neurosci; Kay et al., 2020 Cell). If, on the other hand, the claim is that based on tuning curves prior to the choice point only, i.e. HC and/or PFC theta sequences are trajectory-specific ("splitter sequences") this would be novel, but requires analyses to determine whether the observed sequences are more split than would be expected from a resampled distribution.

We thank the reviewer for the opportunity to clarify the trajectory-specific theta sequences. To determine the content of theta sequences, we used the trajectory-specific tuning curves at the behavioral timescale (i.e., rate maps) as templates of the Bayesian decoder, as we described in the Materials and methods section:

“A memoryless Bayesian decoder was built for different trajectory types (for outbound, C-to-L vs. C-to-R; for inbound, L-to-C vs. R-to-C) to estimate the probability of animals’ position given the observed spikes (Bayesian reconstruction; or posterior probability matrix):

P(X,Tr|spikes)=P(spikes|X,Tr)P(X,Tr)P(spikes), where X is the set of all linear positions on the track for different trajectory types (i.e., Tr; Tr ∈{L, R}, where L represents the L-side trajectory, R represents the R-side trajectory),”

Therefore, for the theta sequences that didn’t sweep beyond the choice point (e.g., Figure 6C, event 1), their content was determined by the tuning curves prior to the choice point only. On the other hand, for the theta sequences that swept beyond the choice point, the tuning curves after the choice point (or non-splitter) also contributed, similar to previous findings (Johnson and Redish, JNeurosci, 2007; Kay et al., Cell, 2020). Therefore, as the reviewer correctly pointed out, our findings extend these previous findings to show trajectory-specific theta sequences in CA1 and PFC. While previous reports show that splitter cells encode distinct trajectories at the behavioral timescale (e.g., Frank et al., 2000; Wood et al., 2000; Ito et al., 2015), our finding suggest that splitter information can also be utilized at fast theta timescales. However, we would like to note that, a recent study (Kay et al., Cell, 2020) has investigated the relationship between CA1 splitter cells and theta-timescale activity, and suggested that a splitter cell’s “firing on the non-preferred path is consistent with representation of a possible future, or hypothetical, location” at the theta timescale, although the decoding of theta sequences using splitter templates was not directly tested, which we have studied here. This previous study had also established theta-timescale representation of real and hypothetical heading directions in the hippocampus, which is similar to our trajectory-specific theta sequences in principle.

To determine whether the observed sequences are significantly more split for different choices, the observation of each sequence was tested against two resampled distributions, in which the trajectory coding information (i.e., spatial shuffling) and cell identities (i.e., cell ID shuffling) were randomly shuffled. This method is also similar to that used in previous studies to determine whether the observed theta sequences are significantly more split for different heading directions (e.g., Wang et al., Science, 2020; Kay et al., Cell, 2020).

We now discuss this point more thoroughly in the manuscript. Specifically, we have included the following in the Discussion section:

“These findings build on previous results that demonstrated theta-timescale activity patterns in the hippocampus representing future locations by recruiting cells encoding positions after the choice point (i.e., non-splitters; Johnson and Redish, 2007; Kay et al., 2020; Papale et al., 2016), and show that trajectory-specific coding of splitter cells at the behavioral timescale is preserved in fast theta sequences for representing future choices prior to the choice point as well.”

Claim 3: This is potentially the most interesting finding, but as currently presented I found it confusing: isn't there an inconsistency between on the one hand, the claim that behavioral-timescale activity supports choice prediction based on pre-choice point (CP) activity in HC (Figure 1H), and on the other, the result that HC pre-CP theta sequences cannot predict choice (Figure 6H)? Is this because the decoding followed by binary classification (as L or R) of theta sequences ignores firing rate differences that could be used to decode future choice? Or is this because the theta sequence analysis throws out some part of the data that is included in the behavioral-timescale analysis? The concern is that if it is the former, a different classifier that uses firing rates would lead to a quite different result; if it is the latter, it seems the current result could be quite sensitive to arbitrary decisions of what data to include. The authors should discuss, and perhaps perform additional analyses, to explain the apparent inconsistency (see minor comments below for suggestions).

We thank the reviewer for pointing this out. The reviewer raised the concern that the difference between behavioral versus theta sequences that we found in HC could potentially be accounted for by the fact that we used different decoders for theta versus behavioral sequences, of which theta-sequence decoder ignored firing rate differences, but only used the sequential structure. First of all, we would like to note that our results of hippocampal behavioral and theta sequences confirm previous reports that HC theta sequences on the center stem that sweep beyond the choice point serially encode alternatives (see also responses to Comment 2; e.g., Johnson and Redish, 2007; Kay et al., 2020), and HC trajectory-dependent activity at the behavioral timescale consistently represents upcoming choice on the center stem (e.g., Takahashi, eLife, 2013; Ito et al., 2015). Furthermore, we would like to point out that it is not surprising that decoding of the same data at different timescales can reveal different information (e.g., Kayser et al., Neuron, 2009; Huxter et al., Nature, 2003; Rich and Wallis, 2016; Zheng et al., 2020). To directly address this concern, we have now performed a decoding analysis for behavioral sequences that is identical to theta sequence decoding. This result shows that the difference of behavioral versus theta sequences was not a simple result of different decoders used.

First, the reviewer is correct in pointing out that the Bayesian decoders for behavioral-timescale activity and theta sequences are indeed not identical. For behavioral-timescale activity, we used a bin-by-bin decoding that takes “firing rate differences” into account: for a behavioral sequence before the choice point that can be divided into N behavioral-timescale bins (bin = 120 ms), of each time bin i (i ∈[1, N]), the location (xi) and choice (i.e., trajectory type, tri) with the maximum decoded probability (i.e., max(p(xi,tri|spikes))) was compared to the actual position and choice of the animal in that bin. On the other hand, for a theta sequence that can be divided into N theta-timescale bins (S = [s1, s2, …, sN], si is the firing pattern in the i-th bin; bin = 20 ms), the posterior probability matrix Pmat (p(X,Tr|spikes), X = [x1, x2, …, xN]) was used to compute a trajectory score for each trajectory type (i.e., Tr, Tr ∈{L, R}; see Materials and methods), and compare with its shuffled distributions, with the significant trajectory type that has the highest sequence score as the decoded trajectory. Therefore, this decoder emphasized the sequential structure of a theta-sequence event, instead of just firing rate differences between the two trajectory types. We would like to note that we used the bin-by-bin decoder of behavioral sequences to compare the decoding accuracy with previously reported results of behavioral sequences (e.g., Ito et al., Nature, 2015; Takahashi, eLife, 2013; Harvey et al., Nature, 2012). For theta sequences, however, a “sequential” decoder was often employed (Feng et al., JNeurosci, 2015; Wang et al, Science, 2020; Zheng et al., Neuron, 2015).

However, we understand the reviewer’s concern about the differences of the decoders. There are two ways to make these two decoders identical: using a bin-by-bin decoder of maximum decoded probability for theta-timescale activity; and using a sequential decoder for behavioral sequences.

For the first approach, previous studies have used a bin-by-bin decoder for theta-timescale activity, just as the behavioral-sequence decoder, and found similar results that hippocampal theta-timescale activity can serially represent actual and alternative choices (Johnson and Redish, 2007; Papale et al., 2016).

Author response image 3
Behavioral-sequence representations of choices in CA1 and PFC.

(A and B) Single-event examples of (A) CA1 and (B) PFC behavioral sequences simultaneously recorded during a C-to-R outbound trial. Left: corresponding linearized spatial firing rate maps (blue colormap for L-side trajectory, red colormap for R-side trajectory). Right: Bayesian decoding with sequence score (r) and p-values based on shuffled data denoted. Cyan lines: actual position. Note that a few single-bin decoding errors of current choices occurred along a sequence. Also note the global remapping of some cells on different trajectory types (CP: choice point). (C and D) Percent of behavioral sequences representing actual or alternative choice before the choice point (CP) for outbound (OUT) and inbound (IN) trials in (C) CA1 and (D) PFC (****p < 1e-4, session-by-session rank-sum paired tests). Error bars: SEMs. Data are presented as in Figure 6G. The choice representation of each behavioral sequence was determined by a sequential decoder, similar to that for theta sequences.

On the other hand, given that we focused on the “sequential” structure of ensemble activity in this study, we have now used a sequential decoder for behavioral sequences, just as the theta-sequence decoder (Author response image 3). To do so, we took "an event" of behavioral sequences, as the firing pattern during a single trial (trajectory) on the center stem, and used it as an input to the Bayesian decoder with spatial shuffling, just as used in theta-sequence decoding, but with time window of 400 ms (instead of 20 ms for theta sequences). Unlike CA1 theta sequences that serially encode actual and alternative choices, CA1 behavioral sequences preferentially encode the actual choice. Therefore, this result is consistent with the one found by the bin-by-bin decoder (Figure 1), suggesting that the difference between behavioral versus theta sequences cannot simply be accounted for by the choice of the decoders. Note that the more robust representations here, compared to bin-by-bin decoding shown in Figure 1, are expected, because this measure took the whole sequence structure (i.e., S = [s1, s2, …, sN]) into account, which is less noisy than using the information within a single bin only (i.e., si). Also, we note that these results are consistent with previous findings that splitters cells show not only rate remapping (firing rate difference) but also global remapping (firing position difference) on different trajectory types (e.g., Wood et al., 2000; Frank et al., 2000; Fujisawa et al., 2008; examples in panels A and B of Author response image 3).

Regarding the latter concern, i.e. whether the observed difference between behavioral and theta sequences is because that the theta sequence analysis throws out some part of the data that is included in the behavioral-timescale analysis, first, we show above that using the same classifier for the distinct timescale sequences still yields the same result. Further, as noted above, it is not surprising that decoding of the same data at different timescales can reveal different information (e.g., Kayser et al., Neuron, 2009; Huxter et al., Nature, 2003; Rich and Wallis, 2016; Kay et al., 2020, Zheng et al., 2020). We also want to note that the proportion of significant theta sequences included in this analysis was determined by the Bayesian decoder with shuffling procedures, with numbers consistent with many previous reports of theta sequences in the hippocampus (Gupta et al., 2012; Drieu et al., 2018; Schmidt et al., 2019), and we did NOT make any arbitrary decisions of what data to include. Finally, we also show that applying additional criteria for significance to the Bayesian decoder, which changed the number of theta sequences used, also revealed the same result (Figure 6—figure supplement 2A, left), arguing against the possibility of sensitivity to data used.

To avoid potential confusion, we have now underscored the different window sizes for decoding in the Results section:

(for behavioral sequences) “To directly assess the cell-assembly representation of the animals’ choices, we used a memoryless Bayesian decoding algorithm (see Materials and methods; decoding bin = 120 ms) (Shin et al., 2019).”

and:

(for theta sequences) “In order to statistically identify theta sequences in CA1 and PFC, the firing pattern within each candidate hippocampal theta cycle (≥ 5 cells active in a given brain region) during active running were analyzed using the Bayesian decoding approach (decoding bin = 20 ms),…”

Claim 4: By itself, the current analyses of the HC/PFC interaction during SWR periods does not seem enough to conclude that there is something important about the interaction unless the authors also test how predictable performance is from HC and PFC content alone (perhaps with a GLM that could determine how much the prediction is improved by adding their measure of HC-PFC coordination). Also, how does this relate to the Shin et al. Neuron 2019 paper showing that CA1-PFC coordination could distinguish between replay of chosen and non-chosen path: is this just a restatement of that same result?

We appreciate the reviewer’s concern. We have investigated hippocampal-prefrontal interactions during SWRs in detail elsewhere in several previous studies (Jadhav et al., 2016; Tang et al., 2017; Shin et al. 2019). In one of these studies (Jadhav et al., 2016), we have used GLMs to predict SWR-related spiking of PFC neurons from the activity of the simultaneously recorded CA1 ensembles, and “these findings demonstrate that there is close coordination between CA1 and PFC activity during SWRs”. Consistent results have been shown by several other subsequent studies (e.g., Wang and Ikemoto, JNeurosci, 2016; Yu et al., eLife, 2017).

For the replay result shown here, it is indeed consistent with our previous finding that CA1-PFC coordination could distinguish between replay of chosen and non-chosen paths (Shin et al., 2019), as the reviewer noted. Here, with additional data, we perform a trial-by-trial prediction of correct and incorrect responses based on replay events (Figure 7—figure supplement 2E-F in the previous version; now elevated to Figure 7F in the revised version), which is a new result and was not achieved in Shin et al. 2019. Finally, given the lack of error prediction by theta and behavioral sequences, the result of replay prediction thus provides further insights into the mechanisms underlying accuracy of memory-guided decision making. Together, these findings offer a consolidated view of multi-timescale sequences in the hippocampus and prefrontal cortex.

Regarding the overall interpretation that HC and PFC interact during memory guided decision making: wouldn't that require that HC and PFC theta sequences should be related? In fact, because PFC but not HC theta sequences predict upcoming choice, would that not be equally or more consistent with these structures not interacting? The analysis in Figure 6L could potentially speak to this point, but without additional analysis it is not clear to me if the reported result isn't just a consequence of PFC but not CA1 sequences being predictive of upcoming choice. Without a more direct demonstration of theta sequence coordination and a relationship to choice, the evidence for interaction seems to be mostly based on the SWR analysis, which as pointed out above needs to be more thorough in order to support an interaction.

The reviewer is correct that PFC theta sequences preferentially encode the actual choice, and PFC but not hippocampal theta sequences predict upcoming choice. It is common that two (even directly) interacting brain regions can encode different task information, which by itself cannot reveal whether these structures are interacting or not. For the CA1-PFC coordination in Figure 6L, we argue that this result cannot be simply accounted for by the fact that the majority of PFC theta sequences depict the actual choice, assuming independence of CA1 and PFC theta sequences. For synchronous CA1 and PFC theta sequences, if the representations of CA1 and PFC theta sequences (denoted as SeqCA1 and SeqPFC, respectively) are stochastically independent, then p(SeqPFC = actual| SeqCA1 = actual) = p(SeqPFC = actual| SeqCA1 = alternative) = p(SeqPFC = actual) = 1- p(SeqPFC = alternative). Given that p(SeqPFC = actual) and p(SeqPFC = alternative) are significantly different from the chance level (50%; p’s < 1e-4, signed-rank tests compared to 50%), stochastic independence would predict that the distributions of sequences coherently representing actual and alternative (i.e., p(SeqPFC = actual| SeqCA1 = actual) and p(SeqPFC = alternative | SeqCA1 = alternative)) both significantly differ from the chance level. However, if they are dependent, different distributions for actual and alternative should be observed, which is indeed the case as shown in Figure 6L, i.e., p(SeqPFC = actual| SeqCA1 = actual) significantly differs from the chance level (p’s = 0.0312), but p(SeqPFC = alternative | SeqCA1= alternative) does not (p = 0.94 and 0.48 for outbound and inbound, respectively; signed-rank tests).

In addition, we would like to note that the idea of CA1-PFC interactions at the theta timescale is also supported by previous evidence showing that PFC cells phase-lock and phase-precess to hippocampal theta oscillations (Jadhav et al., 2016; Jones and Wilson, 2005a; Siapas et al., 2005; Sigurdsson et al., 2010), as well as theta-frequency synchrony and coherent spatial coding between the hippocampus and PFC (Benchenane et al., 2010; Gordon, 2011; Hasz and Redish, 2020; Jones and Wilson, 2005b; Sigurdsson et al., 2010; Zielinski et al., 2019). Furthermore, disruption of prefrontal activity results in impaired theta sequences in the hippocampus (Schmidt et al., 2019). We have discussed these studies in our manuscript.

We have now explicitly mentioned and discussed this issue in the Results section:

“Note that this result cannot be simply accounted for by the fact that the majority of PFC theta sequences depict the actual choice, assuming independence of CA1 and PFC theta sequences (see Materials and methods).”

and in the Materials and methods section:

“For synchronous CA1 and PFC theta sequences, we examined if their representations were coherent or independent (Figures 6J-6L). […] However, if they are dependent, different distributions for actual and alternative should be observed (Figure 6L).”

– Different criteria for selecting splitter cells are used in HC and PFC, why not use the same criteria for both?

These criteria were defined arbitrarily based on the fact that |SI| = 0.384 ± 0.004 and 0.154 ± 0.003 for all CA1 and PFC cells recorded, respectively. We used these criteria to show examples of trajectory specific ensemble sequences in Figures 1F and 1G (the decoding and statistics of behavioral sequences were nevertheless performed using all cells recorded in a given region). We have used more stringent and conservative criteria with trial-label shuffling procedures to define trajectory-dependent cells (i.e., splitter cells; the trial labels (left or right) were randomly shuffled 1,000 times, and cells with a SI significantly higher than its shuffle surrogates (p < 0.05) were defined as splitter or trajectory-dependent cells), and reported statistics based on these shuffles in Shin et al., 2019.

– Goal vs trajectory representation: "maintained representations of current goals"; "modulated by journeys and goals" - the authors should avoid making claims about goal representation: unless the authors analyze error trials, goal representation is confounded with memory of the previous trial. In addition, splitter cells seem to be better described as representing trajectories rather than goals (Grieves et al., eLife 2016).

We agree with the reviewer about careful wording for “goal representation”. To better reflect our findings, “maintained representations of current goals” was rephrased to “maintained representations of current choices”, and “modulated by current journeys and goals” was changed to “modulated by current journeys”.

– The replay results are given relatively high importance in the discussion and feature in the abstract, but are only presented in a supplementary figure; its relative prominence across sections should be made more consistent.

Considering the novelty and importance of the replay results as we discussed in Comment 4 (1. the trial by-trial prediction of correct and incorrect responses based on replay events is a new result that wasn’t able to be achieved in Shin et al., 2019; 2. this result, together with similar properties of behavioral and theta sequences during correct and incorrect trials, provides further insights into the mechanisms underlying correct and incorrect responses, as well as a consolidated view of multi-timescale sequences in the hippocampus and prefrontal cortex), we have now elevated the replay prediction result into the main figure (Figures 7E and 7F). We have one sentence in the Abstract describing this result, i.e., “whereas replay sequences during inter-trial periods were impaired prior to navigation”.

– "In this delayed alternation task, animals had to traverse a delay section ..." please clarify if an actual delay was enforced (i.e. keeping the animal confined to the central stem for some specified time) or if the task was self-paced. This is important because adding delays to continuous alternation tasks is known to make the tasks hippocampus-dependent (e.g. Ainge et al., Hippocampus 2007).

The reviewer is correct that we did not enforce a delay in our task (i.e. keeping the animal confined to the central stem for some specified time). We have further clarified this point in the Results section:

“In this delayed alternation task, animals had to traverse a spatial delay section (i.e., common “center stem”; no enforced delay) of a W-maze on each trial.”

As for hippocampus-dependence of the task, we think that the reviewer is referring to tasks similar to maze-based continuous alternation without delay (e.g., continuous alternation on a Figure-8 or T-maze; Ainge et al., 2007). While the reviewer is correct that these tasks have been shown to be hippocampus independent, we would like to clarify a distinction with the W-maze alternation task. It has been shown in previous studies using lesions or optogenetic manipulations that animals are likely to accomplish such a continuous (T-maze/8-maze) spatial alternation task as a “habit” (Ainge et al., Hippocampus, 2007; Sabariego et al., Neuron, 2019), and such tasks do not depend on the hippocampus (Ainge et al., Hippocampus, 2007; Sabariego et al., Neuron, 2019), medial entorhinal cortex (Sabariego et al., Neuron, 2019), or prefrontal-hippocampal communication via nucleus reuniens (Ito et al., 2015). However, the sensitivity of the W-maze task to hippocampal and prefrontal functions is quite different from these continuous alternation tasks. It has been shown that rats with hippocampal lesions exhibited impaired learning of both inbound and outbound components of the task (Kim and Frank, 2009); disruption of SWRs can impair the learning of the outbound component of the task (Jadhav et al., 2012); contralateral hippocampal-PFC inactivation also impairs learning of the outbound component of the task (Maharjan et al., 2018). A distinguishing feature of the W-maze task is that working-memory (outbound) and reference-memory (inbound) demands are regularly interleaved between trials, and the time in the center arm and well serves as a built-in ‘‘delay’’ period (Kim and Frank, 2009), leading to the hippocampal dependence and the difference in learning this task compared to simple continuous alternation mazes.

– Figure 1D: if the goal is to show a splitter cell, the authors should show an example where the signal splits in the stem, not around the choice point (where we know behavior is likely to differ)

An example where the signal splits in the stem has been shown in Figure 1E. Given that both the center-stem tail of a spatial field with its peak located at the side arm or around the choice point, and a spatial field with its peak located within the center-stem could be used to define splitter cells and contribute to decoding of upcoming choices (e.g., Wood et al. 2000; Ito et al., 2015), we picked an example for each of these situations in Figures 1D and 1E, respectively.

– "approximately 50% of all theta sequences in both CA1 and PFC met these additional stringent criteria": please explain why the Figure 2—figure supplement 1 - panel D shows a different number (around 10%)?

We apologize for not being clear about this. “Approximately 50% of all theta sequences in both CA1 and PFC met these additional stringent criteria” refers to the fact that 50% of all theta sequences detected using spatial shuffles alone (Figure 2G; i.e., ~20% are significant) were determined as significant using both spatial and cell-ID shuffles (Figure 2—figure supplement 1D; i.e., ~10% are significant, which is ~50% of ~20%). We have now clarified this point in the legends of Figure 2—figure supplement 1D.

– The authors should cite and discuss Stout and Griffin 2020 Frontiers Behavioral Neuroscience and Kinsky et al., 2020 Nature Communications as some of their findings are similar to those reported here.

We thank the reviewer for alerting us of these references, which we have now included in our revised manuscript.

– "both theta-associated mechanisms at the two timescales" - isn't there just one timescale for theta-associated mechanisms, i.e. theta sequences?

“Both theta-associated mechanisms at the two timescales” was referred to the behavioral-timescale and theta-timescale mechanisms during locomotor periods, because both rate-map activity and theta sequences have been shown to be modulated by theta oscillations. We have further clarified this point as:

“Theta-associated mechanisms at both behavioral and theta timescales during active navigation supported retention during working memory periods, but did not predict errors in decisions.”

– "we have previously shown a relationship between forward and reverse CA1-PFC replay with past and future trajectories respectively (Shin et al., 2019)." It seems that in the cited paper forward replay was associated with future trajectories while reverse replay was associated with past ones, while the text implies the reverse - shouldn't it be "with future and past trajectories respectively"?

We thank the reviewer for bringing this to our attention and we apologize for this typo. We have now corrected this:

“We have previously shown a relationship between reverse and forward CA1-PFC replay with past and future trajectories respectively (Shin et al., 2019).”

– "resampled the incorrect trials with replacement to match the correct trials" - shouldn't it be the reverse (downsampling correct trials?)

It is correct that we used up-sampling (i.e., randomly sample, with replacement, the minority class to be the same size as the majority class), instead of down-sampling, for the ROC analysis. While both up-sampling and down-sampling are commonly used to resolve class imbalance for classification, downsampling always results in a loss of information, and the cost of misclassification of the minority samples is often higher than that of the majority samples. For these reasons, we chose up-sampling in this analysis.

– "spatial fields" or "place fields" are generally used to refer to specific regions of high firing of a place cell (O'Keefe, 1976). Instead it seems that what the authors mean here is "rate maps", i.e., representations of firing rate as a function of space including all activity of the cell. Indeed, no criteria for place field detection are indicated in this paragraph. The authors could replace "spatial fields" by rate maps when appropriate and only use spatial fields or place fields when considering the actual place fields to avoid any confusion.

We agree with the reviewer that “rate maps” is a more accurate description than “spatial fields” in this situation. We have now replaced “spatial fields” with “rate maps” (or “spatial firing rate maps”), whenever appropriate, in our revised manuscript. For a few analyses, in which specific regions of high firing were clearly defined (e.g., in Spatial field asymmetry section of the Materials and methods), “spatial field” was used, and it is consistent with recent finding that spatially localized fields across multiple neocortical regions depend on an intact hippocampus (Esteves et al., JNeurosci, 2021).

https://doi.org/10.7554/eLife.66227.sa2

Article and author information

Author details

  1. Wenbo Tang

    Graduate Program in Neuroscience, Brandeis University, Waltham, United States
    Contribution
    Conceptualization, Data Curation, Software, Formal analysis, Validation, Visualization, Methodology, Writing - original draft, Writing - review and editing
    Contributed equally with
    Justin D Shin
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0003-4361-6705
  2. Justin D Shin

    Graduate Program in Neuroscience, Brandeis University, Waltham, United States
    Contribution
    Data curation, Validation, Investigation, Methodology, Writing - original draft, Writing - review and editing
    Contributed equally with
    Wenbo Tang
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0002-7959-7772
  3. Shantanu P Jadhav

    1. Graduate Program in Neuroscience, Brandeis University, Waltham, United States
    2. Neuroscience Program, Department of Psychology, and Volen National Center for Complex Systems, Brandeis University, Waltham, United States
    Contribution
    Conceptualization, Supervision, Project administration, Funding acquisition, Methodology, Writing - original draft, Writing - review and editing
    For correspondence
    shantanu@brandeis.edu
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0001-5821-0551

Funding

National Institutes of Health (R01 MH112661)

  • Shantanu P Jadhav

Richard and Susan Smith Family Foundation (Odyssey award)

  • Shantanu P Jadhav

The funders had no role in study design, data collection and interpretation, or the decision to submit the work for publication.

Acknowledgements

This work was supported by NIH Grant R01 MH112661 and the Smith Foundation Odyssey award to SPJ.

Ethics

Animal experimentation: All procedures were approved by the Institutional Animal Care and Use Committee at Brandeis University (protocol #21001) and conformed to US National Institutes of Health guidelines.

Senior and Reviewing Editor

  1. Laura L Colgin, University of Texas at Austin, United States

Publication history

  1. Received: January 4, 2021
  2. Accepted: March 5, 2021
  3. Accepted Manuscript published: March 8, 2021 (version 1)
  4. Version of Record published: March 25, 2021 (version 2)

Copyright

© 2021, Tang et al.

This article is distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use and redistribution provided that the original author and source are credited.

Metrics

  • 1,922
    Page views
  • 361
    Downloads
  • 0
    Citations

Article citation count generated by polling the highest count across the following sources: Crossref, PubMed Central, Scopus.

Download links

A two-part list of links to download the article, or parts of the article, in various formats.

Downloads (link to download the article as PDF)

Download citations (links to download the citations from this article in formats compatible with various reference manager tools)

Open citations (links to open the citations from this article in various online reference manager services)

Further reading

    1. Neuroscience
    Rebecca A Mount et al.
    Research Article

    Trace conditioning and extinction learning depend on the hippocampus, but it remains unclear how neural activity in the hippocampus is modulated during these two different behavioral processes. To explore this question, we performed calcium imaging from a large number of individual CA1 neurons during both trace eye-blink conditioning and subsequent extinction learning in mice. Our findings reveal that distinct populations of CA1 cells contribute to trace conditioned learning versus extinction learning, as learning emerges. Furthermore, we examined network connectivity by calculating co-activity between CA1 neuron pairs and found that CA1 network connectivity patterns also differ between conditioning and extinction, even though the overall connectivity density remains constant. Together, our results demonstrate that distinct populations of hippocampal CA1 neurons, forming different sub-networks with unique connectivity patterns, encode different aspects of learning.

    1. Neuroscience
    Mahmood S Hoseini et al.
    Research Article

    Visual perception in natural environments depends on the ability to focus on salient stimuli while ignoring distractions. This kind of selective visual attention is associated with gamma activity in the visual cortex. While the nucleus reticularis thalami (nRT) has been implicated in selective attention, its role in modulating gamma activity in the visual cortex remains unknown. Here we show that somatostatin- (SST) but not parvalbumin-expressing (PV) neurons in the visual sector of the nRT preferentially project to the dorsal lateral geniculate nucleus (dLGN), and modulate visual information transmission and gamma activity in primary visual cortex (V1). These findings pinpoint the SST neurons in nRT as powerful modulators of the visual information encoding accuracy in V1, and represent a novel circuit through which the nRT can influence representation of visual information.