Report

Fast-Forward Playback of Recent Memory Sequences in Prefrontal Cortex During Sleep

See allHide authors and affiliations

Science  16 Nov 2007:
Vol. 318, Issue 5853, pp. 1147-1150
DOI: 10.1126/science.1148979

Abstract

As previously shown in the hippocampus and other brain areas, patterns of firing-rate correlations between neurons in the rat medial prefrontal cortex during a repetitive sequence task were preserved during subsequent sleep, suggesting that waking patterns are reactivated. We found that, during sleep, reactivation of spatiotemporal patterns was coherent across the network and compressed in time by a factor of 6 to 7. Thus, when behavioral constraints are removed, the brain's intrinsic processing speed may be much faster than it is in real time. Given recent evidence implicating the medial prefrontal cortex in retrieval of long-term memories, the observed replay may play a role in the process of memory consolidation.

According to memory-consolidation theory, the hippocampus is necessary for the retrieval of recently encoded episodic memories. For remote memories, in contrast, the neocortex is sufficient for recall (14). The transfer of memories from hippocampal to neocortical control is widely believed to involve replay during sleep of the neural patterns representing the memory (57). Consistent with this hypothesis, patterns of brain activity during a task appear to be repeated during subsequent sleep in rats, birds, monkeys, and humans (816). In rats, multielectrode hippocampal recordings have shown that the temporal order of replay during sleep is preserved (17, 18).

Among cortical areas, the medial prefrontal cortex (mPFC) apparently plays a unique role in mediating retrieval of consolidated, remote memories. In both rats and humans, activity in the mPFC is greater during retrieval of remote memories than during retrieval of recent memories; an opposite pattern is seen in the hippocampus (1921). In addition, lesions of mPFC lead to deficits in retrieval of remote memories (19, 20, 22, 23). One might thus expect the mPFC to play a key role in memory consolidation. Indeed, disrupting mPFC activity during consolidation impairs subsequent performance (24).

If mPFC does play a role in consolidation, task-induced neural activity patterns in mPFC may replay during subsequent sleep. We therefore examined neural ensemble activity after performance of a spatial sequence task. Two rats were trained to run to a series of locations around the perimeter of a 1.3-m circular platform with electrical brain stimulation as a reward. Sequences, consisting of eight locations, were repeated throughout the course of a 50-min running session, alternating in blocks of three cued and three noncued (i.e., memory-guided) sequences throughout the session. Each day, a rat ran the sequence task continuously during two 50-min blocks. Neural activity was recorded during a pre-task sleep period preceding the first task block and in two post-task sleep periods following each task block. Each sleep session was 20 to 60 min in duration. Both rats were implanted with micro-drives containing 12 independently manipulable four-conductor electrodes (“tetrodes”) (25), allowing simultaneous recording of 40 to 120 neurons within the anterior cingulate and prelimbic cortices. Memory reactivation was assessed during periods of motionlessness during the sleep session.

The reactivation of the task-related neural patterns was initially assessed with a measure called “explained variance” (EV), based on the firing rate correlation matrix for all pairs of concurrently recorded cells (26). Explained variance measures the proportion of variability in the cell-pair firing-rate correlations during task performance that can be accounted for by correlations during subsequent sleep, taking into account the correlations that existed in the initial sleep session. For this measure, 100% would mean that cell-pair correlations during task and subsequent sleep were identical. Previous reactivation studies have found average EV values of ∼15% in rat hippocampus, ∼10% in rat ventral striatum, and between 5 and 11% in monkey cortex (13, 14, 26). Average EV in the current study was 11% across two rats and a total of 66 sessions. The distribution of values was strongly skewed; the average of EV values within the lower quartile was 4% and that in the upper quartile, 22%. Reactivation was significantly stronger in the sleep following the second task period than in the sleep following the first (6% versus 11%, Student's t test, P <0.01). At least in some sessions, reactivation in mPFC was as strong as, if not stronger than, that reported in the hippocampus or other parts of the brain.

Cross-correlations of spike trains from pairs of cells were used to examine the temporal structure of replaying patterns in mPFC. Analyses were limited to the second post-task sleep, which is the period showing strongest EV. Cells in mPFC tend to fire at specific locations along a segment, with the majority firing selectively during either approach to or departure from at least some reward zones. These response characteristics, combined with the repetitive nature of the task, led to distinct peaks and valleys in many of the cross-correlation plots (Fig. 1A), indicating consistent temporal relationships in neuronal firing. For some cell pairs, the increased number of central peaks and valleys seen during the task were evident again in the post-task sleep, but highly compressed in time, suggesting that temporal patterns playing during the task were replaying during sleep at an accelerated rate. Peaks and valleys were small or nonexistent during pre-task sleep, suggesting that the replaying patterns were induced by task-related activity. Similar results were obtained when data from multiple cell pairs within the same session were compared together (Fig. 1B).

Fig. 1.

Cross-correlations between mPFC cell pairs during task and sleep. (A) Example cross-correlations from four cell pairs. Each column shows data from one cell pair during pre-task sleep, task, and post-task sleep. The y axis shows the number of coincident spikes per second within each 10-ms bin. (B) Sorted cross-correlations from simultaneously recorded cell pairs. Each row in each subpanel shows the cross-correlation between a single pair of cells, scaled so that peak and valley range from zero to one. The rows are sorted according to the temporal offset of the maximum peak during the task. In addition, only cell pairs showing a peak z score exceeding 11 during the task were included (∼5 to 15%; see Supporting Online Material for details). Red indicates the highest coincidence rate and blue, the lowest. The time axis during sleep epochs is magnified.

To quantify the extent to which neural patterns during sleep evolved at a faster rate than those during the task, times of cross-correlation peaks during task and sleep were automatically extracted and compared (Fig. 2A). For peaks nearest the origin, a strong linear relation was found between task and sleep peak times, with slopes indicating that replay during sleep is compressed relative to the task. Because this analysis requires many cell pairs with matching task and sleep cross-correlations, compression rates were only computed for sessions where the explained variance exceeded 15%, a total of 13 sessions (rat 1: 10; rat 2: 3). Compression rates varied between 5.4 and 8.1, with a mean of 6.5 (Fig. 2B).

Fig. 2.

Quantification of temporal compression during reactivation. (A) Task cross-correlation peak times plotted against post-task sleep peak times. In each case, two peaks are extracted from each cross-correlation, one on the left of zero and one on the right. Matches in the upper left and lower right quadrants of the graph are thus precluded. Estimating compression rate depended upon finding corresponding peaks. Therefore, cell pairs were limited to those showing strong similarity between task and sleep cross-correlations after accounting for compression (i.e., cell pairs exhibiting strong reactivation). The best-fit regression line was found with robust regression, a technique less sensitive to outliers than normal regression. (B) Histogram of compression rates extracted with the analysis shown in (A).

The temporal extent of replay is limited. The attenuation of a given peak during sleep seems to be proportional to the offset of the corresponding peak during the task, with peaks beyond about 3 s being almost completely lost (Fig. 1A). To quantify the duration of the replay window, we compared sleep and compressed task cross-correlations at different offset times (Fig. 3A). The correlation between sleep and task cross-correlations was strongest near zero and then dropped off rapidly, reaching near chance levels around 750 ms (Fig. 3B). This suggested that replaying patterns were, on average, coherent for a duration of ∼1.5 s (2 × 750 ms) during sleep. Assuming 6.5 times compression, the mPFC was evidently replaying behavioral events spanning ∼10 s during the task (about half of a complete eight-element sequence).

Fig. 3.

Temporal extent of replay window. (A) Cross-correlations from post-task sleep (dashed line) and task (solid line) were compared with the use of a sliding 200-ms window (gray bar). In this example, the task cross-correlation has been compressed by a factor of 6.8. (B) Strength of the correlation between task and sleep cross-correlations in each window plotted against time (as measured with the sleep time frame). Data shown were averaged across all 13 high reactivation sessions, with vertical bars indicating the SEM. Data from negative time windows were averaged with the corresponding positive time windows.

Template matching (18) was used to assess whether the replay evident in pairwise activity was coherent across ensembles of cells. For each of the eight sequence segments, we created a template by averaging binned firing rates of a group of cells across multiple repetitions of the sequence (Fig. 4A, left). As expected, many strong matches to each template were observed during the task. Strong matches were also observed during post-task sleep, but only when sleep bin size was reduced, effectively compressing the template relative to sleep (Fig. 4A, right). Indeed, the highest number of strong matches occurred for compression factors of 6 to 7 (Fig. 4C), in accord with estimates based on matching cross-correlation peaks. Few matches were observed when the template column order was reversed, indicating that replay occurs in the forward direction.

Fig. 4.

Template matching. (A) (Left) Template from one segment of the sequence, showing firing rate (spikes/bin) from multiple cells (y axis) sorted by time of peak firing (x axis). (Middle) Examples of good matches to the template from sleep (bottom) and task (top) periods. B in size for the task and sleep examples are 100 ms and 14 ms respectively, the latter representing a compression factor of 7. (Right) The two graphs show histograms of match strength (z scores) between templates and “target” data within sequentially ordered 14-s windows. Z scores were derived via random shuffling of template columns. Color indicates the number of matches. In the top graph, template and target bin size are the same. In the bottom graph, target bin size is a factor of 7 smaller than in the template. The time axis is discontinuous because first task and second sleep period were omitted. (B) Match strength between template and a portion of post-task sleep data at different compression factors (indicating the extent to which sleep replay is compressed relative to the task). Most of the strong matches (i.e., red vertical bands) had peaks corresponding to compression factors between 6 and 8. (C) For each compression factor, the number of local maxima (i.e., peaks) with a z score exceeding 4 was divided by the total number of peaks (“Ratio of Peaks”). Data are averaged over all eight templates with the SEM indicated. Neither post-task sleep with reversed templates nor pre-task sleep showed strong matches or evidence of a peak compression rate.

Whether the observed mPFC replay during the first hour of sleep represents off-line rehearsal needed for memory recoding remains an open question; however, the time frame coincides with what other research suggests is a critical window for memory processing. Studies in humans have shown that memory is tied to slow-wave activity within the first few hours of sleep (16, 27). In rats, local injection of drugs disruptive to mPFC function leads to learning deficits when given within 2 hours after a task, but not after this time (24, 28).

The hippocampus reportedly replays events at 5 to 20 times their behavioral rate (17, 2931); however, in the hippocampus, spikes representing adjacent place fields occur in rapid succession within a single theta cycle during behavior (32). Relative to this within–theta cycle rate, reactivation during sleep is not accelerated. In contrast, reactivation in rat mPFC is clearly compressed five to eight times, and a similar effect may be present in primary visual cortex (31). Fast replay in neocortex may reflect the speed of the brain's intrinsic dynamics (e.g., conduction speeds, synaptic delays, etc.) when not constrained by behavioral events. However, the episodes of sequential replay during sleep are limited to windows of a few hundred milliseconds. This may reflect a time limitation imposed by the duration of cortical “up” states induced by slow-wave oscillations (33) during which replay presumably occurs, or it may reflect cumulative drift error in the sequence replay.

In conclusion, two independent analytical methods, cross-correlation matching and template matching, both clearly showed accelerated replay of task-related neural activity patterns within the mPFC during sleep. That this happens (i) in an area implicated in remote memory retrieval and (ii) during a time window in which processes critical to consolidation are unfolding suggests that accelerated replay may be an important part of the process whereby hippocampus-dependent memories become cortex-dependent.

Supporting Online Material

www.sciencemag.org/cgi/content/full/318/5853/1147/DC1

Methods

SOM Text

Fig. S1

Table S1

References

References and Notes

View Abstract

Navigate This Article