CA1 place cells are considered crucial for spatial memory, but data is limited regarding whether their representations of space evolve over time scales of weeks or more1
. Some theories suggest place cells should retain stable place fields for long-term retention of familiar environments1
. Alternatively, dynamic aspects of place coding may facilitate distinct memory traces of different events occurring in the same environment2
Due to technical limitations, it has been only partially explored if CA1 representations of familiar environments are stable or evolve over time. Electrical recordings from many tens of cells are feasible3
, but it is challenging to record from the same cells longer than a few days. Data on place fields' stability has largely been from small numbers of cells recorded over at most a week4-10
. These studies have demonstrated cells with stable place fields, but the data have been too sparse to assess how coding evolves at the ensemble level.
To study long-term coding dynamics, we combined (): a viral vector (AAV2/5-CamKIIα-GCaMP3) to express the Ca2+
in pyramidal cells; a chronic mouse preparation for time-lapse imaging of CA1 over weeks12
; and a miniaturized (<2 g) microscope for Ca2+
-imaging in hundreds of cells in freely behaving mice13
. We thereby tracked somatic Ca2+
dynamics of 515–1040 pyramidal cells in individual mice as they repeatedly visited a familiar track over 45 days.
Ca2+-imaging in hundreds of place cells in freely behaving mice
We first verified CA1 neurons' place coding attributes as mice explored various arenas. The data revealed up to 740 cells (range: 73–740 cells; n
= 13 mice) undergoing Ca2+
excitation in single fields of view (; Supplementary Fig. 1; and Movie 1
dynamics generally displayed quiescent periods interrupted by prominent transients. This concurs with in vitro
studies showing GCaMP3 reports spike bursts well but for solitary spikes has weak signals easily masked by background fluorescence or noise11
. We used established computational means14
to extract individual cells and their dynamics from each session's Ca2+
-imaging data, without regard to mouse behavior (Online Methods
As expected of place cells, during active exploration many pyramidal cells exhibited Ca2+
-excitation when the mouse occupied a specific portion of its arena (). When we placed mice in two different arenas at the same location in the room but with distinct shape, color and orientation cues, a subset of cells exhibited re-mapping2
, showing spatially distinct patterns of Ca2+
-excitation in the two arenas. Consistent with prior work, some cells had place fields in only one arena. Thus, one can optically detect CA1 place cell activity in freely behaving mice, in accord with a Ca2+
-imaging study in mice exploring a virtual reality15
To study place cells over weeks, we trained mice to run back and forth on a linear track; Ca2+
-imaging occurred on ten sessions over 45 days (). As in prior studies in linear environments16
, many cells had clear place-coding properties that usually depended strongly on the mouse's running direction (). Overall, ~90% of cells had at least one Ca2+
transient during running.
Basic aspects of CA1 place codes are stable over weeks
For detailed analyses we focused on a subset of mice (n
= 4) and used a conservative definition of place field by requiring statistically significant mutual information between a cell's Ca2+
excitation events and the mouse's location17
. With this definition, ~20% of cells had place fields for left, right, or both running directions (). The set of place fields fully covered the track, with the ends represented more densely than the interior (). The mean place field size was ~27% of the 84-cm track, within the range published for mice15,16,18,19
. For each place field we detected Ca2+
activity in 17 ± 14% of running passes (n
= 1656 place fields, mean ± s.d.; range: 2-87%). Across days 5–35 the percentages of cells on each day with place fields for right (12 ± 1%; mean ± s.e.m) or left (12 ± 1%) motion did not vary (Kruskal-Wallis ANOVA; P
= 0.77 and 0.88 for right and left, respectively; n
= 7 sessions; n
= 4 mice) (). Nor were there changes in the distributions of place fields' locations or sizes () (Kolmogorov-Smirnov test; P
=0.06–0.99 for locations and 0.02–0.99 for sizes, both compared to a significance threshold of 2.4·10-3
that includes the Dunn-Sidák correction for the 21 pairwise comparisons). Notably, we saw no discernible changes to cells' morphologies nor substantial changes in mean Ca2+
-transient amplitudes or baseline fluorescence within or across sessions (Supplementary Fig. 2
). Thus, photobleaching was negligible, and neither GCaMP3 expression nor illumination had perceptibly deleterious effects on cell health.
To register repeated observations of individual cells, we first examined the precision of image registration (Supplementary Fig. 3
). Bootstrap analyses showed errors in aligning cells' locations across sessions were <1 μm. This precision more than sufficed, as even the closest cells had ≥6 μm between centroids. Over the full study each mouse yielded 515–1040 cells (n
= 4 mice), more than the maximum (740) seen in one session but consistent with anatomical data.
A majority of cells was active in one or two sessions (57 ± 1%; mean ± s.d.; n = 2960 cells; n = 4 mice). 2.8 ± 0.3% were active in all 10 sessions (). Yet, each session had the same percentage (31 ± 1%) of active cells out of the full tally (Kruskal-Wallis ANOVA; P = 0.46) (
inset). Cells came in and out of this active subset day-by-day, but the overlap in active subsets from any two days was only moderately time-dependent, declining from ~60% for sessions 5 days apart to ~40% for 30 days apart ().
Place fields are spatially invariant and temporally stochastic while preserving a stable representation at the ensemble level
Comparisons between any two sessions revealed ~15–25% overlap in the subsets of cells with statistically significant place fields, declining from ~25% for sessions 5 days apart to ~15% for 30 days (). Notably, when individual cells did show place fields in more than one session, the place fields' locations were generally identical (). This is a compelling, independent validation of our image registration protocol. Though cells came in and out of the place-coding ensemble, place fields' invariant locations plus the slowly declining overlap in place-coding ensembles led to spatial representations that retained a clear resemblance while decaying over time ().
We next sought factors that influenced cells' recurrences in the place-coding ensemble. If cell physiological or coding parameters were key influences, Ca2+
-activity or place-coding parameters might correlate with recurrence probabilities. But if network dynamics were more important, the data might reveal no relationships between cells' characteristics and recurrence probabilities. Notably, the numbers of sessions in which cells had Ca2+
activity or statistically significant place fields were uncorrelated with their rates and amplitudes of Ca2+
activation (Supplementary Fig. 4
). Cells with high place-coding stability in single sessions had virtually the same recurrence odds as other cells (Supplementary Fig. 5
). Neither inclusion of Ca2+
transient amplitudes in the computations of place fields nor variations in how we extracted cells from the raw data altered these findings (Supplementary Figs. 6, 7
). Overall, we failed to find parameters predictive of which cells recur in either the active or place-coding ensembles.
Given place fields' invariant locations, did the ~15–25% overlap between different days' coding ensembles suffice to retain a stable spatial representation? To address this, we used Bayesian decoding techniques to study how well we could reconstruct the mouse's location from the Ca2+
-imaging data (Supplementary Fig. 8
and ). We created a set of decoders of a common mathematical structure, trained each decoder on a portion of one day's data, and tested it on other data. When test and training data were from the same day, estimates of mouse location were excellent (median error nearly always <7 cm) and highly significant compared to shuffled test data (P
; Kolmogorov-Smirnov test). We then asked how well a decoder trained on data from one day would perform on data from other days. In comparisons between decoders using the same number of cells, performance declined only modestly with the interval between training and testing and remained very significant for 30-day intervals (P
; Kolmogorov-Smirnov test). Thus, the ~15% commonality in place-coding subsets across 30 days sufficed to deduce the mouse's trajectory using a decoder trained on data of 30 days prior.
Though GCaMP3 does not faithfully report single spikes11
, our approach can sense isolated spike bursts. To evade analyses of place coding by using only solitary spikes, cells would have to avoid burst spiking across entire sessions while still encoding spatial information. We do not exclude this possibility but consider it unlikely, especially given the key place coding role ascribed to bursts20
and lack of correlation here between cellular Ca2+
activity and involvement in place coding. Improved Ca2+
sensors should reveal a greater portion of spiking activity and could amend our findings with GCaMP3. Our general approach will allow long-term tracking of large neural ensembles in multiple brain areas beyond CA1.
Overall, our data indicate retention of spatial information in CA1 combines stable place field locations with ~15–25% odds an individual cell will recur in the ensemble place code. Prior long-term recordings had stressed place field stability and usually focused on tens or fewer cells, far less than the ~3500 we studied. In vivo
imaging allows reliable tracking of CA1 cells over months12
, here revealing the fluctuating membership of place coding ensembles. This supports prior reports of individually stable place fields but shows CA1 coding has day-to-day dynamism at the cellular level while preserving spatial information in the ~15–25% overlap between coding ensembles from any two days. Conversely, each episode in a familiar arena has a unique signature via the ~75–85% of cells that do not overlap when comparing coding ensembles from any two sessions (). Note that our data show the existence of these non-overlapping signatures but do not imply any functional significance. One possibility is that coding turnover is a long-term form of the spike-rate re-mapping seen over shorter intervals2
and might help distinguish traces of distinct events occurring in the same environment.