Extended Data Fig. 5: Firing sequences in simultaneously recorded time-cells are similar to the population time-cell sequences pooled across all days; and decoding elapsed-time from time cells.

(a) Temporal tuning curves for all the significant time-cells, pooled across all experimental days and bats (the 3 panels correspond to the 3 different locations in the room: balls A, B and Start). These plots are identical to those shown in main Fig. 1g, and were plotted here again to facilitate comparison with panel b. (b-d) Simultaneously-recorded time-cells. (b) Three examples of internally-generated temporal sequences, for ensembles of neurons that were recorded simultaneously: These examples depict similar sequences (with a similar time-span) to the population in panel a. These 3 ensembles were recorded on 3 different recording-days, in the 3 different locations in the room (balls A, B, Start). We could not obtain larger numbers of simultaneous neurons because of the limited number of tetrodes in this study (n = 4 tetrodes; we obtained up to 12 simultaneously recorded significant time-cells per day). (c) All the days × locations (for all bats) in which we had ≥ 2 simultaneously recorded time-cells (n = 57 days × locations). The 3 panels represent the 3 locations in the room. x-axis, preferred time for each neuron (circles); y-axis, experimental day; horizontal lines in each panel represent groups of simultaneously recorded time-cells. Experimental days are sorted according to the total span of preferred-times for the time-cells recorded on that day. Green: the 3 examples in panel b of internally-generated firing sequences. The red numbers on the right indicate the identity of the bat (no. 1–4) from which the cells were recorded. (d) Distributions of time-differences (∆T) between the preferred-times for all the cell-pairs recorded simultaneously on the same day (gray bars; n = 151, 109 and 23 cell-pairs on landing balls A, B and Start respectively), and all the cell-pairs recorded on different days (black lines; n = 12800, 9288 and 3614 cell-pairs on landing balls A,B and Start respectively), plotted separately for the 3 locations in the room. The gray and black distributions were statistically indistinguishable (two-sided Kolmogorov-Smirnov tests: P = 0.126, P = 0.128 and P = 0.208, for balls A, B and Start, respectively). This demonstrates that the pooled sequences (main Fig. 1g) are reliably representing the within-day sequences – indicating that time-cells in the bat hippocampus form internally-generated firing sequences. (e) Bayesian maximum-likelihood decoding of elapsed time. Left panel: Confusion matrix showing the decoded time (y-axis) versus the actual elapsed time (x-axis), using all the time cells, in all three locations. The probabilities in each time-bin were divided by the uniform chance probability. Right panel: Temporal decoding error for each time bin (200-ms bins were used here), computed between 0–8 s, for three cell groups: red line, all the time-cells (n = 274 cells × positions); peach line, contextual time-cells only (cells that were time-tuned in only one location; n = 123 cells × positions); blue line, pure time-cells only (cells that were time-tuned on both A and B; n = 88 cells × positions). Note the temporal decoding error was < 0.6 s for all the time bins up to 8 s – indicating that these neurons carry robust information about elapsed time, up to 8 s after landing. (f) Cross-decoding of elapsed time: For each trial we trained a decoder on responses at the other location. Only pure time-cells with preferred-time difference of ΔT ≤ 1 s between locations were used to train the decoder. The confusion matrix shows the decoded time (y-axis) versus the actual elapsed time (x-axis); the decoded probabilities in each time-bin were divided by the uniform chance probability. (g) Bayesian maximum-likelihood decoding of the origin of flight history – namely, decoding from where did the bat fly to the Start ball – this decoding was performed based on the firing of time-cells when the bat was on the Start ball. Left panel: the identity of the previous landing ball (ball A or B) can be decoded (classified) above chance level during the first ~4 seconds after landing on the Start ball. To assess the statistical significance of decoding in each time bin, we compared the observed classification accuracy to a shuffle test where we randomly permuted the true identities of balls A and B from which the bat flew. We repeated the shuffling 1,000 times and calculated the classification accuracy for each of the 1,000 shuffle-repeats (permutations) in each time bin. Asterisks denote time bins in which the empirically-observed classification accuracy showed significance at 95% [two-sided] compared to the distribution of classification accuracy of the shuffle tests (the observed classification accuracy was higher than the accuracy of 997.5 of the shuffles – Bonferroni-corrected for multiple comparisons for the number of time bins; P < 0.0025). Right panel: the number of time-cells, in each time bin, which showed significant difference in their firing-rate between trials when the bat flew from ball A to the Start ball versus from ball B to the Start ball. These results support the notion that time-cells encode relevant behavioral information. (h) Violin plots showing the distributions of peak firing-rates for pure time-cells, contextual time-cells, and non-time cells (n = 151, 123, and 603 cells × positions, respectively). Dots, individual neurons (cells × positions); red circles, median for each cell group. Peak firing-rate plotted in this panel is the peak of the temporal response (temporal tuning-curve).