Supplementary Materials

Order of supplementary materials

  • Materials and Methods
  • Tables
  • Figures
  • References
  • Movies

Original abstract

Sequential firing of hippocampal neurons during both running and rest is believed to contribute to episodic-memory encoding. In particular, sharp-wave (SPW) sequences, which normally occur during rest, have been suggested to serve as a physiological substrate of memory. This hypothesis arose because (i) elimination of SPW sequences impairs learning and memory and (ii) the similarity of some SPW sequences to average running sequences (such as place-cell and episode-cell sequences) creates the impression that SPWs “replay” an animal’s experience of running. Using these average running sequences as templates, it has been shown that SPWs can replay the running sequences in both forward and backward directions relative to the templates. This led to the conjecture that SPW sequences are the sequences that are activated bidirectionally (i.e., forward and backward).

We used a novel method to test this bidirectionality conjecture by directly comparing pairs of SPW sequences without the use of average running sequences. Assuming that SPW sequences can be activated in both forward and backward directions, correlations among SPW sequences should be both positive and negative. Surprisingly, our analysis of correlations among SPW sequences revealed a very significant number of positive correlations but only a chance-level number of negative correlations. This lack of negative correlations among SPW sequences suggests that SPW sequences are activated unidirectionally, not bidirectionally as previously conjectured. This same method was also robust enough to reproduce the seemingly contradictory findings that SPW sequences are positively and negatively correlated with running sequences. More than suggesting that backward SPW replay does not exist, this analysis questions the entire “replay” framework since SPW sequences are statistically correlated positively with each other regardless of their similarity to running sequences.


All procedures were approved by the Janelia Farm Research Campus Institutional Animal Care and Use Committee. All data used in this study were previously used in Wang et al., 2015. All experimental procedures were described in detail in Wang et al., 2015. Here we describe data analysis methods specific for this work. No statistical methods were used to predetermine sample sizes, but our sample sizes are similar to those reported in previous publications. We used non-parametric statistical methods to reduce the effect of the small sample size.

Sequence detection

A sequence was defined to be the spike train generated by a collection of neurons within a time window. We used the spikes of all pyramidal cells that were identified during spike sorting. For this analysis, we had three main sequence types: arm-run sequences, wheel-run sequences, and SPW sequences.

An arm-run sequence contains all of the neuronal spiking that occurred while an animal was traveling between the running wheel and the opposite end of a maze arm. The start (resp., end) time of an arm run was determined to be the moment when the animal crossed into (resp., out of) the arm. Similarly, a wheel-run sequence contains all of the neuronal spiking that occurred while an animal’s was running in the wheel. The specific start and end times of a wheel run were determined by thresholding the wheel’s angular speed. A wheel run was only considered if its duration was long enough (approximately 8–12 seconds) to open the mechanical doors separating the delay area from the track’s arms.

SPW events were identified in the local-field potential (LFP) based on their signature shape: positive-going wave in the deep CA1 pyramidal layer and negative-going wave in the superficial CA1 pyramidal layer (Fig. 1B). Specifically, we subtracted the raw LFP traces from the deep and superficial parts of the CA1 layer. Then we calculated the average LFP in a 0.5 sec time window around each time point and subtracted the running average from the signal. We smoothed the resulting signal with a Gaussian kernel (SD = 5ms) and computed z-score. We dete