Hippocampal synchrony and neocortical desynchrony cooperate to encode and retrieve episodic memories

theta and ATL/ PTPR alpha/beta power (both p fdr > 0.5). These results indicate that hippocampal “slow” gamma synchronisation precedes ATL alpha/beta desynchronisation during the retrieval of episodic memories – a reversal of the dynamic observed during episodic memory formation. Lastly, we examined how the neocortical-hippocampal dynamics differed between encoding and retrieval. To this end, the subsequent memory effect (SME; remembered minus forgotten cross-correlation at encoding) for ATL alpha/beta power and hippocampal “fast” gamma power was contrasted with the retrieval success effect (RSE; remembered minus forgotten cross-correlation at retrieval) for ATL alpha/beta power and hippocampal “slow” gamma power in a random effects, non-parametric, permutation-based t-test. This revealed an interaction whereby ATL desynchrony preceded hippocampal synchrony during encoding (p fdr = 0.040; 100-200ms) but hippocampal synchrony preceded ATL desynchrony during retrieval (p fdr = 0.036; 200-300ms) [see fig. 5 for difference line plot; see supp. fig. 3 for separate hit/miss line plots]. No interaction was observed when analysing the PTPR-hippocampus cross-correlation. These results support those reported in the previous two

magnetic stimulation impairs both episodic memory formation and retrieval, suggesting that alpha/beta desynchronisation plays a causal role in these processes 19,24 .In short, these studies suggest that neocortical alpha/beta desynchronisation underpins the processing of event-related information, allowing for the formation and later recollection of highly detailed episodic memories.
Within the hippocampus, synchronised gamma activity, nested within ongoing theta oscillations, is hypothesised to form a neural code capable of binding event-related information into a cohesive episode 6,7,25,26 .More specifically, each theta cycle provides a window where discrete elements of the event (coded by individual gamma cycles) can be organised, related and maintained (see fig. 1b).Critically, it is hypothesised 3 that gamma cycles lock to the part of the theta cycle optimal for long-term potentiation (LTP), enhancing the synaptic strengthening between neural populations coding for each element of the event and thereby optimising encoding.During later retrieval, the cuing of one element of this theta-gamma code is thought to reactivate the code in its entirety 25 , reinstating the memory.
Studies in both animals [27][28][29] and humans 30,31 support these ideas.Interestingly, two distinct gamma rhythms are thought to couple to theta, with faster gamma rhythms (60-100Hz)   supporting episodic encoding and slower gamma rhythms (30-45Hz) supporting episodic retrieval 7,32,33 .Evidence suggests that "fast" gamma boosts connectivity between CA1 and the entorhinal cortex (allowing information to flow into the hippocampus for representational binding), while "slow" gamma boosts connectivity between CA1 and CA3 (facilitating pattern completion) 34 .In conjunction, these findings and theories would suggest that the synchronisation of theta to "fast" and "slow" gamma rhythms support the hippocampal ability to associate and reactivate discrete elements of an episodic memory.
. CC-BY-NC-ND 4.0 International license available under a not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.It is made The copyright holder for this preprint (which was this version posted April 21, 2018.; https://doi.org/10.1101/305698doi: bioRxiv preprint Hippocampal-neocortical interactions during human episodic memory formation and retrieval  Nine patients, implanted with hippocampal depth electrodes for the treatment for medicationresistant epilepsy, completed a simple associative memory task (see fig. 1c) where they related life-like videos or sounds to words that followed.Following a short distractor, participants attempted to recall the previously presented videos/sounds using the words as cues.Electrophysiological analysis was centred on verbal stimulus presentation at both encoding and retrieval.By keeping external stimulation consistent between encoding and retrieval, any differences in oscillatory dynamics must be driven by internal influences.We conducted these analyses in three ROIs (see fig. 1d): 1) the hippocampus (a hub for representation binding), 2) the anterior temporal lobe (ATL; a hub for semantic-based information processing 36 ), and 3) the posterior temporal/parietal region (PTPR; a hub for retrieval-related attentional processes [37][38][39][40] ).Foreshadowing the results below, we show that ATL alpha/beta desynchronisation precedes hippocampal "fast" gamma synchronisation during successful memory formation, and that hippocampal "slow" gamma synchronisation precedes ATL alpha/beta desynchronisation during successful memory retrieval, revealing the first empirical evidence of an interaction between these two mechanisms during human episodic memory formation and retrieval.

Results
Alpha/beta oscillations dominate the neocortex; theta oscillations dominate the hippocampus We first sought to empirically define the peak frequencies in our three regions of interest.
Broadband spectral power (1-100Hz) was computed across a 1500ms window starting at the onset of the verbal stimulus (at encoding and retrieval).To help identify spectral peaks, the 1/f noise was then subtracted from the data [41][42][43] (see methods for details).Subsequently, the resulting power spectra were collapsed over time and trials, and split into hippocampal and neocortical ROIs.Across participants, a distinct slow-theta peak could be observed in the hippocampus at ~2.5Hz and an alpha/beta peak could be observed in the two neocortical regions between 8-20Hz (see figure 2a).We defined the peak frequency of each ROI for each participant individually and conducted all subsequent analyses on these peak frequencies (see supplementary table 1 for individual peak frequencies).
Hippocampal theta-gamma phase-amplitude coupling correlates with the successful formation of episodic memories We then investigated whether functional coupling between hippocampal theta phase and gamma power can be observed during these same time windows (for power-related effects, see supplementary materials).To test this, we computed the phase-locking value between the hippocampal theta phase and the hippocampal gamma envelope of each participant (separately for "fast" and "slow" gamma bands) 46 .Importantly, if the hippocampus supports episodic binding, we would anticipate that theta-gamma coupling is only prevalent during the presentation of the second (i.e.verbal) stimulus, as both stimuli need to be presented before inter-item binding can occur.When analysing phase-amplitude coupling between theta and "fast" gamma, a significant increase was observed for later remembered, relative to later forgotten, stimuli during the encoding of the verbal stimulus (p = 0.008; see fig 4a-c), but not during the encoding of the dynamic stimulus (p = 0.301), supporting the idea that theta- gamma coupling facilitates inter-item binding.Control analyses revealed no difference in phase-amplitude coupling between remembered and forgotten trials during the encoding pre-stimulus interval (p > 0.5) or during the presentation of the verbal stimulus at retrieval (p > 0.5).When analysing phase-amplitude coupling between theta and "slow" gamma, no difference between remembered and forgotten items was observed in any time window (verbal stimulus at encoding: p > 0.5; dynamic stimulus at encoding: p = 0.277; pre-stimulus interval at encoding: p > 0.5; verbal stimulus at retrieval: p > 0.5).These findings demonstrate that gamma power locks to theta phase during the formation, but not retrieval, of episodic memories.While we anticipated observing similar coupling during retrieval based on earlier accounts 7 , its absence conforms to other claims that theta-gamma coupling is more prevalent during encoding relative to retrieval 6 .
Phase-amplitude coupling is an analytical method susceptible to several confounds 47 .To avoid a cumbersome results section, we have resolved concerns about trial number imbalances, event-related potentials, asymmetric waveforms, power differences between conditions, meaningful phase/power-giving frequencies and bandwidth in the supplementary materials.
Hippocampal theta/gamma synchronisation and neocortical alpha/beta desynchronisation cooperate during the encoding and retrieval of human episodic memories So far, we have demonstrated that both neocortical desynchronisation and hippocampal synchronisation are prevalent during episodic memory processes.Critically however, the synchronisation/desynchronisation framework 3 would predict that, during encoding, these two markers correlate such that the degree of neocortical desynchronisation can predict the degree of hippocampal synchronisation.On a cognitive level, this would signify information processing within the neocortex preceding representational binding in the hippocampus.To test this theory, we cross-correlated the neocortical alpha/beta power time-series with the hippocampal theta and gamma power time-series.This analysis offsets the neocortical timeseries relative to the hippocampal time-series in an attempt to identify at what time lag the two time-series most strongly correlate.The cross-correlation was computed for every trial, and the subsequent memory effect (SME) was calculated by subtracting the mean crosscorrelation across forgotten items from the mean cross-correlation across remembered trials.
By calculating the SME, any correlation between the two time-series that is driven by shared noise (originating from a shared reference) is removed, as this correlation is consistent across remembered and forgotten trials.Furthermore, the SME highlights memory-specific dynamics in neocortical-hippocampal links, rather than general, memory-unspecific connectivity.Relative to later forgotten items, later remembered items showed a significant negative cross-correlation between ATL alpha/beta power and hippocampal "fast" gamma power (pfdr = 0.037; see fig. 5 for difference line plot; see supp.fig. 3 for separate hit/miss line plots).Critically, this cross-correlation suggests that alpha/beta power decreases precede "fast" gamma power increases by approximately 100-200ms.No link was observed between ATL alpha/beta power and hippocampal "slow" gamma power (pfdr = 0.257).Alpha/beta power in the PTPR did not cross-correlate with hippocampal "fast" or "slow" gamma power (pfdr = 0.167 and pfdr > 0.5 respectively).No link was observed between hippocampal theta and ATL/ PTPR alpha/beta power (both pfdr > 0.5).These results indicate that a unique connection exists between the ATL and the hippocampus during episodic memory formation, where ATL alpha/beta desynchronisation precedes hippocampal "fast" gamma synchronisation.
We then investigated whether this relationship reverses (i.e.hippocampal synchronisation precedes neocortical desynchronisation) during episodic memory retrieval.On a cognitive level, this would represent pattern completion in the hippocampus preceding information reinstatement in the neocortex.To test this, we repeated the cross-correlation analysis in the same manner as above for epochs covering the presentation of the retrieval cue and then calculated the retrieval success effect (RSE) by subtracting the mean cross-correlation across forgotten items from the mean cross-correlation across remembered trials.The RSE was calculated for the same reasons as the SME (see above).Relative to forgotten items, remembered items showed a significant negative cross-correlation between ATL alpha/beta power and hippocampal "slow" gamma power (pfdr = 0.048; see fig. 5 for difference line plot; see supp.fig. 3 for separate hit/miss line plots), where an increase in hippocampal gamma power would precede a decrease in ATL alpha/beta power by 200-300ms (see sup. fig. 2 for separate remembered and forgotten cross-correlations).No link was observed between ATL alpha/beta power and hippocampal "fast" gamma power (pfdr > 0.5), nor between PTPR alpha/beta power and hippocampal "fast" or "slow" gamma power (pfdr > 0.5 and pfdr = 0.217 respectively).No link was observed between hippocampal theta and ATL/ PTPR alpha/beta power (both pfdr > 0.5).These results indicate that hippocampal "slow" gamma synchronisation precedes ATL alpha/beta desynchronisation during the retrieval of episodic memories -a reversal of the dynamic observed during episodic memory formation.
Lastly, we examined how the neocortical-hippocampal dynamics differed between encoding and retrieval.To this end, the subsequent memory effect (SME; remembered minus forgotten cross-correlation at encoding) for ATL alpha/beta power and hippocampal "fast" gamma power was contrasted with the retrieval success effect (RSE; remembered minus forgotten cross-correlation at retrieval) for ATL alpha/beta power and hippocampal "slow" gamma power in a random effects, non-

Discussion
To successfully encode and recall episodic memories, we must be capable of 1) representing detailed multisensory information, and 2) binding this information into a coherent episode.Numerous studies have suggested that these two processes are accomplished by neocortical oscillatory desynchronisation and hippocampal oscillatory synchronisation respectively 3,5,7,25 .Here, we provide the first empirical evidence that these two processes co-exist and interact.During successful episodic memory formation, alpha/beta desynchronisation in the anterior temporal lobe (ATL) reliably precedes "fast" hippocampal gamma synchronisation (60-80Hz) by 100-200ms.In contrast, "slow" hippocampal gamma synchronisation (40-50Hz) precedes alpha/beta desynchronisation by 200-300ms during successful episodic memory retrieval.These findings demonstrate that the cooperation between neocortical desynchronisation and hippocampal synchronisation underpins the formation and retrieval of episodic memories.
Our central finding demonstrates that ATL alpha/beta desynchronisation and hippocampal theta/gamma synchronisation cooperate during the formation and retrieval of episodic memories.This result draws together a multitude of conflicting studies, some which indicate that synchronisation benefits memory e.g. 27,30,31and others which indicate that desynchronisation benefits memory e.g. 12,23,48, and provides a possible empirical resolution to the so-called "synchronisation-desynchronisation conundrum" 3 .These findings are in line with previous observations demonstrating that hippocampal gamma synchrony precedes hippocampal alpha desynchrony during associative memory retrieval 49 .However, we are the first to show that this sequence reverses during encoding, and to link these two mechanisms across brain regions (via simultaneous hippocampal-neocortical recordings unavailable to 49 ).We speculate that the delay in hippocampal response relative to ATL alpha/beta desynchronisation during encoding reflects the need for the ATL to process semantic details prior to the hippocampus binding this information into a coherent representation of the event 25,26 .In contrast, we posit that the ATL delay in response relative to hippocampal gamma synchrony during retrieval reflects the need for the hippocampal representational code to be reactivated prior to reinstating highly-detailed semantic information about the event 50 .Intriguingly, the observed delay between alpha and gamma-band activity was notably greater (100-300ms) than what has been reported in a previous rodent study (50-100ms) 51 .Seemingly, our observed interaction does not reflect direct communication between the neocortex and hippocampus, but rather reflects communication via intermediary regions (e.g. the entorhinal cortex).Ultimately, this sequential flow of information would result in a delay in communication between the ATL and hippocampus.Anatomically speaking, this cascade of communication may act upon the "direct intrahippocampal pathway" -a route with reciprocal connections between the ATL and hippocampus via the entorhinal cortex 52,53 (parsimoniously, the absence of connections between the PTPR and hippocampus via this pathway may explain why no similar PTPR-hippocampus crosscorrelation was observed).These anatomical connections would allow the ATL and hippocampus to cooperate during episodic memory formation and retrieval, facilitating the flow of neocortical information into the hippocampus during encoding and the propagation of hippocampal retrieval signals into the neocortex during retrieval.
We also uncovered the first empirical evidence of distinct gamma rhythms supporting human episodic memory formation and retrieval 7,32 .Specifically, we found that "fast" gamma oscillatory activity (60-80Hz) dominates encoding while "slow" gamma oscillatory activity (40-50Hz) dominates retrieval, generalising earlier rodent findings e.g. 34to humans.Critically, as the "fast vs. slow" distinction was only present for successfully recalled items, we posit that this oscillatory differentiation is not driven by changes in task (as the same difference would then be observed for forgotten items), but by differences in the encoding/retrieval processes themselves.Earlier rodent studies have suggested that the distinction between the two gamma bands reflects a difference in CA1 coupling 34 ; "fast" gamma oscillations support CA1-entorhinal cortex coupling, facilitating the transfer of information into the hippocampus, while "slow" gamma oscillations support CA1-CA3 coupling, facilitating the reactivation of stored information.We speculate that these patterns of connectivity extrapolate to humans and explain the observed differences in gamma frequency relating to episodic memory formation and retrieval.
In combination, the cross-correlation and gamma-band analyses produce a detailed picture of information flow during episodic memory formation and retrieval.Based on earlier frameworks 3,7 and models 4 , we postulate that the link between neocortical alpha/beta desynchronisation and hippocampal "fast" gamma synchronisation during memory formation reflects the flow of semantic information (processed in the desynchronised ATL) to entorhinal cortex 26 via the direct intrahippocampal pathway 52,53 , where "fast" gamma synchronicity between the entorhinal cortex and CA1 passes this information onto the hippocampus 34,54 .In contrast, the link between hippocampal "slow" gamma synchronisation and neocortical alpha/beta desynchronisation during memory retrieval reflects the flow of reactivated representational codes from CA3 to CA1 (via "slow" gamma synchronicity 34,54 ), which propagates out into the neocortex 50 via reciprocal connections in the direct intrahippocampal pathway, reinstating semantic details in the desynchronised ATL.However, future research with direct recordings from these hippocampal sub-regions in humans is needed to empirically test this proposed flow of information during episodic memory formation and retrieval.
In addition to these novel findings, we uncovered robust evidence for an increase in theta-"fast" gamma phase-amplitude coupling during the formation of episodic memories.
Research into phase-amplitude coupling has been plagued by physiological 55 and analytical 47 confounds that can produce spurious results.We have attempted to address these concerns (see supplementary materials) in order to present a convincing demonstration of hippocampal theta-gamma phase-amplitude coupling in human episodic memory.Given the presence of phase-amplitude coupling during the encoding of the verbal (but not dynamic) stimulus, this coupling could be interpreted as a mechanism for relational binding.Binding can only occur after more than one element has been presented, so any marker of binding would only be observable during the presentation of the second (i.e.verbal) stimulus.Concerns that this coupling is driven by stimulus modality (i.e.verbal vs. visual/audio) are unfounded as no coupling is observed during presentation of the same verbal stimuli at retrieval.Similarly, concerns that the observed hippocampal coupling reflects the maintenance of the previously presented dynamic stimulus 56,57 are unfounded as no hippocampal coupling was observed during the presentation of the retrieval cue, a period when patients must maintain the recalled dynamic stimulus for later response.Rather, it would appear that the hippocampal theta-gamma phase-amplitude coupling observed during successful memory formation is uniquely linked to the binding of disparate elements into a coherent episodic memory.
Lastly, we found that ATL alpha/beta desynchronisation accompanies the successful formation and retrieval of episodic memories, supporting a wealth of research preceding our findings (for reviews, see 3,5 ).Importantly, while alpha/beta desynchronisation could be reliably predicted by verbal stimulus onset during episodic memory formation, the same was not true for retrieval.We speculate that this is due to greater temporal variability in neocortical desynchrony during retrieval relative to encoding.As retrieval requires hippocampal pattern completion prior to the neocortical information reinstatement (an intermediary step absent for encoding), any temporal variability in the pattern completion process would impair our ability to time-lock neocortical desynchrony to stimulus onset.
Under this assumption however, we would also expect ATL desynchrony to lock to hippocampal activity relating to pattern completion, such as hippocampal gamma activity 33,49 .
Our cross-correlation results support this idea, demonstrating that ATL alpha/beta desynchronisation can be reliably predicted by preceding hippocampal "slow" gamma synchronisation during episodic memory retrieval.In short, these findings suggest that ATL alpha/beta desynchronisation accompanying episodic memory encoding and retrieval may time-lock to different events.
Three questions remain however: First, why does neocortical alpha/beta desynchronisation cross-correlate with hippocampal gamma synchronisation, but not hippocampal theta synchronisation?We argue that it is not theta power (our metric of synchronisation) but theta phase that is most important for memory formation and retrieval.This stance conforms to several existing theories of episodic memory that posit that hippocampal theta phase, not power, facilitates representational binding and organisation 6,7,25 .
Second, do similar bi-directional streams of information flow exist between the hippocampus and other neocortical regions?As it was not medically necessary, electrode coverage did not expand to every neocortical region linked to episodic memory.Therefore, we could not test this theory.We speculate, however, that similar bi-directional links do exist.For example, hippocampal gamma synchronisation may co-ordinate with alpha/beta desynchronisation in the visual cortex to facilitate the encoding and retrieval of visual memories 19 .Speculating further, hippocampal gamma synchronisation may be the metaphorical spark that lights the fuse of memory replay, coded in desynchronised neocortical alpha phase patterns 18 .
Third, why is human hippocampal theta (~2.5Hz) "slow" in comparison to rodent hippocampal theta (~8Hz), but similar "slowing" is not observed within the gamma band?
The notion that human hippocampal theta is slower than that of rodents is not controversial 58,59 .It has been proposed that the "slowing" of hippocampal theta oscillations is beneficial as it compensates for conduction delays within the hippocampus (a by-product of massive brain scaling resulting from evolution).In contrast, the "slowing" of hippocampal gamma oscillations may be detrimental for learning.Speculatively, a slower gamma rhythm would mean that interacting neurons no longer fire at a rate optimal for spike-timing dependent plasticity (STDP; a form of long-term potentiation), limiting synaptic strengthening.In support of these ideas, empirical evidence suggests that theta rhythms "slow" to a much greater degree than gamma rhythms as brain size increases (see fig. 2B in 59 ).In short, theta rhythms "slow" to facilitate communication over greater distances, while gamma rhythms remain consistent to preserve STDP.
In summary, we deliver the first empirical evidence that neocortical desynchronisation and hippocampal synchronisation cooperate during the formation and retrieval of episodic memories, providing evidence that may help resolve the so-called "synchronisationdesynchronisation conundrum" 3 .Furthermore, we provide the first evidence that distinct describing the curvature of the 1/f characteristic.The 1/f curve (Ax) was then subtracted from the logtransformed power-spectrum (B).This approach removes the 1/f curve while retaining oscillatory peaks in the power spectrum (see fig. 2A for 1/f-corrected power-spectra).

Peak frequency analysis
To identify dominant frequencies within the hippocampus and neocortex, the raw signal recorded at every contact for each epoch was convolved with a 5-cycle wavelet (0 to 1500ms post-stimulus [padded with real data for lower frequencies], in steps of 25ms; 1Hz to 100Hz, in steps of 0.5Hz).The 1/f noise was subtracted using the method described above to help pronounce the peaks in the power-spectrum.The data was then smoothed using a Gaussian kernel (200ms; 1Hz) to attenuate inter-and intra-individual differences in spectral responses 61,62 and to help approximate normally distributed data (an assumption frequently violated in small samples).The data was averaged across all time-points, trials and contacts (separately for the hippocampus, ATL and PTPR).Peaks of 1/f corrected absolute power were then visually identified for each participant.To allow group comparisons (e.g. in figure 2), the power-spectrum of each participant was z-transformed using the mean and standard deviation across trial/contact/time-averaged frequencies prior to plotting.
To identify the memory-related difference in the dominant gamma bands, the power spectra for "remembered" trials were calculated in an identical manner, except that the Gaussian kernel was expanded to account for the greater variability of high-frequency oscillatory responses (200ms, 5Hz).
The absolute power for the averaged retrieval epochs was subtracted from the absolute power for the averaged encoding epochs and the encoding-related/retrieval-related gamma peaks were visually identified for each participant.
To quantify the difference in "fast"/"slow" gamma, the power-spectra for encoding and retrieval were collapsed in seven 10Hz bins ranging from 30Hz to 100Hz and then contrasted in a random effects, non-parametric permutation-based t-test (5000 randomisations; for details, see Maris and Oostenveld,   2007).The multiple comparison issue was solved using the false-discovery rate correction 64 .To confirm that this difference was related to memory as opposed to a difference in task, this analysis was repeated using the "forgotten" trials.

Spectral power analysis
For all spectral power analyses (i.e.encoding and retrieval epochs), the time-series data was convolved with 5-cycle wavelets (from 0ms to 1500ms post-stimulus, in steps of 25ms; 1Hz to 100Hz, in steps of 0.5Hz).The 1/f noise was subtracted using the method described above to attenuate condition-related differences in spectral noise 44 .The data was then concatenated across trials and ztransformed using the mean and standard deviation of power across time for each channel-frequency pair.Subsequently, the data was smoothed using a Gaussian kernel (200ms; 1Hz), to attenuate intraand inter-individual variability in spectral and temporal responses to the stimuli 61,62 , and to help approximate normally distributed data (an assumption frequently violated in small samples).The timefrequency resolved data was then averaged over channels to provide a time-series for each "ROI" x "peak frequency band" pair.
For statistical analysis, trials were split into two groups based on whether the stimuli were remembered or forgotten.Then, the time-series were collapsed into seven time bins of 200ms ranging from stimulus onset to 1400ms after onset.The two conditions were then contrasted in a random effects, non-parametric permutation test (5000 randomisations; for details, see Maris and Oostenveld,   2007).The multiple comparison issue was solved using the false-discovery rate correction 64 .

Phase-amplitude coupling analysis
Theta-gamma phase amplitude coupling within the hippocampus was assessed for every trial using phase-locking value 46 .The consistency between the angles of the peak theta phase (±0.5Hz) and the peak gamma (±5Hz) power envelope was computed for windows of 1000ms (reducing concerns of non-stationarity introduced by long epochs), with an overlap of 750ms between bins, for each individual trial (for further details, see 46 ).By calculating phase-locking value on a single-trial level, issues with trial number imbalances between conditions can be side-stepped as each phase-locking value is calculated on the same number of samples.That said, single-trial analysis could mean gamma is locked to different phases of the theta cycle on each trial, and therefore would not serve any mechanistic purpose.To alleviate these concerns, we tested for non-uniformity in gamma-powerto-theta-phase-locking (i.e.phase preference), and have included individual polar plots detailing the distribution of the preferred phase for gamma-locking (see figure 4d  ).The first 500ms after stimulus onset was excluded to alleviate concerns that any observed phase-amplitude coupling may be driven by non-stationarity induced by the event-related potential.The observed phase-locking values for each contact were averaged to provide a single measure of phase-amplitude coupling in the hippocampus per trial.For statistical analysis, trials were split based on whether they were later remembered or later forgotten and contrasted in a random effects, non-parametric permutation-based t-test (5000 randomisations; for details, see Maris and Oostenveld, 2007).

Cross-correlation analysis
For all cross-correlation analyses (i.e.encoding and retrieval epochs), the time-series data was convolved with 5-cycle wavelets (from 0ms to 1500ms post-stimulus, in steps of 10ms; 1Hz to 100Hz, in steps of 0.5Hz).The 1/f noise was subtracted using the method described above to attenuate condition-related differences in spectral noise 44 .The data was then concatenated across trials and was z-transformed using the mean and standard deviation of power across time for each channelfrequency pair.Subsequently, the data was smoothed using a Gaussian kernel (50ms; 1Hz), to attenuate intra-and inter-individual variability in spectral and temporal responses to the stimuli 61,62 , and to help approximate normally distributed data (an assumption frequently violated in small samples).The smoothing kernel used in earlier analyses covered 200ms, but here it has been reduced to 50ms as excessive smoothing would obscure the temporal dynamics of the neocorticalhippocampal cross-correlation.For each "trial x channel combination" pair, the cross-correlation between the hippocampus and the ATL, and the cross-correlation between the hippocampus and PTPR, was computed using the Matlab function crosscorr() with a lag of 300ms (meaning the correlation between hippocampus and neocortex was considered for every offset from where the neocortex preceded the hippocampus by 300ms to where the neocortex lagged behind the hippocampus by 300ms).This returned a time-series of Pearson correlation values describing the relationship between hippocampus and neocortex at all considered lags.These correlation values were then averaged over channels and split into two groups: remembered and forgotten.These two groups were individually averaged over trials for each participant, collapsed into bins of 100ms, and then contrasted in a random effects, non-parametric permutation-based t-test (5000 randomisations; for details, see Maris and Oostenveld, 2007).The multiple comparison issue was solved using the false-discovery rate correction 64 .We term the "remembered > forgotten" difference in crosscorrelation for encoding data "the subsequent memory cross-correlation" and the difference for retrieval data "the retrieval success cross-correlation".
To test the "encoding-retrieval" x "lag-lead" interaction, we contrasted the subsequent memory crosscorrelation with the retrieval success in a random effects, non-parametric permutation-based t-test (5000 randomisations; for details, see Maris and Oostenveld, 2007).The multiple comparison issue was solved using the false-discovery rate correction 64 .

Figure 1 .
Figure 1.The sync-desync framework.(a) this framework explains the encoding (left) and retrieval (right) of associative episodic memories.Incoming stimuli are independently processed by relevant sensory regions of the neocortex.These neocortical representations are then passed onto the hippocampus where they are bound together.At a later stage, a partial cue reactivates the hippocampal associative link, which in turn reactivates neocortical patterns coding for the memory representation, giving rise to conscious recollection.(b) a pictographic representation of hypothesised oscillatory dynamics.Reduced oscillatory synchronisation (blue line) within the neocortex allows individual neurons (blue dots) to fire more freely and create a more flexible neural code.Gamma cycles lock to a part of the theta phase (red line) optimal for long-term potentiation (LTP), allowing individual elements to be organised and bound in relation to one another.(c) the trial outline for encoding (left) and retrieval (right) tasks.During encoding, participants are tasked with forming an associative link between a life-like dynamic stimulus (either a video or sound) and a subsequent verbal stimulus.During retrieval, participants are presented with verbal stimuli from the previous encoding block and asked to retrieve the associated dynamic stimulus.Electrophysiological analysis was conducted during the presentation of the verbal stimulus at encoding and retrieval (blue outline).(d) electrode coverage of the anterior temporal lobe (blue), posterior temporal/parietal regions (green) and hippocampus (red).

.
CC-BY-NC-ND 4.0 International license available under a not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.It is made The copyright holder for this preprint (which was this version posted April 21, 2018.; https://doi.org/10.1101/305698doi: bioRxiv preprint Hippocampal-neocortical interactions during human episodic memory formation and retrieval 5

Figure 2 .
Figure 2. Dominant frequencies in neocortex and hippocampus.(a) the mean 1/f corrected power spectrum (with shaded standard error of mean) across all encoding and retrieval trials reveals a theta peak (~2.5Hz) in the hippocampus and an alpha/beta peak (8-20Hz) peak in the two neocortical ROIs.(b) the difference in gamma power between encoding and retrieval reveals a peak in encodingrelated, "fast" gamma at 60-80Hz and a peak in retrievalrelated, "slow" gamma at 40-50Hz (grey line pfdr < 0.05).(c) raw signal during encoding (top) and retrieval (bottom) from a hippocampal contact of "patient 1".

Figure 3 .
Figure 3.The difference in neocortical desynchronisation between later remembered and later forgotten items during episodic memory formation (top) and retrieval (bottom) in two regions of interest: the anterior temporal lobe (left) and posterior temporal/parietal region (right) [red line, pfdr < 0.05].

Figure 4 .
Figure 4. Hippocampal phase-amplitude coupling during episodic memory formation and retrieval.(a) the difference in phase-amplitude coupling between hits and misses for theta locked to encoding-related gamma (red) and retrieval-related gamma (blue) [*p < 0.05].(b) a co-modulogram showing the difference in phaseamplitude coupling between hits and misses as a function of theta phase and encoding-related, "fast" gamma power during the presentation of the verbal stimulus at encoding.(c) a hippocampal recording from "patient 1"depicting an increase in gamma amplitude (red; filtered with band-pass FIR at 74±5 Hz) during the trough of the ongoing theta cycle (grey; filtered with band-pass FIR at 2.5±0.5 Hz).(d) preferred theta phase for gamma power locking in two patients (proportion of trials; hits in red; misses in grey).

Figure 5 .
Figure 5. Cross correlation between the hippocampal gamma power and neocortical alpha/beta power (anterior temporal lobe in blue; posterior temporal/parietal region in green).A negative lag indicates that neocortical power fluctuations precede hippocampal fluctuations; a positive lag indicates the reverse.During encoding (left), ATL power decreases precede hippocampal "fast" gamma power increases.During retrieval (middle), ATL power decreases follow hippocampal "slow" gamma power increases.The contrast of activity between encoding and retrieval (right) confirms this interaction [red line, pfdr < 0.05].
figure 4 for all patients).Across subjects, a Rayleigh test revealed that later remembered items demonstrated a significant deviation from uniformity (z = 3.714, p = 0.018), suggesting gamma power peaked at a preferred phase of the theta cycle during successful memory formation.No deviation from uniformity was observed for later forgotten items (z = 1.354, p = 0.267).Phase-amplitude coupling was computed for four time windows (word at encoding [500ms to 2500ms], dynamic stimulus at encoding [500ms to 2500ms], word at retrieval [500ms to 2500ms], and encoding prestimulus interval [-2000ms to 0ms]).The first 500ms after stimulus onset was excluded to alleviate