Oscillations support short latency co-firing of neurons during human episodic memory formation

  1. Frédéric Roux
  2. George Parish
  3. Ramesh Chelvarajah
  4. David T Rollings
  5. Vijay Sawlani
  6. Hajo Hamer
  7. Stephanie Gollwitzer
  8. Gernot Kreiselmeyer
  9. Marije J ter Wal
  10. Luca Kolibius
  11. Bernhard P Staresina
  12. Maria Wimber
  13. Matthew W Self
  14. Simon Hanslmayr

  • Curated by eLife

    eLife logo

    Evaluation Summary:

    Roux and colleagues measured spiking activity and local field potentials predominantly from the hippocampus and also a few surrounding structures in the medial temporal lobe from patients with pharmacologically intractable epilepsy while the patients performed a cued-recall task. They report differences in local spike-field coherence measurements between hits and misses in the gamma frequency band and differences in both local and distal spike-field coherence measurements between hits and misses in the theta frequency band. The authors further report differences in the timing of spikes between pairs of neurons, with hits correlated with putative downstream neurons firing about 30 ms after putative upstream neurons and misses correlated with delays of about 60 ms. Overall, these are interesting observations that provide intriguing data to further think about how neurons in the medial temporal lobe correlate with recognition memory.

    (This preprint has been reviewed by eLife. We include the public reviews from the reviewers here; the authors also receive private feedback with suggested changes to the manuscript. Reviewer #1 and Reviewer #2 agreed to share their names with the authors.)

Reviewed by

Read the full article See related articles

Listed in

Log in to save this article

Abstract

Theta and gamma oscillations in the medial temporal lobe are suggested to play a critical role for human memory formation via establishing synchrony in neural assemblies. Arguably, such synchrony facilitates efficient information transfer between neurons and enhances synaptic plasticity, both of which benefit episodic memory formation. However, to date little evidence exists from humans that would provide direct evidence for such a specific role of theta and gamma oscillations for episodic memory formation. Here, we investigate how oscillations shape the temporal structure of neural firing during memory formation in the medial temporal lobe. We measured neural firing and local field potentials in human epilepsy patients via micro-wire electrode recordings to analyze whether brain oscillations are related to co-incidences of firing between neurons during successful and unsuccessful encoding of episodic memories. The results show that phase-coupling of neurons to faster theta and gamma oscillations correlates with co-firing at short latencies (~20–30 ms) and occurs during successful memory formation. Phase-coupling at slower oscillations in these same frequency bands, in contrast, correlates with longer co-firing latencies and occurs during memory failure. Thus, our findings suggest that neural oscillations play a role for the synchronization of neural firing in the medial temporal lobe during the encoding of episodic memories.

Article activity feed

  1. Version published to 10.7554/elife.78109 on eLife
  2. eLife

    Author Response

    Reviewer #1 (Public Review):

    Figures 2 through 6. There is no description of the relationship between the findings and the anatomical location of the electrodes (other than distal versus local). Perhaps the non-uniform distribution of electrodes makes these analyses more complicated and such questions might have minimal if any statistical power. But how should we think about the claims in Figures 2-6 in relationship to the hippocampus, amygdala, entorhinal cortex, and parahippocampal gyrus? As one example question out of many, is Figure 2C revealing results for local pairs in all medial temporal lobe areas or any one area in particular? I won't spell out every single anatomical question. But essentially every figure is associated with an anatomical question that is not described in the results.

    To address the reviewer’s point we now report the distribution of spike-LFP pairs across anatomical regions for each Figure 2-6. The results split by anatomical regions are reported in Figure 2 – figure supplement 7, Figure 3 – figure supplement 7, Figure 4 – figure supplement 1, Figure 5 – figure supplement 2, and Figure 6 – figure supplement 3. We also calculated a non-parametric Kruskal-Wallis Test to statistically examine the effect of anatomical regions on the results shown in each figure. Generally, these new results show that the effects are similar across regions, apart from two exceptions (i.e. Figure 4 – supplement 1; and Figure 5 – supplement 2). However, we would like to stress that these results should be taken with a huge grain of salt because the electrodes were not evenly distributed across regions (i.e. ~75% of observations pertain to the hippocampus), and patients as the reviewer correctly points out. This leads to sometimes very low numbers of observations per region and it is difficult to disentangle whether any apparent differences are driven by regional differences, or differences between patients. Detailed results are reported below.

    Manuscript lines 207-212: “In the above analysis all MTL regions were pooled together to allow for sufficient statistical power. Results separated by anatomical region are reported in Figure 2 – figure supplement 7 for the interested reader. However, these results should be interpreted with caution because electrodes were not evenly distributed across regions and patients making it difficult to disentangle whether any apparent differences are driven by actual anatomical differences, or idiosyncratic differences between patients.”

    Manuscript lines 255-258: “Finally, we report the distal spike-LFP results separated by anatomical region in Figure 3 – figure supplement 7, which did not reveal any apparent differences in the memory related modulation of theta spike-LFP coupling between regions.”

    Manuscript lines 264-266: “PSI results separated by anatomical regions are reported in Figure 4 – figure supplement 1, which revealed that the PSI results were mostly driven by within regional coupling.”

    Manuscript lines 399-303: “We also analyzed whether the memory-dependent effects of cross-frequency coupling differ between anatomical regions (see Figure 5 – figure supplement 2). This analysis revealed that the results were mostly driven by the hippocampus, however we urge caution in interpreting this effect due to the large sampling imbalance across regions.”

    Manuscript lines 343-346: “As for the above analysis we also investigated any apparent differences in co-firing between anatomical regions. These results are reported in Figure 6 – figure supplement 3 and show that the earlier co-firing for hits compared to misses was approximately equivalent across regions.”

    Figure 1

    1A. I assume that image positions are randomized during a cued recall?

    Yes, that was the case. We now added that information in the methods section.

    Manuscript lines 526: “Image positions on the screen were randomized for each trial.”

    What was the correlation between subjects' indication of how many images they thought they remembered and their actual performance?

    We did not log how many images the patients thought they remembered. Specifically, if the patients answered that they remembered at least one image, then they were shown the selection screen where they could select the appropriate images. Therefore, we cannot perform this analysis. We report this now in the methods section. However, albeit interesting, the results of such an analysis would not affect the main conclusions of our manuscript.

    Manuscript lines 523-524: “The experimental script did not log how many images the patient indicated that they thought to remember.”

    1B. Chance is shown for hits but not misses. I assume that hits are defined as both images correct and misses as either 0 or 1 image correct. Then a chance for misses is 1-chance for hits = 5/6. It would be nice to mark this in the figure.

    Done as suggested (see Figure 1).

    The authors report that both incorrect was 11.9%. By chance, both incorrect should be the same as both correct, hence also 1/6 probability, hence the probability of both incorrect seems quite close to chance levels, right?

    Yes, that is correct, however, across sessions the proportion of full misses (i.e. both incorrect) was significantly below chance (t(39)=-1.9214; p<0.05). Nevertheless, the proportion of fully forgotten trials appears to be higher than expected purely by chance. This is likely driven by a tendency of participants to either fully remember an episode, or completely forget it, as demonstrated previously in behavioural work (Joensen et al., 2020; JEP Gen.). We report this now in the manuscript.

    Manuscript lines 132-136: “Across sessions the proportion of full misses (i.e. both incorrect) was significantly below chance (t39=-1.92; p<0.05). However, the proportion of fully forgotten trials appears to be higher than expected purely by chance. This is likely driven by a tendency of participants to either fully remember an episode, or completely forget it, as demonstrated previously in behavioral work (25).”

    1C. How does the number of electrodes relate to the number of units recorded in each area?

    The distribution of neurons per region is shown in the new Figure 1D (see above). It approximately matches the distribution of electrodes per region, except for the Amygdala where slightly more neurons where recorded. This is because of one patient (P08) who had two electrodes in the left and right Amygdala and who contributed at lot of sessions (i.e. 9 sessions, comparing to an average of 4.44 per patient).

    Line 152. The authors state that neural firing during encoding was not modulated by memory for the time window of interest. This is slightly surprising given that other studies have shown a correlation between firing rates and memory performance (see Zheng et al Nature Neuroscience 2022 for a recent example). The task here is different from those in other studies, but is there any speculation as to potential differences? What makes firing rates during encoding correlate with subsequent memory in one task and not in another? And why is the interval from 2-3 seconds more interesting than the intervals after 3 seconds where the authors do report changes in firing rates associated with subsequent performance? Is there any reason to think that the interval from 2-3 seconds is where memories are encoded as opposed to the interval after 3 seconds?

    Zheng et al. used a movie-based memory paradigm where they manipulated transitions between scenes to identify event cells and boundary cells. They show that boundary cells, which made up 7.24% of all recorded MTL cells, but not event cells (6.2% of all MTL cells) modulate their firing rate around an event depending on later memory. There are quite a few differences between Zheng et al’s study and our study that need to be considered. Most importantly, we did not perform a complex movie-based memory paradigm as in Zheng et al. and therefore cannot identify boundary cells, which would be expected to show the memory dependent firing rate modulation. This alone could contribute to the fact that no significant differences in firing rates in the first second following stimulus onset were observed. Such an absence of a difference of neural firing depending on later memory is not unprecedented. In their seminal paper, Rutishauser et al. (2010; Nature) report no significant differences in firing rates (0-1 seconds after stimulus onset, which is similar to our 2-3 seconds time window) between later remembered or later forgotten images. This finding is also in line to Jutras & Buffalo (2009; J Neurosci) who also show no significant difference in firing rates of hippocampal neurons during encoding of remembered and forgotten images.

    The 2-3 seconds time interval, which corresponds to 0-1 seconds after the onset of the two associate images, is special because it marks the earliest time point where memory formation can start, therefore allowing us to investigate these very early neural processes that set the stage for later memory-forming processes. While speculative, these early processes likely capture the initial sweep of information transfer into the MTL memory system which arguably is reflected in the timing of spikes relative to LFPs. It is conceivable that these initial network dynamics reflect attentional processes, which act as a gate keeper to the hippocampus (Moscovitch, 2008; Can J Exp Psychol) and thereby set the stage for later memory forming processes. This interpretation would be consistent with studies in macaques showing that attention increases spike-LFP coupling, whilst not affecting firing rates (Fries et al., 2004; Science). We modified the discussion section to address this issue.

    Manuscript lines 468-474: “Interestingly, these early modulations of neural synchronization by memory encoding were observed in the absence of modulations of firing rates, which is consistent with previous results in humans (16) and macaques (12), but contrasts with (43). Studies in macaques showed that attention increases spike-LFP coupling whilst not affecting firing rates (44). It is therefore conceivable that these initial network dynamics reflect attentional processes, which act as a gate keeper to the hippocampus and thereby set the stage for later memory forming processes (45).”

    Lines 154-157 and relationship to the subsequent analyses. These lines mention in passing differences in power in low-frequency bands and high-frequency bands. To what extent are subsequent results (especially Figures 3 and 4) related to this observation? That is, are the changes in spike-field coherence, correlated with, or perhaps even dictated by, the changes in power in the corresponding frequency bands?

    To address this question we repeated the analysis that we performed for SFC for Power in those channels whose LFP was locally coupled to spikes in gamma, and distally coupled to spikes in theta. Furthermore, we correlated the difference in peak frequency between hits and misses between Power and SFC. If power would dictate the effects seen in SFC then we would expect similar effects of memory in power as in SFC, that is an increase of peak frequency for hits compared to misses for gamma and theta. Furthermore, we would expect to find a correlation between the peak frequency differences in power and SFC. None of these scenarios were confirmed by the data. These results are now reported in Figure 2 – figure supplement 5 for gamma, and Figure 3 – figure supplement 5 for theta.

    Manuscript lines 195-199: “We also tested whether a similar shift in peak gamma frequency as observed for spike-LFP coupling is present in LFP power, and whether memory-related differences in peak gamma spike-LFP are correlated with differences in peak gamma power (Figure 2 – figure supplement 5). Both analyses showed no effects, suggesting that the effects in spike-LFP coupling were not coupled to, or driven by similar changes in LFP power.”

    Manuscript lines 248-253: “As for gamma, we also tested whether a similar shift in peak theta frequency is present in LFP power, and whether there is a correlation between the memory-related differences in peak theta spike-LFP and peak theta power (Figure 3 – figure supplement 5). Both analyses showed no effects, suggesting that the effects in spike-LFP coupling were not coupled to, or driven by similar changes in LFP power.”

    Do local interactions include spike-field coherence measurements from the same microwire (i.e., spikes and LFPs from the same microwire)?

    Yes, they do. Out of the 53 local spike-SFC couplings found for the gamma frequency range, 11 (20.75%) were from pairs where the spikes and LFPs were measured on the same microwire. We assume that the reviewer is asking this question because of a concern that spike interpolation may introduce artifacts which may influence the spectrograms and consequently the spike-LFP coupling measures. This was also pointed out by Reviewer #2. To address this concern, we split the data based on whether the spike and LFP providing channels were the same or different. The results show that (i) the spectrogram of SFC is highly similar between the two datasets, with a prominent gamma peak present in both and no significant differences between the two; (ii) restricting the analysis to those data where the LFP and spike providing channels are different replicated the main finding of faster gamma peak frequencies for hits compared to misses; and (iii) limiting the SFC analysis further to only ‘silent’ channels, i.e. channels where no SUA/MUA activity was present at all also replicated the main finding of faster gamma peak frequencies for hits compared to misses.

    These analyses suggest that the SFC results were not driven by spike interpolation artefacts.

    Manuscript lines 199-203: “To rule out concerns about possible artifacts introduced by spike interpolation we repeated the above analysis for spike-LFP pairs where the spike and LFP providing channels are the same or different, and for ‘silent’ LFP channels (i.e. channels were no SUA/MUA activity was detected (see Figure 2 – figure supplement 6). “

    60 Hz. It has always troubled me deeply when results peak at 60 Hz. This is seen in multiple places in the manuscript; e.g., Figures 2B, 2E. What are the odds that engineers choosing the frequency for AC currents would choose the exact same frequency that evolution dictated for interactions of brain signals? This is certainly not the only study that reports interesting observations peaking at 60 Hz. One strong line of argument to suggest that this is not line noise is the difference between conditions. For example, in Figure 2B, there is a difference between local and distal interactions. It is hard for me to imagine why line noise would reveal any such difference. Still ...

    The frequency for AC currents in Europe is 50 Hz, not 60 Hz as in the US. Therefore, our SFC effects are well outside the range of the notch.

    Figure 6. I was very excited about Figure 6, which is one of the most novel aspects of this study. In addition to the anatomical questions about this figure noted above, I would like to know more. What is the width of the Gaussian envelope?

    The width of the Gaussian Window used in the original results was 25ms. We chose this time window because in our view it represents a good balance between integrating over a long-enough time window and thus allowing for some jitter in neural firing between pairs of neurons, whilst still being temporally specific. Finding the right balance here is not trivial because a too short time window underestimates co-firing, and a too long time window may not provide the temporal specificity necessary to detect co-firing lags (Cohen & Kohn, 2011; Nat Neurosci). To test whether this choice critically affected our results, we repeated the analysis for different window sizes, i.e. 15, 35, and 45 ms. The results show that the pattern of results did not change, with hits showing earlier peaks in co-firing compared to misses. Critically, the difference in co-firing peaks was significant for all window sizes, except for the shortest one which presumably is due to the increase in noise because of the smaller time window over which spikes are integrated. These issues are now mentioned in the methods section, and the results for the different window sizes are reported in Figure 6 – figure supplement 4.

    Manuscript lines 346-347: “The co-firing analyses were replicated with different smoothing parameters (see Figure 6 – figure supplement 4).”

    Manuscript lines 894-898: “We chose this time window because it should represent a good balance between integrating over a long-enough time window and thus allowing for some jitter in neural firing between pairs of neurons, whilst still being temporally specific (57). To test whether this choice critically affected our results, we repeated the analysis for different window sizes, i.e. 15, 35, and 45 ms (see Figure 6 – figure supplement 4).”

    Are these units on the same or different microwires?

    All units used for the analysis shown in Figure 6 come from different microwires. This was naturally the case because the putative up-stream neuron was distally coupled to the theta LFP, and the putative down-stream neuron was locally coupled to gamma at this same theta LFP electrode. This information is listed in Figure 6 – source data 1 which lists the locations and electrode IDs for all neuron pairs shown in figure 6.

    How do the spike latencies reported here depend on the firing rates of the two units?

    To address this question we first tested whether firing rates (averaged across the putative up-stream and down-stream neurons) differ between hits and misses. If they do, this would be suggestive of a dependency of the spike latency differences between hits and misses on firing rates. No such difference was observed (p>0.3). Second, we correlated the differences between hits and misses in Co-firing peak latencies with the differences in firing rates. Again, no significant correlation was observed (R=-0.06; p>0.7), suggesting that firing rates had no influence on the observed differences in co-firing latencies. These control analyses are now reported in the main text.

    Manuscript lines 347-350: “No significant differences in firing rates between hits and misses were found (p>0.3), and on correlations between firing rates and the co-firing latencies were obtained (R=-0.06; p>0.7), suggesting that firing rates had no influence on the observed co-firing differences between hits and misses.”

    What do these results look like for other pairs that are not putative upstream/downstream pairs?

    As we reported in the original manuscript in lines 352-355 we did not find a memory dependent effect on co-firing latencies if we select neuron pairs solely on the basis of distal theta SFC. Within this analysis the distally theta coupled neuron would be the up-stream neuron and the neuron recorded at the site where the theta LFP is coupled would be the down-stream neuron. This null-result suggests that in order for the memory dependent difference in co-firing lags to emerge, the down-stream neurons need to be coupled to a local gamma rhythm in order for the memory effect on co-firing latencies to emerge. However, within this previous analysis there is still a notion of up-stream and down-stream neurons because neuron pairs were selected based on distal theta phase coupling. We therefore repeated this analysis for all pairs of neurons in a completely unconstrained fashion such that all possible pairs of neurons that were recorded from different electrodes were entered into the co-firing analysis. This analysis also revealed no difference in co-firing lags, neither for positive lags nor for negative lags. Instead, what this analysis showed is tendency for hits to show a higher occurrence of simultaneous or near simultaneous firing, which is in line with Hebbian learning. These results are now reported in Figure 6 – figure supplement 1.

    Manuscript lines 333-335: “In addition, a completely unconstrained co-firing analysis where all pairs possible pairings of units were considered also showed no systematic difference in co-firing lags between hits and misses (Figure 6 – figure supplement 1).”

    Reviewer #2 (Public Review):

    Roux et al. investigated the temporal relationship between spike field coherence (SFC) of locally and distally coupled units in the hippocampus of epilepsy patients to successful and unsuccessful memory encoding and retrieval. They show that SFC to faster theta and gamma oscillations accompany hits (successful memory encoding and retrieval) and that the timing of the SFC between local and distal units for hits comports well with synaptic plasticity rules. The task and data analyses appear to be rigorously done.

    Strengths: The manuscript extends previous work in the human medial temporal lobe which has shown that greater SFC accompanies improved memory strength. The cross-regional analyses are interesting and necessary to invoke plasticity mechanisms. They deploy a number of contemporary analyses to disentangle the question they are addressing. Furthermore, their analyses address limitations or confound that can arise from various sources like sample size, firing rates, and signal processing issues.

    Weaknesses:

    Methodological:

    The SFC coherence measures are dependent in part on extracting LFPs derived from the same or potentially other electrodes that are contaminated by spikes, as well as multiunit activity. In the methods, they cite a spike removal approach. Firstly, the incomplete removal or substitution of a signal with a signal that has a semblance to what might have been there if no spike was present can introduce broadband signal time-locked to the spike and create spurious SFC. Can the authors confirm that such an artifact is not present in their analyses? Secondly, how did they deal with the removal of the multiunit activity? It would be suspected that the removal of such activity in light of refractory period violation might be more difficult than well-isolated units, and introduce artifacts and broadband power, again which would spuriously elevate SFC. Conversely, the lack of removal of multiunit activity would seem to for a surety introduce significant broadband power. One way around this might be that since it is uncommon to have units on all 8 of the BF microwires, to exclude the microwire(s) with the units when extracting the LFP to avoid the need to perform spike removal.

    The reviewer raises a valid concern which we address as follows. Firstly, an artefact introduced into SFC by linear interpolation would be a problem for those local SFCs where the spike providing channel and the LFP providing channel are identical. Out of the 53 local spike-SFC couplings found for the gamma frequency range, only 11 (20.75%) were from pairs where the spikes and LFPs come from the identical microwire. It is unlikely that this minority of data would have driven the results. Furthermore, it is unlikely that the interpolation would introduce a frequency shift of SFC that is memory dependent, because the interpolation is more likely to cause a general increase in broadband SFC (as opposed to having a frequency band specific effect). However, to address this concern, we split the data based on whether the spike and LFP providing channels were the same or different. The results show that (i) the spectrogram of SFC is highly similar between the two datasets, with a prominent gamma peak present in both and no significant differences between the two; (ii) restricting the analysis to those data where the LFP and spike providing channels are different replicated the main finding of faster gamma peak frequencies for hits compared to misses.

    Secondly, we followed the reviewer’s suggestion and repeated the SFC analysis for ‘silent’ microwires, i.e. microwires where no single or multi-units were detected. This analysis replicated the same memory effects as observed in the analysis with all microwires. Specifically, we found an increase in the local gamma peak SFC frequency for hits compared to misses, as well as an increase in distal theta peak SFC frequency for hits compared to misses. These results are reported in the main manuscript and in Figure 2 – figure supplement 6 for gamma, and figure 3 – figure supplement 6 for theta.

    Manuscript lines 199-203: “To rule out concerns about possible artifacts introduced by spike interpolation we repeated the above analysis for spike-LFP pairs where the spike and LFP providing channels are the same or different, and for ‘silent’ LFP channels (i.e. channels were no SUA/MUA activity was detected (see Figure 2 – figure supplement 6).”

    Manuscript lines 253-255: “We also repeated the above analysis for spike-LFP pairs by only using ‘silent’ LFP channels (i.e. channels were no SUA/MUA activity was detected (see Figure 3 – figure supplement 6) to address possible concerns about artefacts introduced by spike interpolation.”

    In a number of analyses the spike train is convolved with a Gaussian in places with a window length of 250ms and in others 25ms. It is suspected that windows of varying lengths would induce "oscillations" of different frequencies, and would thus generate results biased towards the window length used. Can the authors justify their choices where these values are used, and/or provide some sensitivity analyses to show that the results are somewhat independent of the window length of the Gaussian used to convolve with the times series.

    The different choices in window length for the Gaussian convolution reflect the different needs of the two analyses where these convolutions were applied. In one analysis we wanted to get a smooth estimate of spike densities that we can average across trials, similar to a peri-stimulus spike histogram. For this analysis we used a window length of 250 ms which we found appropriate to yield a good balance between retaining smooth time courses whilst still being temporally sensitive. Importantly, for the statistical analysis of the firing rates, spike densities were averaged in much larger time windows than 250 ms (i.e. 1 – 2 seconds) therefore our choice of window length for spike densities would not have any bearing on the averaged firing rate analysis.

    In the other analysis, which is more central for our manuscript, we used a cross-correlation between spike trains to estimate co-firing lags in the range of milliseconds. Therefore, this analysis necessitated a much higher temporal precision. We used a Gaussian Window with a width of 25ms because it represents a good balance between integrating over a long-enough time window and thus allowing for some jitter in neural firing between pairs of neurons, whilst still being temporally specific. Finding the right balance here is not trivial because a too short time window would be prone to noise and underestimates co-firing, whereas a too long time window may not provide the temporal specificity necessary to detect co-firing lags (Cohen and Kohn, 2013; Nat Neurosci). To test whether this choice critically affected our results, we repeated the analysis for different window sizes, i.e. 15, 35, and 45 ms. The results show that the basic pattern of results did not change, with hits showing earlier peaks in co-firing compared to misses. Critically, the difference in co-firing peaks was significant for all window sizes, except for the shortest one which is likely due to the increase in noise because of the smaller time window over which spikes are integrated. These issues are now mentioned in the methods section, and the results for the different window sizes are reported in Figure 6 – figure supplement 4.

    Manuscript lines 346-347: “The co-firing analyses were replicated with different smoothing parameters (see Figure 6 – figure supplement 4).”

    Manuscript lines 894-898: “We chose this time window because it should represent a good balance between integrating over a long-enough time window and thus allowing for some jitter in neural firing between pairs of neurons, whilst still being temporally specific (57). To test whether this choice critically affected our results, we repeated the analysis for different window sizes, i.e. 15, 35, and 45 ms (see Figure 6 – figure supplement 4).”

    Conceptual:

    The co-firing analyses are very interesting and novel. In table S1 are listed locally and distally coupled neurons. There are some pairs for example where the distally coupled neuron is in EC and the downstream one in the hippo, and then there is a pair that is the opposite of this (dist: hippo, local EC). There appear to be a number of such "reversal", despite the delay between these two regions one would assume them to be similar in sign and magnitude given the units are in the same two regions. It seems surprising that in two identical regions of the hippo the flow of information or "causality", could be reversed, when/if one assumes information flows through the system from EC to hippo. This seems unusual and hard to reconcile given what is known about how information flows through the MTL system.

    The reviewer is correct that the spike co-firing analysis suggests a bi-directional flow of information between the hippocampus and surrounding MTL regions (e.g. entorhinal cortex; see Figure 6 – figure supplement 3). However, this bi-directional flow of information is not incompatible with neuroanatomy and the memory literature. The entorhinal cortex serves as an interface between the hippocampus and the neocortex with superficial layers providing input into the hippocampus (via the perforant pathway), and the deeper layers receiving output from the hippocampus (van Strien et al., 2009; Nat Rev Neurosci). Therefore, on a purely anatomical basis we can expect to see a bi-directional flow of information between the hippocampus and the entorhinal cortex, albeit in different layers. Importantly, reversals as shown in our Figure 6 – source data 1 involved different microwires and therefore different neurons (i.e. the entorhinal unit in row 1 was recorded from microwire 3, whereas the entorhinal unit in row 2 was recorded from microwire 8). It is conceivable that these two neurons correspond to different layers of the entorhinal cortex and therefore reflect input vs. output paths. Moreover, studies in humans demonstrated that successful encoding of memories depends not only on the input from the entorhinal cortex into the hippocampus, but also on the output of the hippocampal system into the entorhinal cortex, and indeed on the dynamic recurrent interaction between these input and output paths (Maass et al. 2014; Nat Comms; Koster et al., 2018; Neuron). Our bi-directional couplings between hippocampal and surrounding MTL regions (such as the EC) are in line with these findings. We have added a discussion of this issue in the discussion section.

    Manuscript lines 447-452: “Notably, the neural co-firing analysis indicates a bidirectional flow of information between the hippocampus and surrounding MTL areas, such as the entorhinal cortex (see Figure 6 – figure supplement 3; Figure 6 – source data 1). This result parallels other studies in humans showing that successful encoding of memories depends not only on the input from surrounding MTL areas into the hippocampus, but also on the output of the hippocampal system into those areas, and indeed on the dynamic recurrent interaction between these input and output paths (43, 44).”

  3. eLife

    Evaluation Summary:

    Roux and colleagues measured spiking activity and local field potentials predominantly from the hippocampus and also a few surrounding structures in the medial temporal lobe from patients with pharmacologically intractable epilepsy while the patients performed a cued-recall task. They report differences in local spike-field coherence measurements between hits and misses in the gamma frequency band and differences in both local and distal spike-field coherence measurements between hits and misses in the theta frequency band. The authors further report differences in the timing of spikes between pairs of neurons, with hits correlated with putative downstream neurons firing about 30 ms after putative upstream neurons and misses correlated with delays of about 60 ms. Overall, these are interesting observations that provide intriguing data to further think about how neurons in the medial temporal lobe correlate with recognition memory.

    (This preprint has been reviewed by eLife. We include the public reviews from the reviewers here; the authors also receive private feedback with suggested changes to the manuscript. Reviewer #1 and Reviewer #2 agreed to share their names with the authors.)

  4. eLife

    Reviewer #1 (Public Review):

    This is an interesting study with observations that provide intriguing data to further think about how neurons in the medial temporal lobe correlate with recognition memory.

    Figures 2 through 6. There is no description of the relationship between the findings and the anatomical location of the electrodes (other than distal versus local). Perhaps the non-uniform distribution of electrodes makes these analyses more complicated and such questions might have minimal if any statistical power. But how should we think about the claims in Figures 2-6 in relationship to the hippocampus, amygdala, entorhinal cortex, and parahippocampal gyrus? As one example question out of many, is Figure 2C revealing results for local pairs in all medial temporal lobe areas or any one area in particular? I won't spell out every single anatomical question. But essentially every figure is associated with an anatomical question that is not described in the results.

    Figure 1
    1A. I assume that image positions are randomized during a cued recall?
    What was the correlation between subjects' indication of how many images they thought they remembered and their actual performance?
    1B. Chance is shown for hits but not misses. I assume that hits are defined as both images correct and misses as either 0 or 1 image correct. Then a chance for misses is 1-chance for hits = 5/6. It would be nice to mark this in the figure.
    The authors report that both incorrect was 11.9%. By chance, both incorrect should be the same as both correct, hence also 1/6 probability, hence the probability of both incorrect seems quite close to chance levels, right?
    1C. How does the number of electrodes relate to the number of units recorded in each area?

    Line 152. The authors state that neural firing during encoding was not modulated by memory for the time window of interest. This is slightly surprising given that other studies have shown a correlation between firing rates and memory performance (see Zheng et al Nature Neuroscience 2022 for a recent example). The task here is different from those in other studies, but is there any speculation as to potential differences? What makes firing rates during encoding correlate with subsequent memory in one task and not in another? And why is the interval from 2-3 seconds more interesting than the intervals after 3 seconds where the authors do report changes in firing rates associated with subsequent performance? Is there any reason to think that the interval from 2-3 seconds is where memories are encoded as opposed to the interval after 3 seconds?

    Lines 154-157 and relationship to the subsequent analyses. These lines mention in passing differences in power in low-frequency bands and high-frequency bands. To what extent are subsequent results (especially Figures 3 and 4) related to this observation? That is, are the changes in spike-field coherence, correlated with, or perhaps even dictated by, the changes in power in the corresponding frequency bands?

    Do local interactions include spike-field coherence measurements from the same microwire (i.e., spikes and LFPs from the same microwire)?

    Figure 6. I was very excited about Figure 6, which is one of the most novel aspects of this study. In addition to the anatomical questions about this figure noted above, I would like to know more. What is the width of the Gaussian envelope? Are these units on the same or different microwires? How do the spike latencies reported here depend on the firing rates of the two units? What do these results look like for other pairs that are not putative upstream/downstream pairs?

  5. eLife

    Reviewer #2 (Public Review):

    Roux et al. investigated the temporal relationship between spike field coherence (SFC) of locally and distally coupled units in the hippocampus of epilepsy patients to successful and unsuccessful memory encoding and retrieval. They show that SFC to faster theta and gamma oscillations accompany hits (successful memory encoding and retrieval) and that the timing of the SFC between local and distal units for hits comports well with synaptic plasticity rules. The task and data analyses appear to be rigorously done.

    Strengths:

    The manuscript extends previous work in the human medial temporal lobe which has shown that greater SFC accompanies improved memory strength. The cross-regional analyses are interesting and necessary to invoke plasticity mechanisms. They deploy a number of contemporary analyses to disentangle the question they are addressing. Furthermore, their analyses address limitations or confound that can arise from various sources like sample size, firing rates, and signal processing issues.

    Weaknesses:

    Methodological:
    The SFC coherence measures are dependent in part on extracting LFPs derived from the same or potentially other electrodes that are contaminated by spikes, as well as multiunit activity. In the methods, they cite a spike removal approach. Firstly, the incomplete removal or substitution of a signal with a signal that has a semblance to what might have been there if no spike was present can introduce broadband signal time-locked to the spike and create spurious SFC. Can the authors confirm that such an artifact is not present in their analyses? Secondly, how did they deal with the removal of the multiunit activity? It would be suspected that the removal of such activity in light of refractory period violation might be more difficult than well-isolated units, and introduce artifacts and broadband power, again which would spuriously elevate SFC. Conversely, the lack of removal of multiunit activity would seem to for a surety introduce significant broadband power. One way around this might be that since it is uncommon to have units on all 8 of the BF microwires, to exclude the microwire(s) with the units when extracting the LFP to avoid the need to perform spike removal.

    In a number of analyses the spike train is convolved with a Gaussian in places with a window length of 250ms and in others 25ms. It is suspected that windows of varying lengths would induce "oscillations" of different frequencies, and would thus generate results biased towards the window length used. Can the authors justify their choices where these values are used, and/or provide some sensitivity analyses to show that the results are somewhat independent of the window length of the Gaussian used to convolve with the times series.

    Conceptual:
    The co-firing analyses are very interesting and novel. In table S1 are listed locally and distally coupled neurons. There are some pairs for example where the distally coupled neuron is in EC and the downstream one in the hippo, and then there is a pair that is the opposite of this (dist: hippo, local EC). There appear to be a number of such "reversal", despite the delay between these two regions one would assume them to be similar in sign and magnitude given the units are in the same two regions. It seems surprising that in two identical regions of the hippo the flow of information or "causality", could be reversed, when/if one assumes information flows through the system from EC to hippo. This seems unusual and hard to reconcile given what is known about how information flows through the MTL system.

  6. Version published to 10.1101/2021.01.28.428480v3 on bioRxiv
  7. Version published to 10.1101/2021.01.28.428480v2 on bioRxiv
  8. Version published to 10.1101/2021.01.28.428480v1 on bioRxiv