Control of working memory maintenance by theta-gamma phase amplitude coupling of human hippocampal neurons

Abstract

Retaining information in working memory (WM) is a demanding process that relies on cognitive control to protect memoranda-specific persistent activity from interference. How cognitive control regulates WM storage, however, remains unknown. We hypothesized that interactions of frontal control and hippocampal persistent activity are coordinated by theta-gamma phase amplitude coupling (TG-PAC). We recorded single neurons in the human medial temporal and frontal lobe while patients maintained multiple items in WM. In the hippocampus, TG-PAC was indicative of WM load and quality. We identified cells that selectively spiked during nonlinear interactions of theta phase and gamma amplitude. These PAC neurons were more strongly coordinated with frontal theta activity when cognitive control demand was high, and they introduced information-enhancing and behaviorally relevant noise correlations with persistently active neurons in the hippocampus. We show that TG-PAC integrates cognitive control and WM storage to improve the fidelity of WM representations and facilitate behavior.

Figure 1.. Task, recordings sites, and behavior.

(a) Example trial. Each trial began with a fixation cross followed by either one (load 1) and three (load 3) consecutively presented pictures, each presented for 2 s (separated by a variable blank screen of up to 200 ms as indicated by a small dot). After a variable maintenance period of on average 2.7 s duration, a probe picture was presented. The task was to decide whether the probe picture has been part of the pictures shown during encoding in this trial (correct answer “No” in the example shown). Pictures were drawn from five categories: people, animals, cars/tools, fruit, landscapes. (b) Recording locations. Each colored dots represent the location of a micro-wire bundle across all 44 sessions shown on a standardized MNI152 brain template (left) and in a 3D model using the brainnetome atlas (right). (c) Proportions of neurons recorded in each brain area. The three frontal areas (pre-SMA, dACC, vmPFC) are jointly referred to as medial frontal cortex (MFC). (d) Behavioral results. Patients made fewer errors and responded faster in load 1 as compared to load 3 trials. Each line connects the two dots belonging to the same session. RT was measured relative to probe stimulus onset. *** p < 0.001, permutation-based t-test.

Figure 2.. Assessment of theta-gamma phase amplitude coupling.

(a) Average normalized modulation indices for all phase-amplitude pairs across all n=1949 channels. The strongest phase-amplitude coupling was observed for phases in the theta range (3–7 Hz) and amplitudes from two gamma bands: a low (30–55 Hz) and a high gamma band (70–140 Hz). (b) Theta-gamma PAC differed by load. Log-normalized modulation indices were averaged within the theta-high gamma band and compared between the two load conditions in each significant PAC channel (p < 0.05) in each region. PAC channels were common in the hippocampus and amygdala (23% and 31%, respectively) but not in the three frontal areas (6.5%, combined and labeled as MFC (medial frontal cortex)). Only in the hippocampus did theta-high gamma PAC differ significantly as a function of load, with PAC higher in load 1 vs. load 3 (left). Z-scored PAC values were shifted into a positive range by an offset of 1 and log-transformed for illustrative purposes only. All statistics are based on non-transformed z-values. (c) Average normalized modulation indices for significant PAC channels from the hippocampus separately for each load condition. (d) Theta-gamma PAC was significantly negatively correlated with reaction times in the hippocampus, but not in the amygdala. See Table S1 for GLM results that controls for load differences in RT. (e) High-amplitude gamma events were more frequent in load 3 than in load 1 within PAC channels from the hippocampus. (f) The longer the gamma events, the weaker was PAC. Shown is the relationship between gamma event duration and PAC. The difference between the durations of high-amplitude gamma events between load 3 and load 1 was negatively correlated with the difference in the modulation index between the two loads across all hippocampal PAC channels. *** p < 0.001; * p < 0.05; ns = not significant, permutation-based t-tests; mixed-model GLMs.

Figure 3.. Firing rates and local SFC of category neurons in the MTL.

(a) Example category neuron recorded in the hippocampus. Category neurons were selected based on higher firing rates for one category than for all other categories during encoding period 1 (ANOVA + post-hoc t-test, both p < 0.05). The preferred category of this neuron was ‘animals’. (b) Firing rates averaged across the maintenance period separately for preferred and non-preferred categories for all category neurons from hippocampus and amygdala. Category neurons remained active as compared to baseline and retained their selectivity during the maintenance period of the task, with FR persistently higher for preferred than non-preferred categories. Each dot is a neuron (n=270). Firing rates are shown as percent change to baseline (−0.9 to −0.3 s prior to the first picture onset). (c) Firing rates of category neurons during the maintenance period were higher in load 1 as compared to load 3 when their preferred category was held in WM. Each dot is a neuron (n=270). (d) In correct trials, FR of category neurons was higher as compared to incorrect trials across both load conditions. Each dot is a neuron (n=246). 24 neurons were excluded due to insufficient data in the incorrect condition. (e) We computed local spike-field coherence between spikes and LFPs recorded in the same area for all category neurons and compared preferred versus non-preferred trials. (f) When paired with local PAC channels, spikes of category neurons in the hippocampus were significantly more strongly phase-locked to local gamma LFPs in the high gamma range during the maintenance period when the preferred category of a neuron was held in WM. No significant differences were found for the amygdala or non-PAC channels (see Fig. S3e). (g) Gamma (70–140 Hz) SFC for hippocampal category neurons was significantly stronger for preferred vs non-preferred trials in both load conditions. No main effect of load or interaction was found. Each dot is a neuron-LFP channel pair (n=151). (h) Theta-gamma PAC was significantly positively correlated with firing rates of category neurons in the hippocampus, but not in the amygdala. See Table S2 for GLM results that control for load differences in PAC. *** p < 0.001; ** p < 0.01; * p < 0.05; ns = not significant; (cluster-based) permutation t-test/ANOVA; mixed-model GLMs.

Figure 4.. PAC neuron selection and local activity.

(a) Example showing binning used for PAC neuron selection for a neuron from the hippocampus. Theta phase, binned into ten groups, and gamma amplitude, median split into low and high, were used to predict spike counts of each neuron from the MTL during the maintenance period. Only if the model containing the two factors and their interaction predicted spike counts significantly better than two other models that lacked the interaction or the gamma amplitude main effect term, respectively, a given neuron was selected as PAC neuron. In this example neuron from the hippocampus, spike count was higher during high gamma amplitudes (gamma main effect) and differed in their theta phase distribution between high and low gamma amplitudes (interaction effect), resulting in selecting this neuron as a PAC neuron. In the analysis, we separated the theta phase into sine and cosine terms to account for the circularity of phase values, which is not shown here for simplicity. (b) Proportions of all recorded neurons that qualified as PAC neurons (yellow, green). (c,d) PAC neurons were not selective for category. Even during picture presentation (encoding), image category could not be efficiently decoded from firing rates of “PAC only” neurons. Error bars reflect the standard deviation of 1,000 decoding repetitions. Black horizontal lines indicate mean decoding accuracy of 1,000 randomly shuffled category labels. Decoding was performed for pseudo-populations of category or PAC neurons, respectively. (e,f) Firing rates of PAC neurons were positively correlated with single-trial estimates of theta-gamma PAC in the hippocampus, not the amygdala (see Table S3 for GLM results). (g,h) Firing rates of PAC neurons during the maintenance period differed between correct and incorrect trials in the hippocampus but not amygdala. Firing rates are shown as percent change to baseline (−0.9 to −0.3 s prior to the first picture onset). (i-l) Firing rates as well as theta, gamma SFC between PAC neurons and local LFP recordings did not differ as a function of load in both MTL areas. Theta and gamma SFC, however, were both significantly stronger than shuffled surrogates in the hippocampus as well as the amygdala. In (e-l), each dot is a neuron. *** p < 0.001; ** p < 0.01; * p < 0.05; ns = not significant; permutation-based t-tests; mixed-model GLMs.

Figure 5.. Remote connectivity of PAC neurons in the MTL to frontal theta LFPs

. (a) We computed long-range SFC between MTL spiking activity and LFPs recorded from all three frontal regions. (b) Spikes of PAC neurons in the hippocampus were more strongly synchronized with theta-band LFPs recorded in the vmPFC during the maintenance period during load 3 compared to load 1 trials. This was not the case for pre-SMA and dACC (see Fig. S5a). (c) Category neurons from the hippocampus, or (d) PAC neurons from the amygdala did not show significant SFC differences between loads relative to vmPFC LFP. (e) Hippocampal PAC cells (n=175 connections, cyan line) yielded the strongest long-range theta SFC difference between load 3 and load 1 trials among 10,000 random selections of hippocampus-vmPFC connections. T-values correspond to comparisons between load 3 and 1 for an average of SFC values in the significant theta range (2.5–4.3 Hz). (f) Remote theta-band SFC between spiking activity of PAC neurons and LFPs recorded in the vmPFC was significantly stronger for fast than for slow RT trials. Each dot is a neuron-channel connection (n=167; some neuron-channel pairs were excluded due to inefficient spike count in at least one of the conditions). *** p < 0.001; * p < 0.05, H = Hippocampus; A = Amygdala; (cluster-based) permutation t-tests.

Figure 6.. Noise correlations of PAC neurons within MTL.

(a) Trial-averaged, bin-wise correlation coefficients for all possible pairs of category and PAC neurons in the hippocampus and amygdala. In both regions, correlation coefficients were significantly higher than zero on average. (d) Repeat of the correlation analysis for all possible PAC-category neuron pairs in the hippocampus. Shuffling trial labels for 1,000 times resulted in far lower correlations between pairs of neurons than unshuffled trial labels (cyan line; mean of correlation coefficients across all pairs. (c) Single-session example for optimized decoding accuracies for firing rates during the maintenance period for hippocampal neurons. Category decoding accuracies were computed with intact or removed noise correlations among neurons. Purple dots indicate when a category cell, yellow dots when a PAC neuron was added to the optimized decoding ensemble. White dots show decoding accuracies for each individual neuron. (d) (left side) Adding hippocampal PAC neurons to the optimized decoding ensemble significantly enhanced decoding performance of WM content when noise correlations were kept intact. When noise correlations were removed, in contrast, PAC cells did not improve decoding performance. This indicates that, for PAC cells, noise correlations rather than individual firing rates shaped the geometry of category representations. (right side) Amygdala PAC neurons contributed more to decoding memory content with intact noise correlations. Note, however, that amygdala PAC cells contributed to decoding also after removing noise correlations during the maintenance period. Each dot represents a neuron. (e) Noise correlations among hippocampal PAC and category neurons were stronger in fast than slow RT trials (median split) for trials in which the category neuron’s preferred category was correctly maintained (left). This effect was especially strong in load 3 trials (right), not in load 1 trials (middle). (f) In the amygdala, we did not observe a significant difference in noise correlations between fast and slow trials when the preferred category was maintained for pairs of PAC and category neurons. In (e,f) each dot represents a cell pair. (g) Pairs of PAC and category neurons in the hippocampus (pink line) showed a significantly stronger difference in noise correlations between fast and slow RT trials than randomly selected pairs of any recorded neuron (from the same session) and category neurons in 10,000 repetitions. *** p < 0.001; ** p < 0.01; * p < 0.05; ns = not significant; H = Hippocampus; permutation-based t-tests.