Cholinergic blockade during encoding, but not retrieval, impairs recollection memory.

a. Overview of methods. In order to study the neural mechanisms behind the effects of cholinergic blockade on human memory, we delivered scopolamine, a cholinergic blocker, to patients with intracranial encephalographic (iEEG) electrodes during an associative recognition memory task. b. Diagram of the drug paradigm. In each session (drug or placebo), patients completed four blocks of the memory task. The blocks had the same task structure, starting with encoding followed by a math distractor and retrieval. In Block 2, subjects received a substance injection (scopolamine or saline), after encoding but prior to the retrieval period. Block 2’s paradigm allowed us to examine whether scopolamine impacts retrieval independently from encoding. c. Diagram of the task structure. Briefly, during the encoding period of the task, subjects are presented with a sequence of words pairs. Then, after completing a math distractor consisting of simple algebra equations, subjects initiate retrieval where they are asked to indicate whether word pairs presented are intact, rearranged or new in relation to the word pairs presented during encoding. d. Bar plot showing recall probabilities for encoding (placebo = all blocks, scopolamine = Blocks 3 and/or 4) and retrieval (placebo = all blocks, scopolamine = Block 2) across all trials. We found a significant decrease in memory performance when scopolamine is present at encoding, but not when it is present during retrieval alone (encoding Block 3 only: t = 42.0, p = 9.77X 10−3; encoding Block 4 only: t = 41.0, p = 1.37X 10−2; encoding Blocks 3 and 4: t = 44.0, p = 3.91X 10−3; retrieval: t = 32.0, p = 0.15; Wilcoxon rank sum tests). e. Bar plot showing change in memory as a function of block for all trials (t(9) = −2.42, p = 1.5X 10−2, LME model), and separately for recollection, familiarity, and novelty trials (recollection: t(9) = −2.72, p = 6.5X 10−3; familiarity: t(9) = −1.01, p = 0.31; novelty trials: t(9) = −0.50, p = 0.61; LME models). f. Correlation between percent change in memory (with respect to the placebo session) and dosage during recollection trials for Blocks 3 and 4 combined (r = −0.80, p = 0.01, two-tail Pearson’s correlation).

Cholinergic blockade disrupts theta power during retrieval.

a. Anatomical locations of all hippocampal electrodes analyzed in the study. b. Retrieval power effects for an example electrode during Block 4. Left: line plot displaying the raw iEEG trace (black) and corresponding slow theta (2–4 Hz) filtered signal (blue) from an example trial. Middle: time-frequency spectrograms displaying mean normalized power across all trials for the example electrode during the placebo and scopolamine sessions. Time 0 s denotes the onset of the retrieval cue, and contrast denotes changes in normalized power. Right: mean slow theta power over time for placebo (blue) and scopolamine (red). Shading denotes ± SEM. c. Group-level mean slow theta power across all trials for Blocks 1, 2, 3 and 4 for placebo (blue) and scopolamine (red). Significant disruptions to slow theta power following scopolamine are present in Blocks 2 through 4 (p′s < 0.05, LME models, multiple-comparison corrected). Shading denotes ± SEM. d. Correlations between scopolamine dosage and change in slow theta power for all blocks (Block 1: r = 0.19, p = 0.94; Block 2: r = −0.17, p = 0.17; Block 3: r = −0.29, p = 0.02; Block 4: r = −0.55, p = 1.80X 10−6; one-tail Pearson’s correlations). e. Scatter plot showing change in slow theta power for all electrodes across blocks (t = −3.20, p = 1.52X 10−3, one-sided t-test). f. Change in power across blocks for different frequency bands. Only the slow theta frequency band shows a significant decrease in power across blocks. Dashed line denotes 95% significance level. g. Bar plot showing changes in mean slow theta power as a function of subregion (Blocks 3 and 4: anterior hippocampus (AH): t = −2.32, p = 0.02; posterior hippocampus (PH): t = −2.36, p = 0.02; and entorhinal cortex (EC): t = −1.46, p = 0.14; LME model, multiple-comparison corrected). Dashed line denotes 95% significance level. h. Bar plot showing changes in mean slow theta power as a function of trial category (Block 3 and 4: recollection: t = −3.35, p = 8.07X 10−4; familiarity: t = −3.26, p = 1.10X 10−3; novelty: t = −3.40, p = 6.77X 10−4; LME model, multiple-comparison corrected). Dashed line denotes 95% significance level.

Dimensionality reduction of spectral power enables decoding of drug condition.

a. Schematic panel outlining the analysis pipeline, including frequency-band data processing, principal component analysis (PCA) decomposition, and classification methods. This unsupervised method allows us to characterize the correlational structure of retrieval-related oscillations across frequencies and brain regions, and identify features associated with pharmacological condition. b. Schematic representation of low and high dimensional PCA decompositions. The separation of trajectories illustrates distinct neural representations across conditions. In low-dimensional problems, one or a few principal components (PCs) capture most of the variance and may be sufficient to support accurate decoding. In contrast, high-dimensional problems typically require many components to capture relevant variance, making decoding more complex. c. Bar plot showing variance explained by the first three PCs, emphasizing the dominant contribution of PC1. Error bars denotes ± SEM across electrodes. d. Histogram displaying mean cumulative variance (average of the first three PCs) across electrodes. The mean cumulative variance for all individual electrodes is greater than 0.9, indicating low-dimensionality in the data decomposition. e. Left: PCA visualization for an example electrode, showing neural trajectories from the placebo (blue) and scopolamine (red) conditions in the first three principal components. We performed dimensionality reduction for each electrode, and decoded condition (placebo or scopolamine) using the resulting PCs. Right: histogram of decoding accuracies across electrodes with a significant decoding accuracy peak at 0.6 (p < 0.001, Wilcoxon signed-rank test). Dashed line indicates chance-level decoding (0.5). f. Left: schematic of frequency-band-level decoder. We decoded condition (placebo or scopolamine) using frequency-band-filtered data across electrodes. Right: line plot comparing decoding accuracies (AUC) using slow theta, fast theta and slow gamma frequency-band-filtered data with increasing number of electrodes. Decoding accuracy increased with number of electrodes and was highest for the slow theta band, exceeding that of both fast theta and slow gamma bands.

Cholinergic blockade impairs theta phase alignment during retrieval.

a. Retrieval inter-trial phase coherence (ITPC, or phase reset) effects for an example electrode during Block 2. Time 0 s denotes onset of retrieval cue, and contrast denotes strength of phase reset. During the placebo session, the electrode shows increased slow theta (2–4 Hz) phase reset following cue onset. This effect is absent during the scopolamine session. b. Polar plots showing phase distribution for a 3 Hz oscillation at 0.3 s for the same example electrode. During the placebo session, the electrode exhibits significant phase clustering during Blocks 2, 3 and 4 (Block 2: z = 14.9, p = 2.27X 10−7, Block 3: z = 9.16, p = 9.15X 10−5, Block 4: z = 4.63, p = 9.54X 10−3; Rayleigh tests), but that effect is absent during the scopolamine session (Block 2: z = 0.23, p = 0.79, Block 3: z = 0.51, p = 0.60, Block 4: z = 2.78, p = 0.06; Rayleigh tests). c. Group-level mean slow theta phase reset is significantly disrupted following scopolamine in Blocks 2 through 4 (p′s < 0.05, LME models, multiple-comparison corrected) across trials. Shading denotes ± SEM. d. Correlations between scopolamine dosage and change in slow theta phase reset for all blocks (Block 1: r = 0.29, p = 0.98; Block 2: r = −0.27, p = 0.02; Block 3: r = −0.30, p = 0.01; Block 4: r = −0.10, p = 0.24; one-tail Pearson’s correlations). e. Group-level phase reset effects at encoding. f. Bar plot showing mean phase clustering across all electrodes during placebo and scopolamine for the encoding and retrieval periods. Scopolamine disrupts slow theta phase reset during both encoding and retrieval periods (encoding: t = −2.07, p = 0.04; retrieval: t = −2.69, p = 0.01; paired t-tests). g. Phase reset effects for an example electrode during placebo. Increased 2–4 Hz ITPC is present at both encoding and retrieval. h. Example electrode effects for a 3 Hz oscillation at 0.5 s. Top: mean phase. Bottom: corresponding phase difference. Phase difference between encoding and retrieval is constant and nearly zero between 0 and 1 s. i. Group-level consistency of slow theta phase differences between encoding and retrieval during placebo. Dashed line denotes Rayleigh test significance (95% confidence level). j. Group-level distribution of slow theta phase differences at 0.5 s (z = 8.05, p = 2.98X 10−4, Rayleigh test).

Cholinergic blockade prevents encoding–retrieval pattern reinstatement.

a. Schematic panel outlining methods for spectral pattern reinstatement analysis. First, we identified matching encoding and retrieval periods for all recollection trials. We then computed the correlation of 4–128 Hz power between encoding and retrieval across 250 ms time bins during the initial 3 s following the encoding or retrieval cue. The resulting similarity matrix quantifies spectral reinstatement by showing whether spectral patterns at different time points during encoding are reinstated at similar, later or earlier time points during retrieval. b. Similarity matrix for example electrode. The “no drug” plot reflects the mean across all blocks within the placebo session, whereas the “scopolamine” plot shows combined data from Blocks 3 and 4. The contrast denotes the Pearson’s correlation coefficient (PCC) of 4–128 Hz power data between encoding and retrieval. Black outlines denote pixels exhibiting greater-than-chance reinstatement compared to surrogate data (95% confidence level). Surrogate data for electrode-level analyses involved shuffling trial labels. c. Bar plot showing mean reinstatement across conditions. Retrieval reinstatement during the no-drug condition is significantly greater than during the scopolamine conditions (no-drug versus scopolamine Block 2: t = 5.04, p = 1.05X 10−6; no-drug versus scopolamine Blocks 3 and 4: t = 5.06, p = 9.68X 10−7; paired t-tests). d. Group-level mean reinstatement matrices for no-drug condition, scopolamine Block 2, and scopolamine Blocks 3 and 4 combined. Black outlines denote pixels exhibiting greater-than-chance reinstatement compared to surrogate data (95% confidence level). Surrogate data for group-level analyses involved shuffling block labels.

Subject demographics and electrode coverage.

Table showing subject demographics and number of electrodes across all regions of interest, including the anterior hippocampus (AH), the posterior hippocampus (PC), and the entorhinal cortex (EC).

Scopolamine dosage correlates with changes in memory during recollection trials.

We found significant effects of scopolamine dosage on memory. These effects were exclusive to items recovered on the basis of recollection rather than familiarity or detected novel items, consistent with a principal impact of cholinergic blockade on associative processes (all: t = −4.48, p = 7.37X 10−6; recollection: t = −3.65, p = 2.58X 10−4; familiarity: t = −0.22, p = 0.82; novelty: t = −0.51, p = 0.61; LME model, multiple-comparison corrected).

Changes in memory following scopolamine for subjects with medium and high doses only.

Bar plot showing recall probabilities for encoding (placebo = all blocks, scopolamine = Blocks 3 and 4) and retrieval (placebo = all blocks, scopolamine = Block 2) across all trials for subjects with medium or high scopolamine doses (> 400 mcg) only. Even after excluding subjects with lowest scopolamine doses (< 400 mcg), we found a significant decrease in memory performance when scopolamine is present at encoding, but not when it is present during retrieval alone (encoding: t = −2.58, p = 0.04; retrieval: t = −1.25, p = 0.26; paired t-tests).

Bar plot showing change in memory across blocks for recollection trials.

There are no significant differences between Block 1 (baseline) and Block 2 (when the drug is present during retrieval only) (t = 0.34, p = 0.74; paired t-tests). This shows that cholinergic blockade during retrieval alone is not sufficient to elicit memory deficits. However, memory is modestly worse in Blocks 3 and 4 (when the drug is present during both encoding and retrieval) compared to Blocks 1 and 2 (t = 2.232, p = 0.056, paired t-tests).

Cholinergic blockade disrupts slow theta power during encoding.

Group-level time-frequency spectrograms displaying changes in mean normalized during encoding for the no-drug (all drug-free blocks combined) and scopolamine conditions. Time 0 s denotes onset of encoding cue. Contrast denotes change in normalized power. Mirroring earlier reports18, in the placebo session there was higher slow theta power during memory encoding, but this effect is absent during the scopolamine session.

Disruptions to retrieval slow theta power in remembered trials only.

Patterns of retrieval slow theta power disruptions during remembered trials are consistent with the disruptions observed across all trials (remembered and forgotten). a. Group-level time-frequency spectrograms displaying mean normalized for the placebo and scopolamine session during Block 2 for remembered trials only. Time 0 s denotes onset of encoding cue. Contrast denotes change in normalized power. b. Mean slow theta power over time for placebo (blue) and scopolamine (red) sessions for Block 2 (p < 0.05, LME model, multiple-comparison corrected). Shading denotes ± SEM. c. Group-level time-frequency spectrograms displaying mean normalized for the placebo and scopolamine session during Blocks 3 and 4 combined for remembered trials only. Time 0 s denotes onset of encoding cue. Contrast denotes change in normalized power. d. Mean slow theta power over time for placebo (blue) and scopolamine (red) sessions for Blocks 3 and 4 combined (p < 0.05, LME model, multiple-comparison corrected). Shading denotes ± SEM.

Disruptions in retrieval phase reset following cholinergic blockade are driven by the anterior and posterior hippocampus.

Bar plot depicting changes in slow theta (2–4 Hz) power for different medial temporal lobe (MTL) subregions. The effects are strongest in the anterior and posterior hippocampus (Blocks 3 and 4: anterior hippocampus (AH): t = −3.97, p = 7.1X 10−5; posterior hippocampus (PH): t = −2.87, p = 4.06X 10−3; and entorhinal cortex (EC): t = −0.94, p = 0.34; LME model, multiple-comparison corrected).

Mean preferred phase at encoding is recapitulated at retrieval.

Polar plots depicting phase distributions and mean phases across all electrodes for encoding (light blue) and retrieval (dark blue). At time 0 s, the preferred phase for encoding and retrieval closely match, with mean a phase difference of only 1.6◦. The mean angles for encoding and retrieval remain significantly clustered until 0.75 s following the cue.

Dimensionality of spectral decoding subspace.

a. Histogram of decoding accuracy across electrodes using only the first principal component (PC1). Mean accuracy (triangle) is significantly above chance (p < 0.001, Wilcoxon signed-rank test). b. Electrode-level decoding accuracy across PCA dimensions (1–3 PCs). While individual decoding performance increases slightly, the improvement of decoding with higher dimensionality is modest. c. Cumulative variance explained across PCA dimensions. Increasing the number of PCs substantially improves the variance explained, indicating that higher components capture more spectral variability but not necessarily more task-relevant information. d. Correlations between change in decoding accuracy and change in variance explained between 1PC and 2PCs (left), and 2PCs and 3PCs (right). The weak or absent correlation between variance gain and decoding improvement underscores that not all variance dimensions are behaviorally relevant. While additional principal components significantly increase explained variance, their contributions to decoding accuracy remain minimal or even negative. These findings support a model where early PCs capture most task-relevant structure, and later PCs primarily reflect noise or irrelevant variability.