续接03

SCR recording

SCR was recorded simultaneously with fMRI scanning using a Biopac MP 150 System (Biopac, Inc., Goleta, CA). Two Ag/AgCl electrodes filled with isotonic electrolyte medium were attached to the center phalanges of the index and middle fingers of the left hand. The gain set to 5, the low pass filter set to 1.0 Hz, and the high pass filters set to DC⁵⁸. Data were acquired at 200 samples per second. Before analysis, the data were transformed into microsiemens (μS) and square root transformed due to non-normality of the data distribution.

SCR analysis

SCR data were analysed offline using Matlab R2014a (MathWorks, Natick, USA). First, data were temporally smoothed with a median filter (that is, 40 samples within a 200 ms window) to reduce scanner-induced noise^59,60. We used Autonomate⁶¹ to analyse event-related SCRs, which has been proven effective in the context of event-related cognitive tasks^2,61. In brief, the electrodermal data were segmented into event-related time windows based on face onset. The face-related SCR was located by identifying rises in the electrodermal data, which constituted the onset of an SCR. Responses that did not fit these criteria were scored as zero. Noisy segments of data (in which an implausible number of candidate SCRs were present) were excluded from further analysis (see more details in the Supplementary Methods). The resulting number of trials used for SCR analyses are reported in the Supplementary Table 6. SCR values were then identified as the maximum value within each time window. If multiple SCRs fell within the same window, the largest response was scored⁶¹. A 2-by-2 repeated-measures ANOVA with memory suppression (Think versus NoThink) and acquisition time (30 min versus 24 h) as within-subject factors was used to examine physiological changes as a function of 4 conditions for cognitive manipulations on aversive memories.

Imaging acquisition

Whole-brain imaging data was collected on a Siemens TRIO 3-Tesla MR scanner in the National Key Laboratory of Cognitive Neuroscience and Learning at Beijing Normal University. Functional images were collected using an echo-planar imaging sequence (axial slices, 33; slice thickness, 4 mm; gap, 0.6 mm; TR, 2000, ms; TE, 30 ms; flip angle, 90°; voxel size, 3.1 × 3.1 × 4.0 mm; flip angle, 90°; FOV, 200 × 200 mm; and 580 volumes), while structural images were acquired through three-dimensional sagittal T1-weighted magnetization-prepared rapid gradient echo (192 slices; TR, 2530, ms; TE, 3.45 ms; slice thickness, 1 mm; voxel size, 1.0 × 1.0 × 1.0 mm³; flip angle, 7°; inversion time, 1100, ms; FOV, 256 × 256 mm).

Imaging preprocessing

Brain imaging data was preprocessed using Statistical Parametric Mapping (SPM8; http://www.fil.ion.ucl.ac.uk/spm). The first 4 volumes of functional images were discarded for signal equilibrium and participants’ adaptation to scanning noise. Remaining images were corrected for slice acquisition timing and realigned for head motion correction. Subsequently, functional images were co-registered to each participant’s gray matter image segmented from corresponding high-resolution T1-weighted image, then spatially normalized into a common stereotactic Montreal Neurological Institute (MNI) space and resampled into 3-mm isotropic voxels. Finally, images were smoothed by an isotropic three-dimensional Gaussian kernel with 4 mm full-width at half-maximum. The data were statistically analysed under the framework of general linear models (GLM)⁶².

Univariate GLM analysis

To assess transient neural activity associated with memory retrieval (that is, Think trials) and suppression (that is, NoThink trials) for newly acquired and overnight aversive memories, separate regressors of interest were modelled for four experimental conditions (see above) and convolved with the canonical hemodynamic response function (HRF) at the first level. In addition, each participant’s motion parameters from the realignment procedure were included to regress out effects related to head movement-related variability. The analyses included high-pass filtering using a cutoff of 1/40 hz to remove high frequency noise⁴, global intensity normalization and corrections for serial correlations using a first-order autoregressive model (AR(1)). Relevant contrast parameter estimate images were initially generated at the individual-subject level, and then submitted to a 2 (Suppression) by 2 (Time) repeated-measures ANOVA for a second-level group analysis treating participants as a random variable. Significant clusters were identified from the group analysis, initially masked using a gray matter mask, and then determined using conservative and well-accepted statistical criteria—that is, a height threshold of P<0.01 and an extent threshold of P<0.05 with family-wise error corrections for multiple comparisons based on nonstationary suprathreshold cluster-size distributions computed using Monte Carlo simulations⁶³.

To further investigate specific neural activity associated with suppression-induced voluntary or intentional forgetting (that is, NTf) and incidental forgetting (that is, Tf), we conducted an additional GLM analysis by including memory Status (forgotten versus remembered, or f versus r) as another variable of interest, similar to Anderson et al.¹¹, together with Time (30 min versus 24 h) and Suppression (NoThink versus Think, or NT versus T). This analysis included eight regressors of interest (that is, NTf_30 min, NTr_30 min, Tf_30 min, Tr_30 min, NTf_24 h, NTr_24 h, Tf_24 h, and Tr_24 h). We used random sub-samplings from the 30 min conditions to artificially match the number of items in the 30 min and 24 h conditions being used for the first level individual analysis. This allowed us to compare neural activity in successful suppression trials between newly acquired and overnight memories with similar statistical power (that is, [NTf_24 h—Tf_24 h] versus [NTf_30 min—Tf_30 min]). Four participants were excluded from this analysis due to lack of at least one forgotten item in each condition. Relevant parameter contrasts were then submitted to 2 (Status: forgotten versus remembered) by 2 Time (Time) by 2 (Suppression) repeated-measures ANOVA for a second-level group analysis treating participants as a random variable. All other settings were same as the above GLM analysis.

To better characterize hippocampal and prefrontal engagement in memory suppression, we performed complementary ROI analyses separately for the left and right entire hippocampus and the middle frontal gyrus (referred to as DLPFC) anatomically defined using the WFU PickAtlas toolbox⁶⁴. Parameter estimates (or β-weights) associated with conditions of interest were extracted from the above anatomically defined ROIs as well as significant clusters in the MTL and PFC regions at the individual level using MarsBar (http://marsbar.sourceforge.net/) and averaged across voxels within each ROI, then plotted in bar graphs for visualization purposes only.

Task-dependent functional connectivity analysis

We examined hippocampus-based functional connectivity changes via PPI analysis⁶⁵. The hippocampal seed was separately defined as a 4-mm sphere centered at the local peak of corresponding clusters showing significant interaction effects between Suppression and Time in the univariate GLM analysis. To accommodate more than two experimental conditions within same model, we employed a generalized form of task-dependent PPI (gPPI)⁶⁶ The physiological activity of given hippocampal seed region was computed as the mean time series of all voxels. They were then deconvolved to estimate neural activity. Next, four PPI regressors, corresponding to each task regressor from the individual level were obtained by multiplying the estimated neuronal activity from the seed region with a vector coding for effects of each condition, forming four psychophysiological interaction vectors. They were further convolved with a canonical HRF to form four PPI regressors of interest. Task-related activations were also included in this GLM to remove out the effects of common driving inputs on brain connectivity.

Contrast images corresponding to PPI effects at the individual-subject level were then submitted to a 2 (Suppression) by 2 (Time) repeated-measures ANOVA for a second-level group analysis. Similar to the univariate GLM analysis above, significant clusters were initially masked by a gray matter mask, and then determined using a height threshold of P<0.01 and an extent threshold of P<0.05 with family-wise error correction for multiple comparisons based on nonstationary suprathreshold cluster-size distributions⁶³.

To investigate brain functional connectivity patterns associated with intentional forgetting and incidental forgetting between 30-min and 24-h conditions, we conducted an additional hippocampal-seeded PPI analysis by taking memory status into account, with a particular focus on NoThink trials that were later forgotten in the post-scan testing phase (ie, [NTf_24 h—Tf_24 h] versus [NTf_30 min—Tf_30 min]). The seed voxels for the connectivity analysis were chosen around the peak coordinate of the hippocampal cluster with a 4-mm sphere of voxels identified from the additional GLM analysis while taking memory status into account. Other settings were same as the above PPI analysis.

Prediction analysis

We employed a machine-learning approach with balanced fourfold cross-validation to mitigate shortcomings of conventional regression models and test for generalizability of the established relationship to out-of-sample individual subjects⁶⁷. For example, we entered memory suppression scores for each individual as dependent variable, and hippocampal activation as independent variable. Then, we estimated r_{(predicted, observed)} to measure how well the hippocampal activations predict the memory suppression scores using a balanced fourfold cross-validation procedure. Data were divided into four folds, and a linear regression model was built using three folds, leaving one fold out. A final r_{(predicted, observed)} was computed based on the average of four repetition of this procedure. Finally, we used a nonparametric testing approach to test for the statistical significance of the model by generating 1,000 surrogate data sets under the null hypothesis of r_{(predicted, observed)} (ref. 68). The statistical significance (P value) of the model was determined by measuring the percentage of generated surrogate data that are greater than the r_{(predicted, observed).}

Multivoxel pattern dissimilarity analysis

To assess multivoxel pattern dissimilarity associated with newly acquired and overnight aversive memories, we modelled each item (collapsing across four repetitions) as a separate regressor, convolved with a canonical HRF implemented in SPM8. This resulted in 36 regressors in total and 9 regressors for each condition. Contrast images for each item versus fixation, generated at the individual level analysis within each condition were then submitted to subsequent inter-item multivariate pattern dissimilarity analysis for the hippocampal ROIs as well as for the whole brain.

ROI-based pattern dissimilarity analysis

For each of four experimental conditions, we extracted voxel-wise brain activation estimates for each item within the same condition from the defined ROIs, and reshaped them into a single dimensional vector for each ROI. Pairwise correlations were then computed among distributed voxels of each ROI, resulting in N × (N−1)/2 pairwise correlation coefficients, with N representing the number of items in each condition. The dissimilarity score was determined by Fisher’s Z transformation of 1 minus the correlation coefficient, separately for each participant^24,25,26. The data were then submitted to a 2-by-2 repeated-measures ANOVA with Suppression and Time as within-subject factors for the second-level analysis to investigate differences in pattern dissimilarity between retrieval and suppression of newly acquired and overnight-consolidated aversive memories.

Whole-brain pattern dissimilarity analysis

We further implemented a searchlight method to measure inter-item multivoxel pattern dissimilarity at the whole brain level^23,24, using a 6-mm spherical region of interest⁶⁹. As with the ROI-based analysis, we computed the inter-item multivoxel pattern similarity for each condition within each searchlight. The analysis was then repeated for a searchlight centered on every voxel in the brain. Searchlight maps for all four conditions were then entered into a 2-by-2 ANOVA with Suppression and Time as within-subject factors on the second level group analysis to determine changes in pattern dissimilarity between retrieval and suppression of newly acquired and overnight-consolidated aversive memories. Significant clusters were identified using a height threshold of P<0.01 and an extent threshold of P<0.05 with family-wise error corrections⁶³.

To further examine the relationship between memory suppression and neural representation patterns, we performed separate regression analyses for the whole-brain pattern dissimilarity maps with suppression-induced forgetting scores in newly acquired or overnight consolidation condition as a covariate of interest. Parallel analyses were also conducted to examine the relationship between hippocampal pattern dissimilarity and DLPFC engagement in the suppression of either newly acquired or consolidated aversive memories, separately. The significant clusters were determined using the same criterion from the above GLM and gPPI analyses.

Estimates of effect size and post-hoc statistical power

Effect sizes for ANOVAs are partial eta squared, referred to as . For paired t-tests, we calculated Cohen’s d using the mean difference score as the numerator and the pooled s.d. from both repeated measures as the denominator⁷⁰. This effect size is referred to in the text as d_av, in which the ‘av’ refers to the use of the average s.d. in the calculation. The post-hoc statistical power was calculated based on the given type I error rate (α=0.05), the corresponding sample size and effect size.

Data availability

The data and codes that support the findings of this study are available from the corresponding author on request.