Stimuli, experimental procedure and evoked responses in Experiment 1

(A) Participants were presented with RDKs in three conditions. In the full sequence trial condition, four successive RDKs were presented. In the start-or end-only condition, only the first or last RDK in the sequence was presented at the beginning of the trial. (B) During MEG scanning, the participants were presented with two functional localizer runs (i.e., functional localizer phase) before providing any sequence information. Next, they were exposed with four full sequence runs (i.e., exposure phase). Finally, in the main phase, full sequence, start-and end-only trials were presented in a pseudo-randomized order. (C) Evoked brain responses as a function of time relative to trial onset in the full sequence, start-and end-only trial conditions, respectively. Bold black lines at the bottom indicate temporal clusters in which they reached significance when compared to the pre-stimulus baseline. Inset figures at the top-right corner show the peak amplitudes during the four corresponding RDK intervals after baseline correction.

Time-resolved decoding probability for each motion direction

(A) Visualization of the time-resolved decoding probability in the three trial conditions. Each row shows the decoding probabilities for the four motion directions at that time point and each column indicates one of the four motion directions. (B) The line plots of the time-resolved decoding probability for the three conditions. Each colored line shows the time course of the decoding probability for each motion direction. For the start-and end-only conditions, we calculated the permutation threshold estimated by randomly shuffling of the labels and re-decoding, only the decoding probability of the cue surpassed the threshold.

Illustration of replay analysis pipeline: Temporally delayed linear modeling (TDLM)

(A) The decoding probabilities were aligned at 200 ms post-stimulus onset according to their corresponding optimal time point per participant and motion direction. Each colored line indicates the decoding results of a classifier applied to a motion direction dataset. The dashed horizontal line indicates the permutation threshold estimated by random shuffling of the labels and re-decoding. (B) Regression models were trained for each participant and motion direction using MEG signals from the functional localizer data. (C) Classifiers were next applied to MEG signals during the post-cue blank period to derive a decoded [time x state] reactivation matrix. An example of backward sequential reactivation of stimuli is shown on the left. (D) Using TDLM, we quantified the evidence for sequential replay of motion directions during the post-cue blank period in start-and end-only conditions. We initially performed a time-lagged regression to create a regression coefficient matrix [states x states] for each time lag by regressing each lagged predictor matrix X(ι1t) onto the state reactivation matrix, Y (i.e., 1st level GLM analysis). The degree to which the reactivations systematically follow a transition matrix of interest was measured by “sequenceness” (i.e., 2nd level GLM analysis). We tested the magnitude of this effect at each time lag independently for all transition lags up to 600 ms. The dashed line represents the corrected nonparametric statistical significance threshold (see Methods for details). The green area indicates the lags when the evidence of sequenceness in the backward direction exceeded the permutation threshold.

Backward replay in both start-and end-only conditions

(A and D) Examples of backward sequential reactivation in start-(A) and end-(D) only conditions from a representative participant. Each row represents the reactivation probabilities for the four motion directions at that time point and each column indicates one of the four motion directions. (B and E) Backward replay of the exposed motion sequence with peak transition lags at 28–40 ms in the start-only condition (B) and at 28 – 36 ms in the end-only condition (E). Horizontal dashed lines represent corrected significance levels from a nonparametric permutation test at the 2nd-level GLM analysis of TDLM. (C and F) Backward replay of the exposed motion sequence predominantly appeared at 1.2 – 1.8 s in the start-only condition (C) and 0.6 – 1.8 s in the end-only condition (F) after the onset of the blank period.

Backward replay is cue-dependent and depends on the amount of exposure

(A) No evidence for replay was found in either the 2nd-only (left) or 3rd-only (right) condition in Experiment 2. Horizontal dashed lines have the same meaning as those in Figure 4B and 4E. (B) In Experiment 3, immediately after the functional localizer phase, participants entered the main phase without the exposure phase. In the end-only condition, backward replay was observed; however, in the start-only condition, no such replay was observed.

Replay events align with ripple power, with source activation in the MTL preceding activation in visual cortex

(A) Top: In Experiment 1, a time-frequency decomposition of sensor-level data revealing a brief increase in high-frequency oscillatory power at replay onset. Bottom: a cluster-based permutation test (cluster forming threshold, t > 3.1; number of permutations = 5000) could identify a significant cluster around 140 Hz. (B) Source localization of ripple-band power 30 ms before replay onset showing significant activation in the MTL (peak Montreal Neurological Institute [MNI] coordinates: X = 21, Y = 21, Z = 13. Neurological orientation.) (C) Source localization of ripple-band power at replay onset showing significant activation in visual cortex (peak MNI coordinates: X = 20, Y = 8, Z = 17). (D) The activation time course of the MTL at its peak MNI coordinate is shown in red, whereas that of visual cortex at its peak MNI coordinate is displayed in green. The MTL reached its peak activation before visual cortex.

Sensor maps and spatial correlation of trained Lasso logistic regression models (related to Figure 3)

(A) Left: Sensor map for each state decoding model in Experiment 1; MEG signals from 72 occipital sensors were selected as features while training the classifier. Right: Correlation matrix among classifiers. No spatial correlation was found among trained classifiers (highest correlation; r < 0.12). (B and C) Same as (A), except the sensor maps and correlation matrix correspond to Experiments 2 (Panel B, highest correlation; r < 0.1) and 3 (Panel C, highest correlation; r < 0.1), respectively.

Decoded feature representations in start-and end-only conditions (related to Figures 3)

Time series of reactivation probability output from the four regression models on an example trial. Time zero corresponds to trial onset. Both participants S08 and S12 were exposed with a fixed motion sequence of 0° → 90° → 180° → 270°.

Sequenceness distribution across participants (related to Figures 4 and 5)

(A) In Experiment 1, of the 18 participants, 14 showed backward replay at 32 ms-lag for the start-only condition (left), and 13 showed backward replay at 32 ms-lag for the end-only condition (right). Each dot represents an individual participant. Backward and forward sequenceness is denoted by blue and yellow dots, respectively. The inset histogram shows the distribution of deviations from the unity line. (B) Same as described in (A), but for the results of Experiment 2 in the 2nd-only (left) and 3rd-only (right) conditions, respectively. (C) Same as described in (A) but for the results of Experiment 3 in the start-(left) and end-only (right) conditions, respectively.

Sequenceness for each of the 24 possible orders (related to Figure 4)

(A) In the start-only condition of Experiment 1, only the backward sequence [4 → 3 → 2→ 1] with lags at 28–40 ms was significantly different (gray panels). The horizontal dashed lines represent significance thresholds derived from state label permutation at the 2nd-level GLM matrix of temporally delayed linear modeling. (B) Same as described in (A) but for the end-only condition in Experiment 1.

Stimuli and experimental procedures in Experiments 2 and 3 (related to Figure 5)

(A) We presented the second (2nd-only condition) or third RDK (3rd-only condition) as a cue in Experiment 2. (B) The procedure of Experiment 3 was identical to that of Experiment 1 except for removing the exposure phase.