Time-dependent neural arbitration between cue associative and episodic fear memories

Abstract

After traumatic events, simple cue-threat associative memories strengthen while episodic memories become incoherent. However, how the brain prioritises cue associations over episodic coding of traumatic events remains unclear. Here, we developed an original episodic threat conditioning paradigm in which participants concurrently form two memory representations: cue associations and episodic cue sequence. We discovered that these two distinct memories compete for physiological fear expression, reorganising overnight from an overgeneralised cue-based to a precise sequence-based expression. With multivariate fMRI, we track inter-area communication of the memory representations to reveal that a rebalancing between hippocampal- and prefrontal control of the fear regulatory circuit governs this memory maturation. Critically, this overnight re-organisation is altered with heightened trait anxiety. Together, we show the brain prioritises generalisable associative memories under recent traumatic stress but resorts to selective episodic memories 24 h later. Time-dependent memory competition may provide a unifying account for memory dysfunctions in post-traumatic stress disorders.

Fig. 1. Illustration of the task design.

a Three sound cues (a/b/c) were presented in three temporal sequences. The first sequence was followed by a car crash (unconditioned stimulus; US), serving as CS+_sequence. The second sequence was never followed by US but element-wise shared the last conditioned-cue ‘c’ with CS+_sequence, hence serving as CS+_element. The third sequence was never followed by US and did not share the conditioned last cue as the CS+_sequence, hence serving as CS−. b The experimental sessions across 2 days. c Schematics of the naturalistic traffic scene played in each trial with and without a car crash (US). The car-truck appears in both types of trials to initiate the US anticipatory epoch. CS conditioned sequence, US unconditioned stimulus. Source images for the car interior and car-truck were purchased from amanaimage and 123RF databases with commercial licences, respectively. The traffic road image was taken from Wikipedia (https://ja.wikipedia.org/wiki/ファイル:四つ橋筋江戸堀1交差点.JPG) with a Creative Commons licence. On the traffic image, a car interior image, a truck car image, and a small yellow dot were superimposed. The sound of a crow call and a standard Japanese traffic light sound effect were downloaded from https://taira-komori.jpn.org/animals01.html and https://necobit.com/necobido/, respectively. We used bicycle bells, background noise, and car crash sounds from the BBC Sound Effects Archives (https://sound-effects.bbcrewind.co.uk/) with commercial licence restrictions. Thus, for demonstrative purposes, in Supplementary Movies 1–4, the sounds of bicycle bells and background noise were replaced with similar commercial licence-free sound files downloaded from https://vsq.co.jp/plus/, the car crash sound was replaced with a similar commercial licence-free file from https://soundeffect-lab.info/. See Supplementary information for the full image and sound credits.

Fig. 2. Skin conductance response (SCR) to conditioned sequences (CS).

a Mean ΔSCR across the two CS+ sequences (sequence and element), after subtracting SCR to CS−. ns: non-significant, **P = 0.009 (two-sided Wilcoxon signed-rank test against mean 0). b Mean ΔSCR to each CS+ (sequence or element), after subtracting SCR to CS−. Interaction: *P = 0.032 (linear mixed effects model); other tests: ⁺P_FDR = 0.094, *P_FDR < 0.05 (two-sided Wilcoxon signed-rank test, FDR corrected for multiple comparisons). Note that in both (a) and (b), ΔSCR larger than 0 at P < 0.1 is colour-coded dark for illustrative purposes. Error bars indicate standard errors of means with N = 42 individual human participants, in both (a) and (b). SCR skin conductance response, CS conditioned sequence. Source data provided as a Source Data file.

Fig. 3. Decoding sequence information from multivoxel activity patterns in HPC and DLPFC.

a The decoder was trained, within each individual, with the averaged activation patterns during the sound cues epoch of the CS+_sequence and CS+_element from each session. Then, the decoder was tested on left-out data from the session’s sound cues epoch, as well as on data from the US anticipatory epoch. TR (repetition time) denotes each fMRI scan with a 2-s interval. Source images for the car interior and car-truck were purchased from amanaimage and 123RF databases with commercial licences. The traffic road image was taken from Wikipedia (https://ja.wikipedia.org/wiki/ファイル:四つ橋筋江戸堀1交差点.JPG.JPG) with a Creative Commons licence. On the traffic image, a car interior image, a car-truck image, and a small yellow dot were superimposed. See Supplementary information for the full image and sound credits. b The accuracy in testing the decoders on the same sound cues epoch (left), or the US anticipatory epoch (right) from the same session. Dashed lines indicate theoretical chance-level decoding accuracy. Error bars indicate standard errors of means with N = 41. Open circles indicate individual participants’ data points. ns: non-significant; Interaction: *P = 0.026 and main effect of ROI: ***P = 0.0005 (linear mixed effects model); between-ROI comparisons: ⁺P_FDR = 0.05, *P_FDR = 0.048, **P_FDR = 0.0017 (two-sided Wilcoxon signed-rank test, FDR corrected for multiple comparisons); tests against chance: *P_FDR < 0.05, **P_FDR < 0.01, ***P_FDR < 0.001 (two-sided Wilcoxon signed-rank test against mean 0.5, FDR corrected for multiple comparisons). CS conditioned sequence, US unconditioned stimulus, HPC hippocampus, DLPFC dorsolateral prefrontal cortex. Source data provided as a Source Data file.

Fig. 4. Information transmission MVPA to assess communication of sequence representations between brain areas.

a Graphic illustration of information transmission analysis. In the illustration, activity patterns in the VMPFC-Amygdala combined ROI (target area) are used to predict the trial-by-trial CS+ sequence likelihood (CS+_sequence versus CS+_element) harboured in the DLPFC or HPC (seed area). b The advantage of DLPFC relative to HPC in transmitting CS+ sequence information with the VMPFC-Amygdala circuit emerges selectively at Long-term test on Day 2. Interaction: *P = 0.023 (linear mixed effects model), between-session comparisons (HPC): **P_FDR < 0.005; between-ROI comparison (Long-term test): ***P_FDR = 0.00038 (two-sided Wilcoxon signed-rank test, FDR corrected for multiple comparisons). Error bars indicate standard errors of means with N = 41. Larger colour squares represent means where faded colour indicates non-significant (chance level) information transmission. Small coloured squares represent individual participants’ data. c The correlation between the seed difference (DLPFC versus HPC) in CS+ information transmission (CS+_sequence versus CS+_element) with VMPFC and trait anxiety was significant (Pearson’s r = −0.49, P = 0.0012, robust regression slope β = −12.4, P = 0.0003). The same correlation was absent with the Amygdala (Pearson’s r = 0.025, P = 0.88, robust regression slope β = 0.36, P = 0.92) with a significant difference from the correlation with VMPFC (r-test z = 2.45, P = 0.015). n.s. non-significant, *P < 0.05, **P < 0.01 (two-sided). CS conditioned sequence, HPC hippocampus, DLPFC dorsolateral prefrontal cortex, VMPFC ventromedial prefrontal cortex, Amyg Amygdala. Source data provided as a Source Data file.

Fig. 5. The role of HPC and DLPFC in transmitting CS+ versus CS− sequence information via VMPFC.

a The analysis used activity patterns in VMPFC to predict the trial-by-trial CS+_element likelihood (versus CS−) harboured in the DLPFC or HPC seed. The network was constantly governed by DLPFC across days, while HPC withdrew on Day 2. Error bars indicate standard errors of means with N = 41. Between-session comparison: ⁺P_FDR = 0.074; between-ROI comparison: **P_FDR = 0.0041 (both two-sided Wilcoxon signed-rank test, FDR corrected for multiple comparisons). b Same as (a), except that it assessed information transmission of CS+_sequence versus CS− representations. The advantage of DLPFC relative to HPC in transmitting sequence information with VMPFC was constant. Larger colour circles represent means where faded colour indicates non-significant (chance level) information transmission. Small coloured circles represent individual participants’ data. Error bars indicate standard errors of means with N = 41. Main effect: **P = 0.0017 (linear mixed effects model); between-ROI comparisons: *P_FDR = 0.024, **P_FDR = 0.005 (two-sided Wilcoxon signed-rank test, FDR corrected for multiple comparisons). CS conditioned sequence, HPC hippocampus, DLPFC dorsolateral prefrontal cortex, VMPFC ventromedial prefrontal cortex. Source data provided as a Source Data file.

Fig. 6. DLPFC-VMPFC communication of CS+_element representation negatively correlates with SCR towards CS+_element on Day 2.

The correlation between the DLPFC-VMPFC information transmission of CS+_element (versus CS−) with ΔSCR to CS+_element (versus CS−) was absent in Acquisition on Day 1 (Pearson r = 0.17, P = 0.29, robust regression slope β = 0.022, P = 0.43) but emerged in Long-term test on Day 2 (Pearson r = −0.37, P = 0.018, robust regression slope β = −0.064, P = 0.035). The difference in the correlations between Day 1 (Acquisition) versus Day 2 (Long-term test) was significant (r-test z = 2.42, P = 0.015). n.s. non-significant, *P < 0.05 (two-sided). SCR skin conductance response, CS conditioned sequence, DLPFC dorsolateral prefrontal cortex, VMPFC ventromedial prefrontal cortex. Source data provided as a Source Data file.