The overall trial structure.

Trials followed a fixed sequence consisting of a 2-s appraisal cue (“Look” or “Decrease”), a 7-s image-viewing period, and a 4-s negative-affect rating. Depending on the condition, participants either viewed neutral images passively (“Look neutral”), viewed aversive images passively (“Look negative”), or used cognitive reappraisal to downregulate negative affect in response to aversive images (“Regulate negative”). With TR of 2s, accounting for the hemodynamic delay of the BOLD response (∼3 TRs to peak), the cue, stimulus, and rating periods are expected to be most strongly represented at approximately TRs 3–4, 4–7, and 8–10.

Temporal dynamics of amygdala subregions during emotional picture viewing and reappraisal, estimated by finite impulse response (FIR) modeling.

BOLD responses were modeled with an FIR basis set spanning 30 s (15 TRs; TR = 2 s) for each of three task conditions. Curves show the group-mean FIR response (± SEM) for Look Neutral (green), Look Negative (blue), and Regulate Negative (red). The time axis is referenced to cue onset: TR 0 = cue onset and TR 1 = stimulus (picture) onset. The colored bar beneath each time course marks the trial structure: cue (0–1 TR), stimulus presentation (1–4.5 TR), and rating (4.5–6.5 TR). Horizontal bars above each time course denote intervals of statistically significant activation, identified by per-time point one-sample t-tests against baseline (paired t-tests for the condition contrast), FDR-corrected across the 15 TRs (Benjamini-Hochberg, q < .05); only positive effects spanning ≥2 consecutive significant TRs are shown. Bar colors match the curves: green = Look Neutral vs. baseline, blue = Look Negative vs. baseline, red = Regulate Negative vs. baseline, and grey = Look Negative vs. Look Neutral. The inverted triangle (▾) on each bar marks the peak TR of that effect. Error bars indicate the standard error of the mean (SEM) across participants. These plotting conventions are used consistently across all time-series figures. Left: Amunts atlas: centromedial (CM), superficial (SF), and laterobasal (LB) amygdala. Right: Nacewicz atlas: central and medial nuclei (CeMe), cortical and anterior amygdaloid area (CoAHA), basolateral and basomedial nuclei (BLBM), and the lateral nucleus (LA). The central insets show the anatomical location of each subregion on a coronal template.

Temporal dynamics of amygdala subnuclei based on FIR analysis, compared directly on the same axis using the Look Negative condition.

A) Amunts atlas, including the Centromedial (CM), Laterobasal (LB), and Superficial (SF) nuclei. B) Nacewicz atlas, including the Central and Medial nuclei (CeMe), Cortical and Anterior Amygdaloid Area (CoAHA), Basolateral and Basomedial nuclei (BLBM), and Lateral Nucleus (LA).

Hemispheric lateralization of FIR-estimated BOLD responses.

Each panel shows group-mean FIR time courses (±SEM) for left (solid blue) and right (dashed red) hemisphere subregions across three conditions: look negative (top), regulate (middle), and look neutral (bottom). The x-axis represents time post-cue onset in TRs (TR = 2s); the y-axis represents BOLD amplitude in arbitrary units. Top) Amunts atlas: Columns correspond to centromedial (CM), superficial (SF), and laterobasal (LB) subregions. Bottom) Nacewicz atlas: Columns correspond to central and medial nuclei (CeMe), cortical and anterior amygdala area (CoAHA), lateral nucleus (LA), and basolateral and basomedial nuclei (BLBM).

Data-driven clustering across the combined amygdala masks from the Amunts and Nacewicz atlases.

A) Spatial locations of k-means-derived clusters of amygdala voxels, shown in comparison with the Amunts and Nacewicz anatomical atlases. B) Mean FIR time series for each data-driven cluster. C) Dice coefficients showing the spatial overlap between data-driven clusters and anatomical amygdala subregions. Asterisks indicate overlaps that were significant relative to a permutation null distribution, with multiple comparisons corrected across the full cluster-by-region matrix using Benjamini-Hochberg FDR at q < 0.05.

Temporal dynamics of amygdala subregions during emotional picture viewing and reappraisal, estimated by finite impulse response (FIR) modeling.

BOLD responses were modeled with an FIR basis set spanning 30 s (15 TRs; TR = 2 s) for each of three task conditions. This analysis reproduces the pooled analysis shown in Figure 2 separately in the two independent cohorts. Region-level dynamics replicated across cohorts. The delayed response peaking after picture offset in CM amygdala reproduced in both cohorts. In the Look Negative condition, CM (Amunts) peaked at TR 8 in AHAB and TR 7 in PIP, and CeMe (Nacewicz) likewise at TR 8 and TR 7, whereas the superficial (SF), laterobasal (LB), cortical (CoAHA), and basolateral (BLBM) subregions peaked early (TR 5) in both samples; the lateral nucleus (LA) was the weakest and most variable region, as in the pooled analysis (peaking at TR 13 in AHAB but TR 7 in PIP, with near-zero AHAB amplitude, 0.09). Group-mean Look Negative evoked-response time courses were highly correlated across cohorts for every subregion except LA (Amunts: CM r = 0.95, SF r = 0.86, LB r = 0.93; Nacewicz: CeMe r = 0.91, CoAHA r = 0.91, BLBM r = 0.93; LA r = 0.57). The Look-Negative-versus-Look-Neutral differentiation of CM/CeMe reproduced robustly in AHAB (CM TRs 4–9; CeMe TRs 4–9) but only at isolated TRs in PIP (CM TRs 7 and 9; CeMe TR 4), while the corresponding differentiation of the superficial/laterobasal and CoAHA/BLBM subregions was clearer in AHAB than in PIP. As in the main analysis, reappraisal (Regulate vs Look Negative) produced no consistent, replicable modulation: the few TRs reaching the FDR threshold in AHAB (CM 3/8/13; CeMe 3/8/9) did not reappear in PIP (CM none; CeMe 7), so no region showed a reproducible reappraisal effect.

Hemispheric lateralization of FIR-estimated BOLD responses.

Each panel shows group-mean FIR time courses (±SEM) for left (solid blue) and right (dashed red) hemisphere subregions across three conditions: look negative (top), regulate (middle), and look neutral (bottom). The x-axis represents time post-cue onset in TRs (TR = 2s); the y-axis represents BOLD amplitude in arbitrary units. Lateralization replicated selectively, in the centromedial subregion. The leftward asymmetry of the Amunts centromedial subregion during negative processing reproduced robustly in both cohorts (CM Look Negative LI: AHAB = +0.194, t = 4.63, p < .001; PIP = +0.162, t = 3.54, p < .001), both bracketing the pooled value of +0.187. The per-TR Left > Right effect likewise reproduced at late TRs in both samples for Look Negative (AHAB: TR 9; PIP: TRs 8–9) and Regulate (AHAB: TRs 9–10; PIP: TR 9), and was absent during Look Neutral in both, as in the pooled analysis. The smaller leftward asymmetry of the superficial subregion reached significance in AHAB (LI = +0.113, t = 2.35, p = .020) but not in PIP (LI = +0.042, p = .43), and the laterobasal subregion was non-lateralized in both cohorts (AHAB LI = +0.032, p = .46; PIP LI = −0.032, p = .52). For the Nacewicz atlas, no subregion showed a laterality index that was significant in both cohorts (CeMe: AHAB p = .076, PIP p = .49; CoAHA: p = .58 / .71; LA reached significance in AHAB only, p = .024 / .68), consistent with the main-text observation that no Nacewicz region was reliably lateralized. Thus the only lateralization effect to replicate across both independent samples was the leftward dominance of the centromedial amygdala — the same subregion carrying the late-sustained temporal profile.

Replication between two datasets on clustering analysis.

Clustering the 958 amygdala voxels on their 45-dimensional FIR feature vectors (k = 4) yielded comparable four-cluster partitions in each cohort. The data-driven partition corresponded to both anatomical atlases well above chance in each sample (AHAB: ARI = 0.258 [Nacewicz] and 0.210 [Amunts]; PIP: ARI = 0.214 and 0.132; all permutation p ≤ .0002), consistent with the pooled values reported in the main text (0.276 and 0.161). The four clusters were themselves reproducible across cohorts: the AHAB and PIP partitions agreed well above chance (ARI = 0.369, permutation p = .0001), and every AHAB cluster had a clear PIP counterpart (cross-cohort Dice = 0.62–0.79; mean best-match Dice = 0.71). Matched-cluster time courses were highly correlated across cohorts for the task conditions (Look Negative: mean r = 0.92, range 0.86–0.97; Regulate: mean r = 0.90), and lower for Look Neutral (mean r = 0.64), as expected given the minimal evoked signal in that condition. Critically, the late-sustained centromedial cluster replicated. In both cohorts, the cluster overlapping the Amunts CM mask did so at an identical Dice of 0.578 (matching the pooled value of 0.58); this same cluster showed the most sustained late Look-Negative-versus-Look-Neutral differentiation in each sample (AHAB: TRs 7–12; PIP: TRs 9–10); and the two cohorts’ centromedial clusters were matched to one another at the highest cross-cohort overlap of any pair (Dice = 0.791), with near-identical temporal profiles (Look Negative r = 0.971). Thus the late-sustained centromedial cluster identified in the pooled analysis emerged independently in each cohort.