Bed nucleus of the stria terminalis connectivity during food cue and taste processing under stress

Abstract

Across species, stress drives alterations in feeding behaviour, including heightened food-seeking and the overconsumption of palatable foods. The bed nucleus of the stria terminalis (BNST) acts as a neural hub linking stress and reward circuits. However, its role in human food cue and taste processing under stress remains unclear. Here, using 7-Tesla fMRI, dynamic causal modelling and in-scanner taste delivery, we demonstrate that beverage cues and taste receipt under stress modulate BNST effective connectivity. Forty-eight participants were presented a palatable and neutral beverage cue before receiving the corresponding beverages under low- and high-stress conditions. Beverage cues under stress downregulated BNST effective connectivity to the nucleus accumbens, orbitofrontal cortex, and dorsal mid-insula (dmINS), with the strength of cue-related downregulation of BNST-to-orbitofrontal effective connectivity related to changes in participants’ subjective stress. In addition, taste receipt under stress downregulated effective connectivity from the dmINS to BNST. These findings provide evidence that stress-related food cue and taste processing modulates BNST reward circuitry, offering a mechanistic perspective on how stress alters responsivity to food cues.

Inhumans, acute stress may alter food-related signalling of the bed nucleus of the stria terminalis (BNST). Here, authors show that 7-Tesla fMRI reveals stress-related changes in BNST connectivity during food cue presentation and taste delivery.

Introduction

Many individuals report increasing their food intake in response to stress^1,2, and stress is associated with eating pathology, particularly binge eating³. Though the neural mechanisms remain unclear, stress enhances responsivity to food cues, motivates food-seeking behaviour^4,5, and shifts preferences towards palatable food⁶ irrespective of weight status⁷. Furthermore, stress can increase cue-triggered craving of sweet rewards in humans⁸ and attenuate perceived taste intensity^9,10, which may in turn influence sweet sensitivity—an individual difference associated with greater liking of sweet foods¹¹. Rewarding cues associated with high-calorie foods, which are abundant in contemporary food environments, can also trigger appetitive drive in the absence of satiety¹², and promote the overconsumption of food under stress¹³. Thus, in this way, the stressed brain appears to be biased towards an increased desire to eat, coupled with a reduced ability to inhibit feeding behaviour, which is precipitated by the presence of rewarding food cues¹.

The bed nucleus of the stria terminalis (BNST) serves as a primary hub for integrating internal bodily cues that guide motivated behaviour—such as feeding—in response to threats and rewards^14,15. The BNST is anatomically connected to the amygdala, lateral hypothalamus, striatum, insula and orbitofrontal cortex (OFC)^16–18, regions implicated in stress, eating and reward responses, respectively. In sated mice, optogenetic activation of the mostly GABAergic (inhibitory) neurons projecting from the BNST to the lateral hypothalamus induces rapid consumption of palatable food^19,20. Human BNST connectivity under food cue and taste processing during stress, however, has not been well characterised.

To date, human BNST research has focused primarily on its role in stress, anxiety and drug reward, including stress-induced drug-seeking behaviour²¹. In active smokers, enhanced cue responsiveness to smoking videos is predicted by reduced BNST, nucleus accumbens (NAcc) and insula activation during acute psychosocial stress²². Other activation-based studies implicate the NAcc, insula and OFC in heightened food cue responsiveness under stress^23,24, and in the encoding of food taste^25–27. Evidence from diffusion-weighted MRI, resting-state fMRI and post-mortem studies have confirmed bidirectional white matter tracts between the BNST, NAcc, insula and OFC in humans^28–32, such that these regions form an anatomical network. Within this network, Banasikowski and Hawken¹⁴ have suggested the BNST may play a role in ‘valence surveillance’³³, monitoring and integrating information about the emotional value of stimuli by relaying hunger and satiety signals from the hypothalamus to the OFC, which is involved in goal-directed decision-making.

Mapping the BNST’s function in humans at standard MRI field strengths is challenging due to its small size³⁴. By virtue of its superior spatial resolution, ultra-high-field 7-Tesla (7 T) fMRI overcomes this barrier, permitting the resolution of neural activity in deep brain structures, including the BNST³². Using such approaches, progress has been made with regard to characterising the food cue-related activity of large-scale brain networks^35–40. However, understanding food cue-related interactions within brain networks has been hampered by the field’s reliance on correlational measures (functional connectivity, for instance), rather than methods that can infer effective connectivity—the signed, directed influence that one brain region exerts over another⁴¹, where a positive (negative) connection reflects a net excitatory (inhibitory) influence at the level of neuronal populations. Although not necessarily superior to simpler functional connectivity approaches in all contexts, dynamic causal modelling (DCM) offers a biophysically grounded framework for inferring effective connectivity, and has proven valuable in addressing such questions⁴².

Here, we leveraged 7 T fMRI, DCM and controlled taste delivery via a gustometer to map the effective connectivity between the BNST and key regions implicated in food cue and taste processing: the NAcc, mid-insula (dmINS), and OFC. We hypothesised that food cue and taste processing under stress would downregulate (decrease the values of) BNST connectivity with other network nodes. We further hypothesised that effective connectivity within the BNST network would be influenced by the reward value of beverage cues and taste. Our findings show that beverage cues under stress have a down-regulatory effect on BNST connectivity to the NAcc, OFC and dmINS. In addition, the strength of the high-stress cue-related downregulation of BNST-to-OFC effective connectivity predicted individual-participant changes in subjective stress, providing evidence of BNST-connectivity–behaviour relationships in the context of food cue processing.

Results

Beverage task and stress induction

All participants (n = 48) completed two runs of a block-design beverage task: one following low-stress induction and another following high-stress induction (Fig. 1; detailed in “Methods”). During each task run, participants were shown images of either a glass of chocolate milk or water. Following each cue presentation, participants received the corresponding beverage inside the scanner via a gustometer.

Fig. 1. In-scanner beverage task.

a The functional scans comprised two consecutive runs of the beverage task. Prior to each beverage task, participants underwent a 2-min low- or high-stress induction. During the former, participants were asked to count backwards aloud from a whole number and given positive feedback whilst observing a 2-min countdown on the screen. During the latter, participants were asked to perform difficult serial subtractions aloud and were informed to start again if their answers were too slow or incorrect whilst observing a 2-min countdown. During high-stress induction, participants were given negative feedback on their performance. b Upon completion of each stress induction, participants rated how stressed and hungry they felt on a horizontal 7-point Likert scale using a handheld button box. c Participants were able to reach for a handheld button box, and a 3D-printed mouthpiece connected to an MRI-compatible gustometer with their free right hand. Skin conductance was measured throughout via electrodes attached to the index and middle finger of participants’ left hand. Heart rate was measured throughout via a pulse oximeter attached to the ring finger of participants’ left hand. d During each 7-min beverage task run, images of a clear glass containing either chocolate milk or water were presented to participants on a screen for 5 s. Participants then received 1 ml of the beverage presented via a mouthpiece whilst the screen displayed the word “TASTE” for 5 s. Participants had 5 s to rate (using the button box) how pleasant the beverage tasted (on a 7-point Likert scale), with 1 being very unpleasant and 7 being very pleasant. After ratings, “RINSE” appeared on the screen for 5 s and participants received 1.5 ml of the tasteless solution via the mouthpiece. Each run consisted of 12 blocks (6 blocks of chocolate milk, 6 blocks of tasteless water solution). The order of chocolate milk and water presentation was pseudorandomised. VAS visual analogue scale. Created in BioRender. Brown, R. (2026) biorender.com/mjfumj9; contains iStock.com/Hyrma and iStock.com/wabeno content used under licence.

Self-reported stress levels were higher after the high-stress induction (mean 5.04 ± 1.58 [s.d.]) compared to the low-stress induction (mean 2.27 ± 1.48, t₍₄₇₎ = 12.41, p = 1.94 × 10⁻¹⁶, two-sided, n = 48; Fig. 2a). There was a main effect of stress (F_{(1, 94)} = 15.01, p = 0.0002, two-sided, n = 48) and beverage (F_(1, 94) = 38.91, p = 1.26 × 10⁻⁸, two-sided, n = 48) on mean pleasantness ratings: chocolate milk was rated more pleasant than water after both high-stress (mean 5.26 ± 1.23 versus mean 3.88 ± 1.26, p = 2.85 × 10⁻⁸, two-sided) and low-stress induction tasks (mean 5.61 ± 1.12 versus mean 4.29 ± 1.04, p = 9.00 × 10⁻⁸, two-sided; Fig. 2b, Supplementary Information). Analysis of cue-level pleasantness ratings revealed a main effect of cue exposure (F_{(5, 940)} = 10.17, p = 1.66 × 10⁻⁹, two-sided), an interaction between cue exposure and stress (F_{(5, 940)} = 2.64, p = 0.02, two-sided), and no significant interaction between stress, beverage type and cue exposure (F_{(5, 940)} = 1.007, p = 0.41, two-sided; Fig. 2c).

Fig. 2. Behavioural ratings and group-level whole-brain general linear model activation during food cue and taste processing.

a Mean self-reported hunger and stress ratings after low-stress and high-stress induction tasks, paired t-tests, *p < 0.05, ****p < 0.0001, two-sided (low stress versus high stress). Error bars show standard error of the mean. Exact p values are reported in the main text. b Mean self-reported chocolate milk and water pleasantness ratings after low- and high-stress induction tasks, two-way repeated measures ANOVA (stress × beverage), main effect of beverage, ****p < 0.0001, Bonferroni-corrected post hoc multiple comparisons, paired t-tests (two-sided), *p < 0.05, **p < 0.01 (low-stress versus high-stress). c Time course showing mean self-reported chocolate milk and water pleasantness ratings during each beverage cue exposure, three-way repeated measures ANOVA (stress × beverage × cue), Bonferroni-corrected post hoc multiple comparisons paired t-tests (two-sided) at each cue exposure, ***p < 0.001, ****p < 0.0001 (chocolate milk versus water during high stress); ^{^^^}p < 0.001; ^{^^^^}p < 0.0001 (chocolate milk versus water during low stress); n = 48 for all. Error bars show standard error of the mean. d Group-level general linear model (GLM) results for cue presentation (CUE), n = 48. e Group-level GLM results for taste receipt (TASTE), n = 48. Group-level activation t-maps are overlaid on a standard anatomical template. Colours indicate t-values. Significant activation was thresholded at voxel-wise pFDR < 0.05, one-sided, and a minimum cluster extent of 10 voxels. Dashed elliptical outlines mark regions of interest (OFC, dmINS, BNST and NAcc). Overlaid masks (red = NAcc; yellow = BNST) indicate the regions used for time-series extraction in subsequent analyses (see “Methods”) and are shown for reference only. LS low stress, HS high stress, BNST bed nucleus of the stria terminalis, dmINS dorsal mid-insula, NAcc nucleus accumbens, OFC orbitofrontal cortex.

Bonferroni-corrected post hoc pairwise comparisons using paired t-tests showed that chocolate milk was rated more pleasant than water at all cue exposures during high stress (cue 1: p = 0.0004; cue 2: p = 0.0019; cue 3: p = 1.85 × 10⁻⁵; cue 4: p = 1.21 × 10⁻⁵; cue 5: p = 0.0002; cue 6: p = 6.33 × 10⁻⁵) and at all cue exposures except cue 1 during low stress (cue 2: p = 0.0002; cue 3: p = 1.85 × 10⁻⁵; cue 4: p = 0.0006; cue 5: p = 4.22 × 10⁻⁵; cue 6: p = 1.85 × 10⁻⁵), two-sided, n = 48 for all. These results suggest that (1) the stress manipulation was successful in increasing participants’ self-reported stress levels, and (2) participants rated the rewarding beverage (chocolate milk) as more pleasant than the neutral beverage (water) throughout the beverage task.

Physiological stress was primarily indexed using tonic electrodermal activity (EDA; detailed in “Methods”), complemented by post hoc analyses of heart rate variability (HRV). EDA values were not significantly different after high-stress induction compared to the low-stress induction (t₍₂₇₎ = −0.312, p = 0.758, two-sided, n = 28, mean difference = −0.006 ± 0.11; Fig. S1). HRV, quantified as the root mean square of successive differences (RMSSD), was lower at the beginning of the high-stress run than at the beginning of the low-stress run (t₍₄₃₎ = −2.25, p = 0.029, two-sided), consistent with greater physiological stress during the high-stress condition. We also used sliding windows (10-s windows stepped every 2 s) to analyse the complete time course of RMSSD across both task runs. Data were segmented into early, middle, and late task periods (first, middle, and final 20 s) to evaluate trends across time. A within-participants repeated-measures ANOVA found a significant condition × time interaction (F_{(2, 86)} = 3.20, p = 0.045, two-sided), indicating that RMSSD evolved differently over time between the low- and high-stress conditions (Fig. S2). Uncorrected post hoc comparisons showed that RMSSD increased significantly more from early-to-middle in the low-stress condition compared to the high-stress condition (t₍₄₃₎ = 2.25, p = 0.029, two-sided). Likewise, RMSSD increased in the low-stress condition but decreased in the high-stress condition when comparing early-to-late windows (t₍₄₃₎ = 2.03, p = 0.048, two-sided). Additional short-term HRV indices (pNN20 and pNN50; the proportion of successive inter-beat intervals differing by >20 ms and >50 ms, respectively) and a non-linear Poincaré-derived measure (SD1/SD2; the ratio of short- to long-term variability) showed a consistent pattern indicative of greater physiological arousal during the high-stress run. Mean heart rate averaged across the entire run did not differ between conditions (t₍₄₇₎ = − 0.37, p = 0.710, two-sided). Full results are reported in the Supplementary Information.

Brain activity during beverage cue presentation and taste receipt

A whole-brain, group-level general linear model (GLM) revealed significant activation in response to beverage cue presentation (‘CUE’), with robust effects present in visuospatial and attentional regions including the lingual gyrus, angular gyrus and middle frontal gyrus. Additional significant clusters were found in the dorsolateral prefrontal cortex, mid-cingulate cortex, middle temporal gyrus, and middle frontal gyrus, with the OFC and dmINS showing activation, overlapping with temporal pole clusters. Subcortical activations included the putamen, thalamus, hippocampus, ventral striatum (NAcc) and amygdala extending to the BNST, with ancillary recruitment of the substantia nigra and ventral tegmental area (VTA; pFDR < 0.05, one-sided, n = 48, voxel-wise false-discovery-rate [FDR] corrected; Fig. 2d).

A second group-level GLM analysis revealed significant activation in response to beverage taste receipt (‘TASTE’), with the strongest effects in sensorimotor and gustatory regions, including the postcentral and precentral gyri, supplementary motor area, cerebellum, rolandic operculum and dmINS. Subcortical engagement was notable in the putamen and thalamus. The OFC showed modest engagement, alongside discernible activation in the amygdala extending to the BNST, and ancillary recruitment of the NAcc (pFDR < 0.05, one-sided, n = 48, voxel-wise FDR corrected; Fig. 2e, for complete results, Tables S1 and S2). Shared activation across cue and taste conditions for the DCM volumes of interest (VOIs) are depicted in Fig. S3.

Bed nucleus of the stria terminalis effective connectivity modulation under stress

As suggested by published guidelines for DCM^43,44, the selection of VOIs for analysis was informed by GLM activation results to ensure that VOIs contained task-relevant activity. The dynamic causal model network structure comprised four regions: the BNST, NAcc, dmINS and OFC (Table 1). A priori, we permitted interactions to conform to a star-shaped connectivity architecture, with the BNST as the central hub (Fig. 3). To examine how beverage cue presentation and taste receipt modulates BNST connectivity under stress, for each participant, four separate dynamic causal models were specified and inverted using concatenated time-series data from the low- and high-stress runs. Each model assessed the distinct modulatory effect of: (1) high-stress cues: the modulatory effect of all beverage cue presentations under high stress, (2) high-stress chocolate milk cues: the modulatory effect of rewarding (chocolate milk) beverage cue presentation under high stress, (3) high-stress taste: the modulatory effect of all beverage taste receipt under high stress, (4) high-stress chocolate milk taste: the modulatory effect of rewarding (chocolate milk) beverage taste receipt under high stress. For the ‘CUE’ models (Models 1 and 2), the driving input (to all regions) was specified as all beverage cue onsets across both runs. For the ‘TASTE’ models (Models 3 and 4), the driving input was specified as all beverage taste receipt onsets across both runs. The modulatory inputs on the strength of each connection to and from the BNST were condition-specific: (1) all beverage cue onsets under high-stress, (2) only chocolate milk cue onsets under high-stress, (3) all beverage taste receipt onsets under high-stress, (4) only chocolate milk taste receipt onsets under high-stress. Four participants were excluded from the ‘CUE’ DCM analyses (n = 44) and one from the ‘TASTE’ DCM analyses (n = 47) as they did not survive the BNST VOI time-series extraction (for details, see “Methods”). Results from additional models testing the effects of the low-stress and water conditions are included in the Supplementary Information (Tables S3 and S4). Results from additional over-complete models—specified identically to the main analyses except that the modulatory inputs were coded to represent all four conditions simultaneously—are also reported in the Supplementary Information (Table S5 and Fig. S4). These analyses suggest that modulation is primarily stress-driven, with minimal evidence for stimulus-specific (water versus chocolate milk) effects.

Table 1.

Group peak coordinates for volumes of interest used in dynamic causal modelling

	MNI coordinates
Region	x	y	z
BNST	22	−8	−11
NAcc	10	19	−8
OFC	43	42	−19
dmINS	37	5	10

Time series extraction for the BNST utilised a consensus mask, created by overlaying the manually segmented BNST mask from Theiss et al.¹¹⁰, onto Pauli et al.’s extended amygdala mask¹¹¹. Both masks were binarised and combined to create a new mask containing only voxels common to both original masks. An inclusive mask was created for the NAcc using the 1-mm³ resolution atlas from the MNI standardised automated anatomical labelling 3 toolbox (AAL3¹¹²). The first eigenvariate across all activated voxels within a 4 mm radial sphere around the subject-specific maxima (p < 0.05, uncorrected), but no more than 8 mm from the group maxima, was extracted. Coordinates shown indicate representative peak locations for localisation; BNST and NAcc time series were extracted from bilateral masks, yielding one time series per node, per subject. BNST bed nucleus of the stria terminalis, dmINS dorsal mid-insula, MNI Montreal Neurological Institute, NAcc nucleus accumbens, OFC orbitofrontal cortex.

Fig. 3. Dynamic causal model specification.

a Intrinsic efferent and afferent (A-matrix) connections (black arrows) were inferred between the BNST and other network regions, alongside self-connections (not shown). Task inputs (green arrows) consisted of boxcar regressors for cue presentation or taste delivery across both low- and high-stress task runs, scaled by input gains (C-matrix). b Inferred modulatory (B-matrix) effects were model specific. In cue-focused models, modulators comprised either the combined chocolate milk (rewarding) and water (neutral) cues presented after high-stress induction (Model 1), or high-stress chocolate milk cues only (Model 2). Likewise, in taste-focused models, modulators comprised either the combined chocolate milk and water taste delivery after high-stress induction (Model 3), or high-stress chocolate milk taste delivery only (Model 4). R right hemisphere, dmINS dorsal mid-insula, BNST bed nucleus of the stria terminalis, NAcc nucleus accumbens, OFC orbitofrontal cortex. Figure made with BrainNet Viewer; contains iStock.com/Hyrma and iStock.com/wabeno content used under licence.

Parametric empirical Bayes (PEB) was used to summarise group-level effects from DCM and to assess the association between the cross-condition change in self-reported stress ratings—change in subjective stress—and the modulation of effective connectivity. We found evidence to suggest that high-stress beverage cue presentation downregulated efferent connectivity from the BNST to the NAcc, OFC, and dmINS, as well as efferent connectivity from the dmINS to the BNST (Fig. 4a). When focusing on chocolate milk cues (Fig. 4b), we found that cue presentation under high stress downregulated efferent connectivity from the BNST to dmINS, and upregulated (increased the value of) efferent connectivity from the NAcc to BNST. By contrast, focusing on taste receipt under stress, we found that taste receipt downregulated dmINS-to-BNST effective connectivity (Fig. 4c); however, this effect did not exceed our imposed posterior probability threshold when considering chocolate milk and water taste receipt separately.

Fig. 4. Inferred effective connectivity modulation and change in subjective stress.

Parametric empirical Bayes (PEB) was used to infer group-level modulatory effects on effective connectivity and their association with change in subjective stress. a–d Arrows indicate the direction of modulatory influence (blue, down-regulatory; red, up-regulatory); values adjacent to arrows indicate maximum a posteriori estimate (Hz). a Model 1 (high-stress cues) revealed down-regulatory modulation of BNST efferents to NAcc, dmINS and OFC, and down-regulatory modulation of dmINS-to-BNST connectivity. b Model 2 (high-stress chocolate milk cues) revealed down-regulatory modulation of BNST-to-dmINS connectivity and up-regulatory modulation of NAcc-to-BNST connectivity. c Model 3 (high-stress taste) revealed down-regulatory modulation of dmINS-to-BNST connectivity. d In Model 1, adding change in subjective stress as a covariate in the PEB model indicated that greater change in subjective stress (z-scored) was associated with stronger downregulation of BNST-to-OFC connectivity during high-stress cue presentation. e Leave-one-out cross-validation enabled assessment of out-of-sample prediction accuracy, comparing predicted and observed change in subjective stress for each left-out participant using the high-stress cue modulation of BNST-to-OFC connectivity inferred in the covariate model (previous panel). Red line shows the out-of-sample estimate with 90% credible intervals (grey shading). Dashed orange line shows the observed value. f Out-of-sample association between observed and predicted change in subjective stress during high-stress cue presentation (r = 0.28, p = 0.031; two-sided). BNST bed nucleus of the stria terminalis, dmINS dorsal mid-insula, NAcc nucleus accumbens, OFC orbitofrontal cortex. Contains iStock.com/Hyrma and iStock.com/wabeno content used under licence.

When focusing on both high-stress beverage cues and treating change in subjective stress as a covariate, we found that downregulation of BNST-to-OFC effective connectivity was associated with change in subjective stress such that individuals with greater change in subjective stress, showed less downregulation (Fig. 4d). Using leave-one-out cross-validation, we showed that this effect was robust: across folds, the group level model of BNST-to-OFC effective connectivity could predict the change in subjective stress of left-out participants (r = 0.28; p = 0.031, two-sided; Fig. 4e, f). Posterior probabilities and modulatory connections shown in Fig. 4a–d, are reported in Table 2.

Table 2.

Inferred modulatory effects of high-stress neutral and rewarding cues and taste on dynamic causal model connectivity parameters

Model	Connection	MAP	Var	Pp
High-stress cue	dmINS → BNST	−0.45	0.034	0.94
	BNST → dmINS	−0.34	0.022	0.91
	BNST → OFC	−0.45	0.016	0.99
	BNST → NAcc	−0.28	0.021	0.87
High-stress chocolate milk cue	NAcc → BNST	0.68	0.032	1.00
	BNST → dmINS	−0.28	0.028	0.86
High-stress taste	dmINS → BNST	−0.33	0.021	0.93

Cue-related models were estimated for 44 participants; taste-related models were estimated for 47 participants. The posterior probability (Pp) threshold was set at Pp ≥0.75 (moderately strong evidence), using a two-sided criterion (that is, the posterior probability that the parameter differs from zero in either direction). Continuous-time modulation is framed in terms of rates (Hz), whereby a task condition (for example, high-stress cues) modulates the degree to which activity in an efferent region changes the continuous rate of change in the state of an afferent region. MAP maximum a posteriori estimate, Var posterior variance, BNST bed nucleus of the stria terminalis, dmINS dorsal mid-insula, NAcc nucleus accumbens, OFC orbitofrontal cortex, Pp two-sided posterior probability (parameter ≠ 0).

Discussion

Here, we demonstrate that beverage cues and taste receipt during acute stress modulated BNST effective connectivity with cortico-striatal regions. High-stress beverage cues downregulated BNST connectivity to the NAcc, OFC, and dmINS, in addition to dmINS-to-BNST connectivity. Greater (more positive) change in subjective stress was associated with less BNST-to-OFC downregulation, an effect shown to be robust via leave-one-out cross-validation. During taste receipt under high stress, dmINS-to-BNST connectivity was also downregulated. Together, these findings suggest that acute stress influences BNST connectivity with cortico-striatal and dmINS regions in a manner that depends on both cue valence and individual stress reactivity.

Broadly, this pattern of BNST-to-OFC downregulation under acute stress aligns with frameworks which posit that the BNST delivers bottom-up afferent interoceptive information to cortical regions, such as the OFC, to attribute valence and to guide behaviour⁴⁵. OFC activity has previously been associated with the encoding of reward predictions necessary for valence-based action^46–49 and OFC-mediated shifts in behaviour under stressful conditions⁵⁰ likely involve homoeostatic input from sensory regions, such as the BNST¹⁴, as extended amygdala lesions have been found to result in inaccurate cue-outcome predictions⁵¹. We also found that the strength of BNST downregulation of the OFC predicted increased change in subjective stress during the task. This indicates that BNST-mediated modulation of cortical regions may be sensitive to experiential stress levels, which could, in turn, viably promote reinforced behaviours, such as food approach⁵². Further investigation on the functional role of this pathway using paradigms that compare goal-directed versus habitual food approach under stress is warranted.

Across models, high stress cues downregulated BNST efferents. From the perspective of Banasikowski and Hawken’s¹⁴ ‘valence surveillance’ account, the BNST relays homoeostatic state-based information to valuation (OFC) and reward (NAcc) systems. Stress-related downregulation of these efferents is consistent with state-dependent gating of that relay, which could bias behaviour toward more cue-driven responding under stress⁵³. Participants with greater change in subjective stress showed less BNST-to-OFC downregulation, suggesting individual variability in this gating. Moreover, our right-lateralised activation pattern aligns with evidence that interoceptive–salience computations and reward-valuation signals show stronger right-hemisphere engagement, indicating that our DCM nodes captured this broader functional asymmetry^54,55. Overall, these results suggest that stress modulates—rather than globally increases—BNST participation in valuation and interoceptive circuits.

Cue presentation under stress had a down-regulatory effect on BNST-to-NAcc connectivity. This result complements previous fMRI studies which consistently have shown dampened or unchanged NAcc response towards cues under stress^56–59. Acute stress is also thought to alter reward processing by amplifying the incentive value of associated cues by increasing dopamine secretion in the NAcc, as evidenced in vivo in animal research⁶⁰. This notion is supported by our finding of the NAcc having an upregulatory effect on the BNST specifically during reward cues and aligns with positron emission tomography studies showing increased dopamine release—reflected in reduced dopamine D2 receptor antagonist [¹¹C]raclopride binding potential—to a psychosocial stressor in humans and non-human primates^61–63. The diverging effects of stress on NAcc activity and dopamine levels has been attributed to differences in stressor magnitude, length, and predictability. Mild-to-moderate, acute, or controllable stressors increase dopamine release, whilst intense, chronic, or unpredictable stressors inhibit dopamine release in the NAcc in rodents⁶⁴. This pattern of BNST-to-NAcc downregulation may reflect stress-induced modulation of striatal responsivity, consistent with preclinical models implicating this circuit in the regulation of reward-seeking behaviour under stress. However, given the limitations of human neuroimaging in measuring cellular-level mechanisms, this interpretation is still speculative and grounded in broader preclinical literature on BNST-NAcc circuits.

Presentation of beverage cues under stress had influences on the reciprocal BNST–dmINS pathway: a down-regulatory effect on both BNST-to-dmINS and dmINS-to-BNST connectivity. Structural connectivity between the BNST and insula has been recently confirmed in humans using probabilistic tractography³¹, with the mid-insula subdivision used as our DCM VOI found to have significant connectivity with the BNST, albeit to a lesser extent than anterior insula subdivisions. This pathway is thought to primarily involve information flow from the insula to the BNST, with weak feedback transmitted from the dorsomedial BNST³³. The insula promotes food-seeking in mice via glutamatergic inputs to GABAergic BNST neurons that project to the VTA⁶⁵. Photoactivation of this insula-to-BNST pathway reinforces food approach behaviour in a dopamine-dependent manner in mice⁶⁵. In response to an acute stressor, activity of dmINS-to-dorsal BNST corticotropin releasing factor (CRF)-positive neurons is associated with active struggling behaviour—a physical stress-coping mechanism in rodents⁶⁶. Relatedly, dmINS has been found in recent 7 T research to function as a primary interoceptive hub tightly coupled with subcortical allostatic centres regions, positioning it as a key node linking bodily state encoding with stress-responsive control⁶⁷. Moreover, antagonising CRF receptors in the BNST decreases frustration stress-induced binge eating of palatable food in rodents⁶⁸, and self-administration of palatable food enhances BNST–VTA synaptic strength¹⁵. In contrast, insula-to-BNST projections have previously been found to induce stress-coping behaviours⁶⁶ and palatable food-seeking⁶⁸ in mice and suggests that this top-down pathway may be engaged to promote appropriate stress-coping responses, potentially including appetitive self-soothing. Both the cue and taste phases of the task involved sensory input, oral motor engagement, and expectancy—all features known to recruit insula–BNST pathways^69,70. Considering the insula’s role in integrating interoceptive and sensory signals, its engagement could plausibly reflect general stress-related modulation rather than reward-specific processing, particularly under conditions of altered interoceptive sensitivity induced by acute stress.

This study had several study limitations. First, our study did not control for satiety. Hunger levels and the nutritional composition of meals can influence brain responses to food cues up to 12 h prior to scanning, and future studies would benefit from exploring the impact of these factors⁷¹. Second, we cannot exclude potential order effects with our study design (low-stress always preceded the high-stress condition), which we implemented to minimise carry-over effects per previous work⁷². Future studies should aim to implement washout periods or counterbalanced designs to better isolate stress-specific contributions. Third, the stress induction period was relatively brief (2 min), and although we observed both subjective and physiological signs of stress, our stress induction task has not been validated against changes in cortisol or adrenocorticotropic hormone, which could provide confirmation of hypothalamic–pituitary–adrenal axis engagement^73,74. Finally, although the BNST differs in size and connectivity between sexes and prior work highlights sex and hormonal influences on stress responsivity and gustatory processing^75–78, our study was not designed or powered to test these effects directly.

In conclusion, this study revealed that the BNST has a stress-mediated down-regulatory role on regions mediating food cue and taste processing. These findings advance our mechanistic understanding of stress-induced neural changes to food anticipation and consumption. We believe translational work in clinical cohorts is now warranted: stress is a reliable precipitant of overeating in those with binge eating disorder⁷⁹ and those with obesity^80,81, and improvements in patient outcomes are critically needed.

Methods

Participants

This study was approved by the Human Research Ethics Committee of The University of Melbourne (HREC ID 22347) and conducted in accordance with all relevant ethical regulations. Fifty-seven healthy adults were recruited into the study via online advertisements. Participants were eligible if they (1) were 18–40 years old, (2) had no history of eating disorder symptomology and/or current diagnosed mental health disorder, (3) were fluent in verbal and written English, (4) had no major hearing impairment or damage, (5) were able to consume dairy products, (6) had no contraindications to MRI (for example, pregnancy, claustrophobia or metal present in their body they cannot remove), and (7) had no medical or neurological condition for which they were on medication. Participants were screened for mental health disorders using the DSM-IV/ICD-10 version of the Mini-International Neuropsychiatry Interview (MINI)⁸². Forty-eight participants (mean age 26.2  ±  5.3 years; 24 male, 24 female; self-reported) successfully completed the scanning protocol and were included in GLM analyses (further demographics, Table S6). Forty-four participants were included for the ‘CUE’ DCM analyses and 47 for the ‘TASTE’ DCM analyses. Participants gave their written informed consent before commencing the study at the Melbourne Brain Centre Imaging Unit (The University of Melbourne, Parkville).

Beverage task and stress induction

To investigate the effect of food cue and taste processing under stress on neural connectivity, all participants completed two runs of a block-design beverage task that utilised an MRI-compatible gustometer (Multistimulator OG001, Burghart Messtechnik, Pinneberg, DE) and PsychoPy⁸³ v2024.2.4. Prior to scanning, all participants received training on how to complete the task, were shown the beverage cues, and explicitly told they would receive the matching beverages in the MRI.

Prior to each beverage run, participants completed a structured stress induction task. The first induction (low-stress condition) involved an easy, self-paced serial subtraction task with no evaluative feedback, while the second (high-stress condition) incorporated a difficult, time-pressured serial subtraction task with negative social evaluative feedback, modelled on the Montreal Imaging Stress Task (MIST)^84–87. In both conditions, participants viewed an on-screen digital clock (a simple rotating arm which remained white in colour for the first minute and turned red for the second minute). There is a strong precedent for using brief, in-scanner math-related tasks to induce stress in the context of fMRI^84–87. Further information on the stress induction procedure is available in the SI Methods.

During each 7-min beverage task run, participants were shown an image (‘CUE’ condition) of a glass containing either chocolate milk (rewarding cue) or water (neutral cue) on an MRI-compatible screen for 5 s (Fig. 1). Following each image, participants received 1 ml of the corresponding beverage (‘TASTE’ condition), whilst the word “TASTE” was displayed for 5 s to prompt participants to taste and swallow the beverage. Chocolate milk (Table S7) or a tasteless water solution (25 mM potassium chloride, 2.5 mM sodium carbonate dissolved in water) were delivered via a 3D-printed mouthpiece⁸⁸ (SI Methods). Following beverage receipt, participants had 5 s to rate—using an MRI-compatible button box—how pleasant the beverage tasted on a 7-point Likert scale, with 1 being very unpleasant and 7 being very pleasant. After each rating, “RINSE” appeared on the screen for 5 s, at which point participants received 1.5 ml of the tasteless solution via the mouthpiece to use as a rinse. Each run consisted of 12 blocks (6 blocks of chocolate milk, 6 blocks of tasteless water solution). The order of chocolate milk and water beverage presentation and corresponding delivery was pseudorandomised. Prior to the start of each run, participants rated their hunger and stress levels on a 7-point Likert scale with 1 being not hungry (stressed) and 7 being very hungry (stressed) (Fig. 1). All participants provided responses for beverage, hunger and stress ratings across both beverage task runs.

Behavioural analyses

Self-reported hunger, stress and chocolate milk pleasantness ratings were tested for normality using the Shapiro–Wilk method. Hunger and stress levels after low- versus high-stress induction tasks were analysed using two-sided paired t-tests. To assess overall beverage pleasantness, mean ratings of chocolate milk versus water were analysed with a two-way repeated measures ANOVA (stress × beverage), followed by Bonferroni-corrected paired t-tests for post hoc comparisons. To examine cue-specific pleasantness dynamics, ratings of each beverage during each cue (taste) block of the low- and high-stress runs were analysed with a three-way repeated measures ANOVA (beverage × stress × cue) with Bonferroni-corrected paired t-tests at each cue exposure. All post-hoc tests were two-sided and n = 48 for all comparisons. Data were analysed and visualised in GraphPad PRISM v10.2.2 (GraphPad Software, Boston, MA, US). Demographic data were analysed using SPSS v29.0.2.0 (IBM Corp., Armonk, NY, USA, 2023).

Physiological measures

EDA data were recorded during both low- and high-stress induction tasks. These data were recorded using MRI-compatible finger electrodes (Ag/AgCl) coated in conductance gel (0.5% saline) and placed on the index and middle finger intermediate phalanges of participants’ left hands. The signal was amplified and sampled at 1000 Hz using PowerLab version 8.0 (AD Instruments).

EDA data underwent manual review for quality by two independent raters (E.G.-H. and M.D.G.) both before and after preprocessing, which included filtering and downsampling. Filtering was performed using a first-order Butterworth bandpass filter with cutoff frequencies set at 0.0159 Hz and 5 Hz. Following filtering, the signals were downsampled to a target rate of 10 Hz. Raters classified files as acceptable, discardable, or requiring the imputation of a missing event marker. Files exhibiting pervasive artifacts, missing both start and end markers, or flat responses (indicative of non-responses) were excluded from further analysis. The inter-rater reliability for participant inclusion decisions was 78.3%, and after resolving discrepancies between raters, data from 28 participants with recordings for both stress-induction tasks remained.

To infer tonic sympathetic arousal from the spontaneous skin conductance fluctuations occurring during both stress-induction tasks, a matching pursuit algorithm⁸⁹ was applied using the psychophysiological modelling (PsPM) toolbox^90,91 (version 6.1.2). The algorithm was configured to model sympathetic arousal without applying additional filtering beyond the initial preprocessing, and a threshold of 0.1 µS was set for peak detection. A paired two-sided t-test was used to compare sympathetic arousal between runs.

The RMSSD in pulse signals extracted using the PhysIO toolbox were used as a measure of HRV during the low- and high-stress conditions. Pulse oximetry was only collected and synchronised during functional image acquisition, meaning HRV estimates could only be derived from the beverage task runs. In accordance with previous research^92,93, a window of 20 s with 2-s steps was used to assess RMSSD differences at the beginning of each run. As stress responses unfold dynamically within a run, collapsing RMSSD across the entire task can obscure periods of heightened arousal and subsequent recovery, whereas computing RMSSD in shorter sliding windows provides the sensitivity needed to detect these transient physiological shifts. Indeed, mean RMSSD did not differ when examining the entirety of low- and high-stress runs (t₍₄₃₎ = − 0.60, p = 0.55, two-sided), whereas significant differences did emerge using a sliding-windows approach.

RMSSD values were extracted continuously across the task period and subsequently segmented into early, middle, and late phases, defined as the first, middle, and final 20 s of the task, respectively. These values were entered into a within-participants repeated-measures ANOVA with condition (low-stress, high-stress) and time (early, middle, late) as factors. This design follows recommendations to model HRV as a time-varying signal, allowing condition effects to be tested while capturing the distinct phases of autonomic reactivity within each run^94,95. Post hoc comparisons between time points within each condition were conducted using paired t-tests to assess directional trends over the course of the task. In addition, successive inter-beat intervals were converted into difference vectors, from which the proportion of successive differences exceeding 20 ms (pNN20) and 50 ms (pNN50) was calculated. The nonlinear Poincaré index (SD1/SD2) was derived from the standard deviations of instantaneous beat-to-beat changes (SD1) and longer-term variability (SD2). Mean values from low- and high-stress runs were compared using paired t-tests.

Functional MRI acquisition and preprocessing

MRI acquisition was conducted at the Melbourne Brain Centre Imaging Unit, The University of Melbourne, using a 7 T research scanner (Siemens Healthcare, Erlangen, DE) fitted with a 32Rx/1Tx-channel head-coil (Nova Medical Inc., Wilmington, MA, US). The functional sequence (beverage task) used a gradient-echo echo-planar imaging (GE-EPI) acquisition with multiband acceleration (factor 6) and GRAPPA (factor 2) in the steady state (repetition time 800 ms; echo time 22.2 ms; flip angle 45°; field of view 20.8 cm; slice thickness 1.6 mm, no gap; matrix 130 × 130; 84 interleaved axial slices aligned to the anterior–posterior commissure line)⁹⁶. The run duration was 7 min, yielding 525 whole-brain EPI volumes per session. A high-resolution T1-weighted, multi-echo MP2RAGE dataset was acquired at the start of the MRI protocol (repetition time 4500 ms; echo times 2.21/4.21/6.15/8.14 ms; inversion times 700/2700 ms; flip angles 6/7°; field of view 24 cm; slice thickness 0.75 mm; matrix 320 × 320; 224 axial interleaved slices)⁹⁷. The denoised, combined-echo 3D MP2RAGE image was used to facilitate functional image co-registration⁹⁸. Participants’ heads were stabilised using foam-padded inserts. Cardiac and respiratory signals were sampled at 50 Hz using a Siemens pulse oximeter and respiratory belt and were used for physiological noise correction.

Structural and functional data were pre-processed using the statistical parametric mapping toolbox (SPM12; v7771, Wellcome Trust Centre for Neuroimaging, London, UK) in a MATLAB 2023a environment (The MathWorks Inc., Natick, MA, US). To correct for motion artefacts, individual participants’ time-series were realigned to the mean image and individualised motion regressors created using the Motion Fingerprint Toolbox⁹⁹. Anatomical images were co-registered to their mean functional image, segmented and normalised according to the International Consortium of Brain Mapping template using the unified segmentation plus diffeomorphic anatomical registration through exponentiated lie algebra (DARTEL)¹⁰⁰. Images were then smoothed with a 3.2 mm full-width-at-half-maximum Gaussian kernel to preserve spatial specificity.

Physiological noise was included as a regressor by using the PhysIO toolbox at the first level¹⁰¹. This approach uses physiological recordings to apply noise correction to fMRI sequences and has previously been shown to benefit the temporal signal-to-noise-ratio of blood-oxygen-level-dependent (BOLD)^102,103. Specifically, we used retrospective image-based correction¹⁰⁴, to model periodic cardiac and respiratory contributions to the BOLD signal using canonical respiratory¹⁰⁵ and cardiac response functions¹⁰⁶. Individualised DARTEL-derived tissue masks were used for anatomical component correction (aCompCor¹⁰⁷) by extracting nuisance components from white matter and cerebrospinal fluid signals.

General linear models of beverage cue and taste processing

First-level (single-participant) GLMs were specified separately for each beverage-task run using the preprocessed fMRI time series and participant-specific nuisance regressors (head motion and physiological noise). Regressors of interest comprised the onsets of beverage cue presentation (‘CUE’) and taste receipt (‘TASTE’) for chocolate milk and water, each convolved with the canonical haemodynamic response function. All remaining task events (for example, pleasantness ratings and the rinse period) were modelled as regressors of no interest. A 128-s high-pass filter was applied to remove low-frequency drift. Serial autocorrelation was modelled using SPM’s FAST approach¹⁰⁸, and models were estimated in SPM using prewhitening and generalised least squares.

For group-level inference and to identify candidate regions for subsequent DCM analyses, subject-wise first-level contrast images were entered into a second-level GLM implemented as a voxelwise paired t-test, testing whether the mean within-participant contrast differed from zero across participants. Separate group-level analyses were performed for ‘CUE’ (all cue events collapsed) and ‘TASTE’ (all taste events collapsed). Whole-brain results were thresholded using voxel-wise FDR correction (pFDR < 0.05) with a minimum cluster extent of 10 voxels.

Dynamic causal modelling

DCM is used to infer directed influences between neuronal populations from observed BOLD time series data. A model is inverted (fitted using a variational Laplace, a Bayesian optimisation technique) to obtain the model structure and connectivity parameters that maximise parsimony: balancing accuracy and complexity⁴². This modelling framework can be utilised to infer the extent to which a manipulated input—an experimental task, for example—upregulates (positively modulates) or downregulates (negatively modulates) the intrinsic excitatory (positive) or inhibitory (negative) source-to-target communication between neuronal populations (nodes). Such continuous-time modulation can be framed in terms of rates (measured in Hz), whereby a task modulates the degree to which activity of an efferent node affects the continuous rate at which the state of afferent node changes. Previous evidence has demonstrated that 7 T fMRI furnishes more reliable effective connectivity inferences compared to lower MRI field strengths¹⁰⁹.

Region selection and time-series extraction

Four regions showing significant activation during ‘CUE’ and ‘TASTE’ were included in dynamic causal models: the BNST, the NAcc, the OFC and the dmINS. The VOIs for each region were extracted at the subject level as described in published guidelines⁴³. For the dmINS and OFC, the first eigenvariate across all activated voxels within a 4 mm radial sphere around the subject-specific maxima (p < 0.05, uncorrected), but no more than 8 mm from the group maxima, were extracted. Time-series extraction for the BNST was run using a consensus mask, created by overlaying the manually segmented BNST mask from Theiss and colleagues¹¹⁰, onto Pauli and colleagues’ extended amygdala mask¹¹¹. Both masks were binarised and combined to create a new mask containing only voxels common to the original masks. The NAcc mask was generated using the 1-mm³ resolution atlas from automated anatomical labelling 3 (AAL3¹¹²) toolbox. The OFC and dmINS were modelled unilaterally, using the strongest task-evoked peaks from the GLM, in line with DCM guidelines suggesting prioritisation of the dominant hemispheric activation (which in this sample, were located in the right hemisphere). The BNST and NAcc were modelled bilaterally given their small size, proximity to the midline, and known bilateral function. Four participants were excluded from the ‘CUE’ DCM analyses (n = 44) and one from the ‘TASTE’ DCM analyses (n = 47) during regional time-series extraction as no significant BNST voxels passed the significance threshold. Individual-level VOI peak locations are depicted in Fig. S5.

To model stress-related differences within a single dynamic causal model per participant, time series from the low- and high-stress runs were concatenated using spm_fmri_concatenate. Regional time series were extracted as GLM-adjusted eigenvariates: nuisance effects were regressed out during VOI extraction by adjusting the time series for regressors not included in the effects-of-interest F-contrast. The resulting residualised regional signals (and associated observation noise) were carried forward into DCM.

Model specification and inversion

Based on previous anatomical literature^16–18, DCM structures were specified such that they assumed bidirectional intrinsic connections between the BNST and the NAcc, OFC and dmINS (Fig. 3). At this first level, models were inverted using concatenated time series data from both low- and high-stress task runs (under standard, default priors). The cue models included all high-stress and low-stress ‘CUE’ blocks (n = 44) as the first input condition, high-stress ‘CUE’ blocks as the second input condition (high-stress cue) and high-stress rewarding ‘CUE’ blocks as the third input condition (high-stress chocolate milk cues). This structure was repeated for the taste models and included all high-stress and low-stress ‘TASTE’ blocks (n = 47). The modulatory effect of food cues, taste, rewarding cues (chocolate milk cues) and rewarding taste (chocolate milk taste) under high stress were tested on all pathways to and from the BNST.

Each model assessed a distinct modulatory effect on BNST effective connectivity: (1) high-stress cues: the modulatory effect of all beverage cue presentations under high stress, (2) high-stress chocolate milk cues: the modulatory effect of rewarding beverage cue presentation under high stress, (3) high-stress taste: the modulatory effect of all beverage taste receipt under high stress, (4) high-stress chocolate milk taste: the modulatory effect of rewarding beverage taste receipt under high stress. For the cue models (Models 1 and 2), the driving input was specified as all beverage cue onsets across both runs. For the taste models (Models 3 and 4), the driving input was specified as all beverage taste receipt onsets across both runs.

Using four separate dynamic causal models instead of a single model with multiple modulatory inputs preserves the theoretical and experimental distinctions between stress conditions and cue- and taste-related processes, preventing interactions that could obscure meaningful differences⁴². It also reduces model complexity (overparameterisation), and helps to ensure that connectivity changes are clearly attributable to specific conditions⁴².

Parametric empirical Bayes

Inversion of dynamic causal models yielded posterior distributions over effective connectivity parameters. These posterior distributions for the modulatory effects served as data in PEB models wherein subject-level effective connectivity represents random deviations from an average (group-level) effective connectivity network. One PEB model (Fig. 4d) included the effect of change in subjective stress, quantified as the change (delta) from low- to high-stress induction⁴⁴ (specifically, the self-reported stress score after high-stress induction minus the self-reported stress score after low-stress induction). After inverting each PEB model, Bayesian model reduction and a greedy search were utilised to evaluate reduced models by omitting connections. Plausible reduced models were averaged using a Bayesian model averaging scheme (per the spm_dcm_peb_bmc routine). Evidence of a nonzero group effect was inferred when the two-sided posterior probability exceeded 0.75 (positive evidence¹¹³).

Leave-one-out cross-validation

Leave-one-out cross-validation was used to test the predictive accuracy of group effect sizes on participants’ individual changes in subjective stress, using the spm_dcm_loo routine. Focusing on the high-stress cue-related modulation of the BNST-to-OFC connection, a PEB model was iteratively inverted for all but one participant to test whether the modulation of effective connectivity parameters was robust enough to predict the change in subjective stress of the left-out participant. The reliability of this prediction was indicated by the correlation between predicted and observed values. An additional subject-level analysis (Fig. S6) showed that individual differences in stress-evoked BNST-to-OFC modulation (encoded in dynamic causal models) was positively associated with changes in subjective stress, providing a complementary, non-PEB perspective on the same effect.

In silico validation of dynamic causal models

The accuracy of DCM can be affected by factors such as signal quality. To address this issue, we conducted an in silico analysis of the main cue- and taste-related dynamic causal models to assess robustness of parameter recovery to estimated noise. These analyses involved three steps: (1) estimation of subject- and region-specific signal-to-noise ratios (SNRs) from empirical BOLD time-series data; (2) simulation of synthetic BOLD time-series using inferred model parameters—including hemodynamic delay parameters—that was combined with region-specific noise reflecting the estimated SNRs; and (3) model inversion using both DCM and PEB procedures on the simulated data. We then evaluated parameter recovery and the correspondence between second-level estimates derived from empirical versus simulated data (SI Results; Table S8; Figs. S7 and S8).

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

Author contributions

Conceptualisation: E.G.-H., T.S., R.M.B., P.S. Methodology: E.G.-H., M.D.G., T.S., R.M.B., P.S., R.K.G., B.A.M., B.J.H. Software: E.G.-H. Formal analysis: E.G.-H., M.D.G., T.S. Investigation: E.G.-H., M.D.G., P.-H.K., T.S. Visualisation: E.G.-H., M.D.G., T.S. Supervision: R.M.B., P.S., S.B.M., T.S. Writing—original draft: E.G.-H., M.D.G., T.S. Writing—review & editing: E.G.-H., M.D.G., T.S., P.-H.K., P.S., R.K.G., B.A.M., B.J.H., S.B.M., R.M.B. Funding acquisition: T.S., R.M.B.

Peer review

Peer review information

Nature Communications thanks Nils Kroemer and the other, anonymous, reviewer(s) for their contribution to the peer review of this work. A peer review file is available.

Data availability

Deidentified effective connectivity, behavioural data and self-reported demographic data are publicly available at https://github.com/evaghreins/Gustometer_Stress_BNST_DCM and deposited in Zenodo under accession code 10.5281/zenodo.18738387.

Code availability

Scripts used to generate the main results of this study are available at https://github.com/evaghreins/Gustometer_Stress_BNST_DCM and deposited in Zenodo under accession code 10.5281/zenodo.18738387. Custom code was written in MATLAB 2023a.

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at 10.1038/s41467-026-71414-y.