Learning about location-dependent threat: neural abnormalities in clinical anxiety

Anxiety disorders are characterized by maladaptive defensive responses to distal or uncertain threats. Elucidating neural mechanisms of anxiety is essential to formulate new treatment strategies tartgeting these circuits. In an fMRI scanner, patients with pathological anxiety (ANX, n=23) and healthy comparisons (HC, n=28) completed a contextual threat learning paradigm, in which they picked owers in a virtual environment comprising a danger zone in which owers were paired with shock and a safe zone (no shock). ANX compared with HC showed 1) global decreased ventromedial prefrontal cortex and anterior hippocampus activation during the task, 2) increased insula and dorsomedial prefrontal cortex activation in the danger zone during the task; and 3) increased amygdala and midbrain/periaqueductal gray activation in the danger zone prior to potential shock delivery. Our ndings suggest that ANX exhibit a heightened reactivity to threat and show decreased activation in modulatory areas responsible for regulating context-appropriate emotional responses. behavioral/physiological learning: task ratings skin conductance for owers safe The neural ower either key the learning circuit, anterior amygdala, ventromedial prefrontal (vmPFC), and vmPFC-hippocampal functional connectivity. (2) Threat appraisal: approach of the ower, the danger the safe zone, recruited sensory and control-related regions, the insula, dorsal anterior cingulate cortex (dACC), extending to the dorsomedial prefrontal (dmPFC), and insula-hippocampal functional connectivity. (3) Threat anticipation: During imminent threat of potential shock, upon picking a ower, a progressively increasing response of the periaqueductal gray matter (PAG) and posterior hippocampus, and insula-dACC coupling observed. previous ndings show that the hippocampus is involved in learning and discrimination of safe and zones, with increased activation with task experience. 8 Therefore, decreased engagement of the anterior hippocampus might reect an inability of patients with pathological anxiety to appropriately use context-specic cues for emotional regulation. These results are in line with research showing hippocampus and medial prefrontal cortex abnormalities in patients suffering from pathological anxiety 23,24 who report an exaggerated response to threats in contexts predicting safety. Reduced activation of the anterior hippocampus, particularly during threat, could impair emotional regulation abilities when shifting contexts. ventromedial


Introduction
When exploring our environment, we might encounter items that require us to learn about their threat value. Learning about potential threatening environments may induce anxiety, an anticipatory response to potential threats, and a fearful response evoked by an imminent acute threat. [1][2][3][4][5] Although anxious and fearful states are normal responses to threats chronic manifestations of these states can be highly debilitating. 6,7 Research shows that patients with chronic anxiety lack the ability to integrate contextual cues to guide learning of threat and safety. [9][10][11] A previous investigation has delineated the neural mechanisms underlying learning and discriminating threats within speci c spatial locations in healthy adults. 8 However, very little is known about how patients with pathological anxiety learn about threat within an environment. Understanding how patients with pathological anxiety learn about threat within speci c spatial location is essential to better understand the disorder and create novel treatments.
Traditional context conditioning paradigms have shown that a defensive response can be triggered not only by the presentation of an aversive stimulus but also by the presentation of the context where an aversive stimulus was previously encountered. Typically, healthy individuals can learn to distinguish between a safe and dangerous context. That is, when healthy individuals associate a cue with an aversive stimulus (conditioned stimulus; CS) in context A (CS+A), they display a threat response that is dependent on the context in which the association was made. However, the same cue (CS) in a different context B (CS-B) elicits a weaker defensive response. Context conditioning requires the use of spatial processing strategies [9][10][11] (i.e., attend to the environment and surrounding landmarks to create a spatial representation of where a threat was encountered). These strategies have been mapped to neural systems that regulate emotion and memory, such as the hippocampus, amygdala, and prefrontal cortex (PFC). [11][12][13][14][15][16][17][18][19][20][21] We developed a virtual-environment paradigm to probe how brain regions interact to shape behavior (i.e. threat learning and discrimination) over time 8 to quantify threat learning relevant to context-speci c threat. The virtual environment depicts a grass eld surrounded by mountains, divided equally into two zones, a safe and a danger zone. In both zones, owers appear and need to be "picked" up. Picking owers in the danger zone potentially causes an electric shock to the wrist (or "bee-sting"), while owers in the safe zone are never associated with shock. To learn threat contingencies, participants must rely on distal environmental cues (e.g., shape of the mountains and clouds, which differ in both zones, and beehives) to locate themselves and learn 'where' they are in the environment and not on the physical properties of the stimuli (i.e., the owers), which are all identical in both zones. 8 Previous ndings with this paradigm in healthy volunteers informed three processes: threat learning, threat appraisal (anxiety-state), and threat anticipation (fear-state). (1) Threat learning: Healthy adults demonstrated behavioral/physiological learning: as the task progressed, shock expectancy ratings and skin conductance for owers increased in the danger but not the safe zone. The neural substrates associated with learning about environmental threats during ower approach, in either zone, engaged key nodes of the learning circuit, including the anterior hippocampus, amygdala, ventromedial prefrontal cortex (vmPFC), and vmPFC-hippocampal functional connectivity. (2) Threat appraisal: approach of the ower, in the danger zone compared to the safe zone, recruited sensory and control-related regions, i.e., the insula, dorsal anterior cingulate cortex (dACC), extending to the dorsomedial prefrontal cortex (dmPFC), and insula-hippocampal functional connectivity. (3) Threat anticipation: During imminent threat of potential shock, upon picking a ower, a progressively increasing response of the periaqueductal gray matter (PAG) and posterior hippocampus, and insula-dACC coupling were observed. 8 Research has shown that individuals with anxiety disorders display a higher defensive response to a CS+, compared to healthy individuals, and they often generalize the threat response to safe cues (CS-), 6,7,12,14 but this not clear if this extends to context. Still, hippocampal dysfunction and decreased hippocampal volume have been associated with anxiety disorders. 12,13,15 For example, studies in both humans and rats suggest that impairment in hippocampal function leads to compensatory learning strategies that do not involve the hippocampus. These abnormal modulation of attention, linked to attention shifts to the cue and not the context, use compensatory neural mechanisms that leads to generalization of threat. 6,7,[10][11][12]14,[16][17][18][19][20]20,21 Using the virtual environment paradigm described above, three main hypotheses are tested. Individuals with an anxiety disorder (ANX; generalized anxiety disorder, social anxiety disorder) compared with a sample of healthy comparisons (HC) will show: (1) Threat learning: poor learning and discrimination associated with compensatory learning strategies that do not involve the hippocampus leading to generalization of threat, (2) Threat appraisal: stronger engagement of the dACC, dmPFC, amygdala, and insula in danger zones when approaching the owers (higher anxiety-state), and (3) Threat anticipation: stronger engagement of the periaqueductal gray matter (PAG) activation in danger zones in anticipation of potential shock delivery (higher fear-state).

Physiological and Behavioral Measures of Threat Learning
Participants explored a virtual environment ( Fig. 1A and B; see Methods for details) that consisted of a mountain landscape and de ned two half-zones recognizable by the unique shape of the mountains in the horizon. For each trial, participants freely explored the arena and were instructed to pick up owers that appeared one at a time in random locations across the environment (approach period). When a participant picked a ower, their position was held stationary for a variable duration (2-8 seconds; stationary period), during which the participant rated the expectancy of receiving a shock (rating of 0-9). Flowers located in one-half of the environment were paired with a shock delivered at the end of the stationary period on 50% of the trials within the danger zone. Flowers in the other half of the environment were never paired with a shock (safe zone). Since all owers were identical, predictive value (danger or safety) could not be attributed to their physical characteristics.
For analysis, the data were divided by zones (safe, danger) and segregated into 4 learning blocks (10 trials in each). All these physiological and behavioral analyses can be found in the supplementary material.
Skin conductance. We measured participants' skin conductance level (SCL), as they navigated towards the ower (approach period), and skin conductance responses (SCR), during the stationary period. A 2x4x2 ANOVA (zone x block x group) of both SCL and SCR revealed a signi cant main effect of zone, due to higher responses in the danger zone as compared to the safety zone (SCL, F(1,49)=16.24, p<0.01; SCR, F(1,49)=4.13, p<0.05; For a full description of skin conductance results, see supplementary material and Figure S1). Shock expectancy. A 2x4x2 ANOVA (zone x block x group) assessing shock expectancy ratings after picking the ower revealed a signi cant zone x block interaction (F(3,147)=31.07, p<0.001). This interaction stemmed from consistently (across the 4 blocks) high expectancy ratings of the shock in the danger zone (danger vs safe: t(50)=0.06, p>0.05), while ratings signi cantly decreased between block 1 and 4 in the safe zone suggesting that there was discrimination learning over blocks for all subjects (block 1 vs 4: t(50)=8.83, p<0.001; For a full description of the results see supplementary material and Figure S1).

fMRI Measures of Threat Learning
Threat appraisal (Anxiety-state): Approaching owers (approach period) in the danger vs. the safety zone Each approach period (approaching a ower) began at trial onset, when the ower appeared in the environment and ended when the ower was "collected." For analyses we excluded the initial orienting period (looking for the ower) of the approach period, only including the last 75% of the approach (active navigation towards the ower).
We compared brain activation (i.e., presumed metabolic activity) between diagnostic groups (ANX vs. HC) as individuals approached owers located in the danger/safe zones of the environment (see Table S1 for full results from this analysis). To assess discrimination learning, and increase signal-to-noise ratio, trials were divided into two blocks comprising the rst-(early) and second-half (late) of the experiment. To directly examine group differences, the rst-level analysis contrasted factors of zone (safe vs. danger) and block (early vs. late), whereas the second-level analysis directly compared groups (group: ANX vs HC). Signi cant peak activation was extracted and analyzed to disentangle the directionality of the results.
An overall group contrast between ANX and HC of approach periods (assessing changes across both zone and block) identi ed two opposing patterns of activity changes in a range of areas comprising posterior cingulate cortex (PCC; p<0.05 FWE), vmPFC, orbitofrontal cortex (OFC), and bilateral anterior hippocampus (p<0.05 FWE small volume corrected; SVC, Fig. 2A). Compared to the HC group, the ANX group demonstrated a greater increase in activity in these areas from early to late blocks (late > early) of the safe zone, and a decrease in activity from early to late blacks of the danger zone. To further understand these distinct patterns of activity, we next performed direct group comparisons on separate components of the task.
When approaching owers in the danger zone (danger > safe), the ANX group (ANX > HC) showed greater activation of bilateral insula (p<0.05 FWE) and dmPFC (p<0.05 FWE SVC; Fig. 2B) compared to the HC group. That is, these areas were more responsive in the ANX group when approaching owers in the danger compared to safe zone. No group differences were observed when looking for areas that showed greater activity when approaching the safe compared to danger zone.
For the HC group, compared to ANX (HC > ANX), approaching owers in the second half of the experiment, compared with the rst half (late > early) showed that, regardless of zone, there was increased activation from early to late blocks in PCC, vmPFC, OFC, and anterior hippocampus (p<0.05 FWE SVC). No other signi cant results were found (p>0.001).
Next, psychophysiological interactions (PPI) analyses were performed for each participant group separately to identify brain regions in which connectivity changed during the danger vs. safe contrast. PPI examined brain connectivity of each signi cant cluster (i.e., seed ROI) from the approaching owers period. In the HC group, a positive association between the dmPFC seed (ROI, MNI coordinated: 9, 26, 45) and bilateral insula (p<0.001 Bonferroni corrected) was found in the contrast (dange > safe On the other hand, in ANX, a negative correlation between the dmPFC-OFC and dmPFC-vmPFC (p<0.001 Bonferroni corrected) was found in the same contrast (danger > safe). No other PPI analysis revealed signi cant results.
Overall, these results suggest that when approaching owers in the dangerous zone, ANX showed reduced activation in vmPFC, PCC, and anterior hippocampus while showing greater activation in the insula and dmPFC. These ndings are further highlighted by negative connectivity between the dmPFC and vmPFC areas. On the other hand, HC displayed greater activation as a function of time in the vmPFC, OFC, PCC, and anterior hippocampus, regardless of the zone suggesting appropriate contextual learning.
Threat anticipation (Fear-state): Held stationary in the danger vs. the safety zone We next examined changes in brain activation when participants' positions were held stationary after picking owers and anticipating a potential shock. We again examined the effects of zone (danger vs. safe zone) and block (early vs. late blocks) in the rst level analysis. The second level analysis consisted of the group comparison (ANX vs. HC; see Table S2 for full results). Analyses followed the same model as for the approach period.
An overall group contrast between ANX and HC during stationary periods (assessing changes across both zone and block) identi ed two opposing patterns of activity changes in a range of areas comprising PCC (p<0.05 FWE), vmPFC, and OFC (p<0.05 FWE SVC). Compared to the HC group, the ANX group demonstrated a greater increase in activity in these areas from early to late blocks (late > early) of the safe zone, and a decrease in activity from early to late blacks of the danger zone. To further understand these distinct patterns of activity, we next performed direct group comparisons on separate components of the task.
When held stationary after picking a ower located in a zone of the environment associated with danger (danger > safe), the ANX group compared to HC (ANX > HC) showed greater activation in dmPFC, dACC, bilateral insula, caudate, thalamus, amygdala, and midbrain areas, including the periaqueductal gray (PAG; p<0.05 FWE SVC; Fig. 3A). A group contrast (ANX > HC) of stationary periods (irrespective of danger or safety) showed increased dmPFC activation from early to late block (late > early) in ANX compared with HC (p<0.05 FWE SVC).
For the HC group, compared to ANX (HC > ANX), owers located in a zone of the environment associated with safety (safe > danger) generated greater activation in the PCC (p<0.05 FWE), vmPFC, OFC, and anterior hippocampus (p<0.05 FWE SVC; Fig. 3B). For HC, compared to ANX (HC > ANX), we also found increased activation during the last half of learning (late > early) regardless of zone in vmPFC and OFC (p<0.05 FWE SVC). No other signi cant results were found (p>0.001).
Given the reported group differences during stationary periods, PPI analyses were next performed for each group separately to identify brain regions in which connectivity changed during danger vs. safe. PPI examined brain connectivity of each signi cant cluster (i.e., seed ROI) from the stationary period. PPI analyses used dmPFC (MNI coordinates: 0, -8, 71) and each amygdala side as seed regions (MNI coordinates: (R) 26, -2, -15, (L) -20, 2, -15). ANX showed increased functional connectivity between the dmPFC-bilateral insula, left amygdala-bilateral insula, and right amygdala-bilateral insula in danger compared to safe zones (p<0.001 Bonferroni corrected). HC showed increased functional connectivity between the dmPFC-OFC, left amygdala-bilateral insula, right amygdala-bilateral insula, left amygdala-vmPFC, right amygdala-vmPFC, left amygdala-OFC, and right amygdala-OFC in danger compared to safe zone during the stationary period (p<0.001 Bonferroni corrected). No other PPI analyses revealed signi cant results.
In summary, during the stationary period in the danger zone (after collecting a ower) ANX, compared to HC, demonstrated reduced activation in vmPFC and PCC over time, with greater activation in the insula, amygdala, and PAG. This was further highlighted by increased connectivity among the dmPFC, amygdala, and insula while lacking any signi cant connectivity from vmPFC areas. On the other hand, HC's recruitment of the vmPFC, OFC, and PCC was seen as a function of time, where those areas were recruited during the stationary periods in both the safe and dangerous zones. Furthermore, HC had signi cant connectivity of the vmPFC and OFC to areas such as the dmPFC and amygdala.

Discussion
The present study examined how patients with pathological anxiety learn to distinguish contextual features that inform the dangerous vs. safe status of stimuli within a virtual environment. This virtual environment task shows how brain regions interact to shape behavior over time. We hypothesized that patients with pathological anxiety would show (1) poor threat learning associated with lower recruitment of the anterior hippocampus; (2) biased threat appraisal with stronger engagement of the dACC, dmPFC, amygdala, and insula (higher anxiety-state); and (3) higher threat anticipation with stronger engagement of the PAG (higher fear-state).
Surprisingly, inconsistent with our hypotheses, patients with pathological anxiety showed no impairment in discriminating between threat and safety within the environment based on their skin conductance and subjective reports. However, consistent with our hypothesis, patients with pathological anxiety, compared to HC, displayed (1) weaker engagement of the anterior hippocampus and vmPFC implicated in memory processing and emotional regulation during both approach and stationary periods (Threat learning and discrimination); (2) higher engagement of the dACC, dmPFC, amygdala, and insula involved in negative valence, emotional expression, and con ict processing during both approach and stationary periods (Threat appraisal); and (3) higher engagement of midbrain areas, including the PAG, during anticipation of potential shock delivery (Threat anticipation; see Fig. 4). These results suggest that patients with pathological anxiety recruit compensatory learning strategies that do not involve the anterior hippocampus or vmPFC, which might explain the increased activation in the salience network (e.g., dmPFC, insula, PAG). In patients with pathological anxiety, reintegrating the anterior hippocampus and vmPFC into the circuitry could rescue appropriate learning strategies to discriminate between safety and danger.

Threat Learning and Discrimination
We observed changes in activation in several brain regions as the experiment progressed ( rst vs. second block) that differed between patients with pathological anxiety and healthy comparisons. Healthy comparisons showed increased activation over time, for regions involved in emotion regulation (ventromedial prefrontal cortex; vmPFC), valuation (orbitofrontal cortex; OFC), spatial memory (hippocampus) and self-referential processing (posterior cingulate cortex; PCC) regardless of safe/danger zone. In patients with pathological anxiety, these brain areas (vmPFC, OFC, hippocampus, and PCC) differentiated between zones, with increased activation in the safe zone and decreased activation in the danger zone over time. The anterior hippocampus is essential for the integration of spatial information in mediating anxiety-like behaviors. 22 As seen in the current study with healthy comparisons, previous ndings show that the hippocampus is involved in learning and discrimination of safe and danger zones, with increased activation with task experience. 8 Therefore, decreased engagement of the anterior hippocampus might re ect an inability of patients with pathological anxiety to appropriately use contextspeci c cues for emotional regulation. These results are in line with research showing hippocampus and medial prefrontal cortex abnormalities in patients suffering from pathological anxiety 23,24 who report an exaggerated response to threats in contexts predicting safety. Reduced activation of the anterior hippocampus, particularly during threat, could impair emotional regulation abilities when shifting contexts.
Patients with pathological anxiety displayed reduced activation in the vmPFC, OFC, and PCC throughout the ower task; particularly, when interacting with owers in the danger zone during the late blocks. These vmPFC/OFC areas are considered to have key roles in regulating contextually appropriate emotions. [25][26][27][28] Previous studies reported that, in healthy individuals, vmPFC, PCC, and anterior hippocampus activation did not differentiate danger and safety during this ower task, but rather increased with experience (from early to late block of the experiment), suggesting a role in learning and discrimination. 8 In the present study, patients with pathological anxiety, unlike healthy comparisons, demonstrated increased engagement for safe owers and decreased engagement for dangerous owers. Given the role of the anterior hippocampus and vmPFC/OFC, the hypoactivation observed in patients with pathological anxiety may suggest that they are unable to use contextual cues to regulate their emotional output, particularly in a potentially dangerous location.
Threat Appraisal (Anxiety-State) Compared to healthy comparisons, patients with pathological anxiety demonstrated greater activation in brain areas involved in threat appraisal (dmPFC, dACC, amygdala, insula) throughout the whole experiment, particularly in the danger zone. Greater engagement of these brain areas likely re ects higher arousal when interacting with the owers, particularly in the danger zone. Activation in the dmPFC, along with amygdala and its connectivity, has been associated with threat arousal 29-32 and may play a key role in integrating threat information to coordinate emotional response with other brain regions. Increased dmPFC activation to threat stimuli has been documented in patients with pathological anxiety, as compared to healthy comparisons, 25,26,30,31 suggesting an overactive appraisal of surroundings and cues. Accordingly, dACC and insula activations have been reported during approach of threat in healthy adults using the ower task, 8 likely re ecting the integration of visceral feelings and cognitive appraisals of threat to trigger fear expression. 33 However, in this study, patients with pathological anxiety showed greater dmPFC, dACC, and insula activation throughout the task, possibly re ecting an overactive threat appraisal and detection network.
Threat Anticipation (Fear-state) During imminent threat, patients with pathological anxiety exhibited greater activation in brain areas related to error detection, con ict resolution, and emotional expression (dACC and amygdala), and regions related to salience and pain judgment (bilateral insula and PAG). Midbrain areas, including the PAG, and amygdala have been reported during imminent threat in healthy adults using the same task, 8 likely re ecting the integration of cognitive appraisals to the emotional anticipation of threat to trigger appropriate fear expression. 33 However, in this study, patients with pathological anxiety, as compared to healthy comparisons, showed greater PAG and amygdala activation during imminent threat, possibly re ecting an overactive threat anticipation. Particularly, increased PAG activation during stationary periods in the danger zone might suggest greater anticipation of the potential shock just prior to its likely delivery.

Brain circuitry connectivity
Unlike healthy comparisons, patients with pathological anxiety failed to exhibit task-related functional connectivity with higher-level cortical areas related to emotional regulation (vmPFC and OFC) throughout the task. During ower-approach, patients with pathological anxiety displayed negative functional connectivity between the dmPFC and both the vmPFC and OFC. During stationary periods in the danger zone, patients with pathological anxiety displayed positive dmPFC-insula and amygdala-insula functional connectivity, while in healthy comparisons these brain areas were functionally connected to the OFC and vmPFC. Coupling between the dmPFC and amygdala has been previously observed with induced threat of shock 25,34,35 and in patients with pathological anxiety 36 during an emotional identi cation task. In pathological anxiety, this circuit may become hyperactive and unable to "turn off," contributing to an attentional bias towards threat. In this study, functional connectivity de cits with vmPFC/OFC areas, in patients with pathological anxiety, could re ect a faulty exible learning spatial strategy. For example, patients with pathological anxiety may fail to integrate the learned contingencies within the environment and surrounding landmarks to create a spatial representation that can modulate emotional output to stimuli while navigating through the safe and danger zones. Therefore, an in exible or exaggerated evaluation of the cue may hinder the use of spatial strategies to regulate emotional output appropriately. In other words, patients with pathological anxiety might have shock-related worry that gets progressively worse as they approach and pick up the ower, and as the experiment progresses.
In conclusion, patients with pathological anxiety show lower activation in the anterior hippocampus and vmPFC/OFC and lower connectivity to emotion expression brain areas (dmPFC, amygdala, and insula), which could explain poor emotional regulation. This possibility is further supported by higher dmPFC activation and other emotion valuation and expression brain areas. Our ndings suggests that patients' threat-related attention may cause de cits in emotion regulation. The current results suggest that disconnected circuitry in brain areas essential for emotional regulation might lead to disrupted emotional output while exploring environments and learning to discriminate cues within these environments. Finding novel psychological intervention or trainings to reintegrate the vmPFC and hippocampus into the learning and discrimination circuitry could help patients with pathological anxiety display appropriate emotional responses while navigating novel environments.

Methods
Participants. Seventy participants, aged 18-50 years, were recruited from the Washington D.C. and Maryland areas. Thirty participants were diagnosed with an anxiety disorder (ANX; generalized anxiety disorder and/or social anxiety disorder) using the Structured Clinical Interview for Diagnostic and Statistical Manual of Mental Disorders 4 th edition, while the other 40 volunteers were healthy comparisons (HC). Before taking part, all participants provided written informed consent and, after completion, were debriefed and reimbursed for their time. The study was approved by the NIH Institutional Review Board. All participants were right-handed, free from neurological impairment, or any psychological disorders (except for an anxiety disorder in the patient group). Four ANX and 10 HC participants were excluded from analyses because they were unable to explain the shock contingencies between the locations at the end of the task (see procedure below). Three ANX and two HC were omitted due to technical issues during scanning or excessive head motion (>20%; see below). Therefore, the nal sample included 23 ANX (mean age=29.6; SD=8.22) and 28 HC (mean age=27.25; SD=8.21). Participants were free of medications and no signi cant differences in any demographic information between groups was found.
Skin conductance. Skin conductance was measured as an index of anxiety via 8mm Ag/AgCl electrodes attached to the medial phalanges of the index and middle ngers of the participant's left hand. Data were acquired using a Biopac EDA100C MRI system (Biopac Systems, Inc., Goleta, CA, USA) at a sampling rate of 1000Hz.
Shocks. Shocks (20 samples) were applied using a Digitimer DS7A electrical stimulator (Digitimer, Welwyn Garden City, UK) to the left hand with intensity up to 50mA for 2ms duration through a silver chloride electrode. Shock intensity was adjusted individually for each participant before starting the experiment. Individual adjustment procedures delivered a series of shocks to each participant, starting at 12mA. Participants were asked to rate the level of pain with each shock on a 1-10 scale. Shock intensity was increased until the level was irritating, but not painful.
Task. A description of the virtual environment task, developed with Unity Software (Unity Technologies, USA) is available in a prior publication. 8 The virtual environment consisted of a circular grassland with a perimeter boundary wall surrounded by distal cues (mountains, sun, and clouds) for orienting, two landmarks (beehives) placed in the grassland, and owers that appeared one at a time in random locations. When participants picked a ower, they rated on a scale from 0-to-9 the likelihood of receiving a shock from the ower picked. After, the virtual character would enter a stationary period for 2000ms-8000ms. After the stationary, participants could move once again, and another ower would appear in the environment. The task included a total of 80 owers-40 owers were paired with a shock on 50% of the trials (danger) while encountering no shock with the remaining 40 (safe).
After every four ower trials, we included a spatial memory trial in which participants learned the location of one of four objects (wooden box, gas can, book, and clock); however, two objects appeared in each half of the environment. No shocks accompanied spatial memory stimuli. After the rst four spatial memory trials, participants' memory for object locations was tested by asking participants to replace the objects where they originally found them (For Object task results see supplementary material).
At the end of the experiment, participants were asked to name the four objects and their locations, as well as explain the contingencies of danger and safety during threat learning, by answering if there was a pattern in the shocks. Participants who were unable to provide the objects' name and position, or explain the contingencies were excluded from the nal analysis to ensure that participants were paying attention to the task.
Behavioral analysis. Data processing and analysis of electrodermal activity (EDA) were performed using MATLAB. EDA data were down-sampled to 200 Hz and then synchronized to the task. EDA was assessed during two periods of the threat learning task. First, mean skin conductance level during each approach quanti ed tonic skin conductance levels (SCL) as participants navigated towards the ower. SCL was quanti ed from the last three-quarters of the approach period from ower appearance until trial completion. Skin conductance level was calculated by measuring the mean skin conductance from the beginning of active approach until right before the ower was picked for each trial. Second, skin conductance responses (SCR) were analyzed during the stationary period to examine phasic changes in anticipation of the shock outcome. SCRs were calculated for every trial by subtracting the minimum skin conductance during the stationary period (baseline) from the maximum response (peak) before the stimulus onset. Any response difference under 0.03 micro-Siemens was scored zero. SCRs were log transformed (log [1+SCR]) to normalize the distribution, and then range correction ([SCR-SCRmin]/[SCRmax-SCRmin]) was applied to control for individual variation in responding. 37 The same correction was applied to the SCLs. For analyses, SCRs and SCLs were averaged into four equal blocks across the duration of the experiment, with each block including ten trials per condition (safe and danger).
Expectancy ratings taken at the beginning of each stationary period were analyzed similarly to skin conductance. Each rating (0-9) was averaged across trials to create four equal blocks separated by safe and danger conditions (10 trials in each block).
Finally, performance on the spatial memory task was analyzed by assessing distance error on each test trial. This distance error was calculated by taking the distance in virtual meters between the participant's response location when replacing the object and its correct location within the environment. Distance error was taken from each trial and averaged into four blocks (1 trial from each object in each block).
All results were analyzed using a General Linear Model (GLM) for repeated measures using 2x2x4 ANOVAs to test differences between zone (safe, danger), group (HC, ANX), and block (1 to 4). Based in a priori hypothesis of ANX overgeneralization 12 , the simple effects within each group were examined using 2x4 ANOVAs to characterize differences between conditions (safe, danger) and block (1 to 4). fMRI acquisition. Blood oxygen level-dependent T2*-weighted functional images were acquired on a 3T Skyra system (Siemens, Germany) using echo-planar imaging (EPI) with a 32-channel head coil. Images were acquired with a 45° oblique angle with the following parameters: 3300 ms TR; 30 ms TE; 1 mm interslice gap, 192 mm eld of view, and 48 axial slices with 2 mm slice thickness resulting in 3mm isotropic voxels. A single echo eld map was recorded for distortion correction of the acquired EPI. After the functional scans, a T1-weighted 3-D structural image (1mm 3 ) was acquired to co-register and display the functional data.
fMRI analysis. Data processing and analysis were performed using SPM12 (http://www. l.ion.ucl.ac.uk/spm). EPI images were rst preprocessed using a bias correction to control for within volume signal intensity difference, unwarping and realignment to correct for movement and slice-time correction. Images were then spatially normalized to the MNI template using parameter estimates from warping each participant's structural image to a T1-weighted average template image. All images were nally smoothed using an 8mm FWHM Gaussian kernel.
The analysis model included 16 regressors of interest. Four separate regressors of interest were created for approach periods by zone (safe or danger) and block ( rst or second half of experiment); these consisted of boxcar functions from the end of the rst quarter of each approach period to the point in which the ower was reached. Four regressors of interest (also by zone and block) were created for the stationary period of each trial, consisting of a boxcar function starting after the participant rated their shock expectancy and lasting for the duration of the stationary period. Three regressors using a stick/impulse function modeled the end of each trial in safe conditions and in danger conditions split by whether participants received a shock or not. Finally, four regressors modeled when participants were replacing objects during the spatial memory task, using a boxcar function covering when a response was made and the approach period to the location where the object was picked ( rst and second half of the experiment). Six regressors of no interest were also added to the model representing movement parameters estimated during realignment. Frames with more than 0.5 mm frame-wise head motion were detected as outliers and modeled using Artifact detection tool (ART). Participants with outliers totaling 20% or more of their total scan were removed from analysis.
Statistical analyses occurred in two stages, individual (General Linear Model analysis) and group level (two sample t-test). At the individual level, an initial control analysis was conducted to examine emotional ower vs. unemotional object approaches to a cue. The neural responses to approach periods were compared between the threat learning task (approaching owers) and the spatial memory task (approaching location to replace the object; See supplementary material). We created a model contrasting approach periods for threat learning (collapsing across safety and danger) with approach during spatial memory across the rst and second half of the experiment using a 2x2 ANOVA (task, block). Approach periods during threat learning were then examined contrasting approach to owers associated with safety or danger and collected during the rst or second half of the experiment using a 2x2 ANOVA (zone, block). Finally, stationary periods during threat learning, periods when the ower was picked, and participants were held stationary, were analyzed in a similar rst-level analysis using a 2x2 ANOVA (zone, block).
At the group level, the contrasts of interest, described above, were entered into two-sample t-tests to compare between the ANX and HC. Within this model, t-contrasts were examined to investigate ANX>HC and HC>ANX differences. To aid and con rm interpretation of signi cant group effects, parameter estimates were extracted from signi cant peak coordinates for the contrasts ANX>HC and HC>ANX using the MarsBaR region of interest toolbox. Post-hoc analysis of these contrasts was conducted using repeated measures 2x2x2 ANOVA (zone or task, block, group) in SPSS.
Family-wise error (FWE; p<0.05) corrected effects across the whole brain are reported for all analyses.
Given the a priori hypotheses and previous ndings, 8 an ROI approach was employed, based on the Automated Anatomical Labeling atlas (AAL) 38 and the WFU Pickatlas toolbox in SPM12. 39 A bilateral mask was de ned for the hippocampus, amygdala, and mPFC (orbitofrontal gyrus, ventromedial prefrontal cortext, dorsomedial prefrontal cortex, and anterior cingulate and medial cingulate cortex).
Statistical threshold was de ned by small volume correction (SVC; p<0.05 FWE).
For any signi cant interaction, the representative time-course was extracted through SPM12 MarsBaR (http://marsbar.sourceforge.net) toolbox, using a 6mm sphere centered on the peak of the activation in the regions of interest, using the rst eigenvariate calculated from singular value decomposition. The extracted values were analyzed in SPSS 24 with a 2x2 ANOVA (task x block) and further analyzed using a one-sample t-test, which was Bonferroni corrected.
'Activation' was used throughout the manuscript to indicate an increase in presumed metabolic activity.
Functional connectivity analyses. Functional connectivity was assessed at the individual level using psychophysiological interactions (PPI) analysis with SPM12. The PPI compares functional connectivity from a single seed region across multiple task conditions. For exploratory analyses, seed regions were selected based on group differences identi ed in the main analysis, such as the vmPFC, OFC, dmPFC, amygdala, and hippocampus. Peak activation from these brain areas in the group level analysis, for approach and stationary periods, were used to create regions of interest (6 mm sphere centered on group level peak activation) for each participant. The seed time series activation for each participant was extracted at the center of the activation peak. The individual t-contrast images of the interaction from the PPI were examined using a group level one-sample t-test. As in a previous study, 8 the PPI were detected using t-test with a threshold of p<0.001 corrected for multiple comparison. Figure 1 (A) Helicopter view of the circular environment that participants (black arrow) explored (red trace) showing how the environment was split into one-half associated with danger (red) and the other with safety. The environment included two beehives (black dots) located at opposite sides of the environment. Participants were required to collect owers, which were generated randomly within the environment.    Illustration of sequential network activation in the ower task between groups. Left Panel: Helicopter view of task phase and schematic of activations over blocks. Middle and Right Panels: Brain activation and functional connectivity of HC (middle) and ANX (right). Green lines/boxes represent activation (and red arrows functional connectivity) that increased from the rst to second half of the experiment. Blue lines/boxes represent activation (and blue arrows functional connectivity) that decreased from the rst to second half of the experiment. Red lines/boxes represent activation (and red arrows functional