Decoding context memories for threat in large-scale neural networks

Abstract

Humans are often tasked with determining the degree to which a given situation poses threat. Salient cues present during prior events help bring online memories for context, which plays an informative role in this process. However, it is relatively unknown whether and how individuals use features of the environment to retrieve context memories for threat, enabling accurate inferences about the current level of danger/threat (i.e. retrieve appropriate memory) when there is a degree of ambiguity surrounding the present context. We leveraged computational neuroscience approaches (i.e. independent component analysis and multivariate pattern analyses) to decode large-scale neural network activity patterns engaged during learning and inferring threat context during a novel functional magnetic resonance imaging task. Here, we report that individuals accurately infer threat contexts under ambiguous conditions through neural reinstatement of large-scale network activity patterns (specifically striatum, salience, and frontoparietal networks) that track the signal value of environmental cues, which, in turn, allows reinstatement of a mental representation, primarily within a ventral visual network, of the previously learned threat context. These results provide novel insight into distinct, but overlapping, neural mechanisms by which individuals may utilize prior learning to effectively make decisions about ambiguous threat-related contexts as they navigate the environment.

Introduction

An important part of human survival involves the ability to adequately differentiate safety from threat when navigating the environment. In an effort to correctly infer the level of threat/danger, humans must often rely on important or salient cues surrounding prior events that shed light on the context in which one finds themselves (Anderson 1974; Hennings et al. 2020). In other words, cues help bring online context, which plays an important, informative role in detecting threat and making decisions (Tulving 2002; Gershman and Niv 2010). For example, consider 2 different conditions in which an individual encounters someone wearing a ski mask. If an individual encounters someone wearing a ski mask while at a ski resort, prior memories may be reinstated, either implicitly or explicitly, that allow the individual to infer that this particular context poses a low level of threat. However, if the individual encounters someone wearing a ski mask in the middle of the summer while walking into a bank, prior memories may be reinstated, either implicitly or explicitly, that allow them to infer that this particular context poses a high level of threat. In other words, in order for the individual in this example to reinstate an appropriate mental context for the current environment, they must be able to track and decode the stimulus features of their environment (e.g. the features of the ski resort vs the features of the bank).

However, detecting the degree to which a context is threatening based on current cues is not always that simple. Often times humans are only given minimal information about the current situation and yet must use what little information is at their disposal in an effort to make inferences about context and the corresponding level of threat (Cochran and Cisler 2019). Additionally, as is often the case in psychopathology (e.g. posttraumatic stress disorder), individuals are often prone to exhibit biased memory retrieval (e.g. responding as though in a threatening context, when actually safe) (Liberzon and Abelson 2016; Hennings et al. 2020). In fact, building upon the mental reinstatement of context literature, a relatively recent area of investigation in cognitive neuroscience centers around understanding whether individuals can accurately infer the danger characteristics of the environment when there is a degree of ambiguity surrounding the context in which an individual finds themselves and, if so, what are the neural mechanisms responsible for such an effect (i.e. how does the brain successfully decipher the correct level of threat) (Rissman and Wagner 2012; Trask et al. 2017; deBettencourt et al. 2019; Frankland et al. 2019; Hennings et al. 2020, 2022; Wammes et al. 2021; Cisler et al. 2023a). The mechanistic implications of this area of investigation are important, as the collective retrieval of stimulus features of the environment and contextual mental representations may inform the mechanisms by which individuals effectively make decisions about threat as they navigate their environment.

To address these questions, we administered a novel 2-day functional magnetic resonance imaging (fMRI) task and leveraged neuroimaging approaches to decode large-scale neural network activity patterns to probe neural mechanisms related to learning (day 1) and inferring (day 2) the correct threat context (i.e. unambiguous and ambiguous contextual conditions) given unique probabilities between stimuli and threat outcomes. In our task (see Fig. 1), adapted from a prior study of nonemotional learning (Chan et al. 2016), healthy participants learned on day 1 the probability between 3 different cues/conditioned stimuli (unique faces) and an aversive outcome (shock), which varied as a function of context (i.e. 4 different rooms clearly differentiated by color). On day 2, participants were not explicitly told which context (i.e. “room”) they were in (i.e. there were no color cues differentiating context) and instead were tasked with explicitly inferring the current context (i.e. “room”) based on their prior experience (i.e. learned associations between cues and aversive outcomes from 24 h earlier). In other words, to successfully accomplish this task, participants had to determine the characteristics of danger in their current environment by tracking the signal value of the different cues and relating to this back to learned information from day 1. Learning the mappings between CS-US pairings in the different contexts (i.e. “rooms”) was presumed to be a relatively straightforward process on day 1 as participants were given explicit information about the context and simply had to learn how threatening each cue was in a specific context by tracking the unique probabilities of shock associated with each of the different cues in each context. However, this process became much more complex on day 2 as participants were not given any explicit cues about the context (i.e. context/room color not explicitly provided; see Fig. 1C) and instead had to infer the correct context (i.e. “room”) based on tracking the signal value of each cue and relating these signal values back to the learned associations with context (i.e. “rooms”) formed on day 1. We postulated that if participants could adequately track the signal value of the different cues during this phase (i.e. identify how dangerous each cue was in a given context), then doing so would help reinstate the correct mental context.

Fig. 1.

Depiction of the mystery room task. There are 2 phases of the task separated into a learning (day 1; panel B) and inference (day 2; panel C) phase. During the learning phase (panel B), the participant’s objective is to learn a mapping between stimulus associations (CS1, CS2, CS3; 3 different neutral faces whose presentation coterminates with a shock of varying probabilities) and rooms (yellow, green, orange, blue; panel A). Participants are explicitly told what room they are in, and the background color and picture configuration for trials completed in that room are the same as the room they are in (i.e. yellow background with distinct room picture when in yellow room). Participants complete 21 trials in a given room and indicate on a trial-by-trial basis whether or not they expect to receive a shock following each of the stimuli. After completing all trials in a given room, participants provide ratings (0 to 10) on how likely they thought they were to receive a shock following each of the stimuli in that room, and were then explicitly taken to a different room. Participants complete 6 runs (2 rooms per run) and enter each room 3 different times for a total of 252 trials. Participants return 24 h later and first complete the inference phase (panel C) during which the participant’s objective is to infer which room they are in based on observed relationships between stimuli and shock outcomes. Each trial begins with the presentation of 1 of 3 neutral faces (CS1, CS2, CS3; same faces as the learn phase) whose presentation coterminates with a shock with varying probabilities (each room has identical properties as the learning phase). Participants are not told which room they are in, and instead, the background color is gray on all trials. Participants complete 21 trials in a given room and indicate on a trial-by-trial basis whether or not they expect to receive a shock following each of the stimuli. After completing the first 11 trials in a given room, participants are asked to rate (0 to 10) how likely it is that they are in each of the 4 rooms. Participants are then told they are still in the same room and then complete an additional 10 trials (as previously described), followed by room likelihood ratings (as previously described). After completing 21 trials in a given room, participants are explicitly taken to a different room. Participants complete 6 runs (2 rooms per run) and enter each room 3 different times for a total of 252 total trials. The order in which participants enter rooms during the infer phase is different than the order in which they enter rooms during the learn phase.

More specifically, our analytic approach allowed us to determine (i) whether individuals can correctly infer threat context (when there is a degree of ambiguity surrounding the context) based on information gleaned during the learning phase, (ii) whether neural representations of threat context and stimulus features of the task during the learning phase could be decoded, (iii) whether neural representations encoding context and the related stimulus features (i.e. shock probabilities) were reactivated during the inference phase, and (iv) whether reactivation of the neural representations of context could accurately predict participant’s presumed threat context (regardless of actual threat context). Given that contemporary models suggest that the human brain processes information across a number of spatially distributed networks, rather than separate brain regions (Bullmore and Sporns 2009, 2012; Bressler and Menon 2010; Meunier et al. 2010; Menon 2011; Avena-Koenigsberger et al. 2017; Pessoa 2017, 2018; Moughrabi et al. 2022; Cisler et al. 2023b), we primarily addressed the aforementioned questions via a large-scale neural network approach, as opposed to univariate voxel-wise ROI approaches. For example, the frontoparietal, default mode, and salience networks (often referred to as a triple network model) are 3 canonical neural networks that are consistently detected during resting-state and cognitively demanding tasks (Smith et al. 2009; Bressler and Menon 2010; Menon 2011) such as the task implemented in this study, which involves learning and inferring threat context. Additionally, striatum, frontoparietal, and ventral visual networks play pivotal roles in the salience tracking, cognitive control of memory retrieval, and mental reinstatement (Zink et al. 2003; Zink et al. 2004; Scimeca and Badre 2012; Rugg and Vilberg 2013; Hill et al. 2023). As such, we hypothesized that frontoparietal, default mode, salience, and striatum networks should track not only the stimulus features of the task but also the actual threat context in which participants find themselves in. Additionally, we hypothesized that when participants make inferences about a presumed threat context, there will be corresponding mental representations of that context within the ventral visual, salience, and frontoparietal networks.

Materials and methods

All procedures were approved by the Health Sciences Institutional Review Board at the University of Wisconsin—Madison, and all participants provided written informed consent to participate. The work described in this manuscript has been carried out in accordance with the Code of Ethics of the World Medical Association (Declaration of Helsinki) for experiments involving humans.

Study participants

Participants consisted of 38 healthy adults (15 males, 23 females, M age = 30.11 ± 7.86) recruited from community-based advertising including newspaper and online advertisements (e.g. Craigslist, Facebook, campus event pages), and flyers posted at approved locations on the University of Wisconsin—Madison campus and at health clinics throughout the greater Madison, WI area. One additional participant completed the study but was removed from analyses due to excessive head motion (i.e. no imaging runs were retained after identifying TRs in need of censoring, see below). Participants underwent structured clinical assessments (Weathers et al. 2013, 2018; Tolin et al. 2018) to confirm the absence of current mental health disorders and the Wechsler Abbreviated Scale of Intelligence (2nd edition) to rule out neurocognitive disabilities (Weschsler 2011). Exclusion criteria consisted of magnetic resonance imaging (MRI) contraindications, estimated full-scale IQ < 75, current mental health disorders, major medical disorders, loss of consciousness greater than 30 min, and age less than 21 or greater than 50.

Mystery room task

The novel task used in this study was designed to probe neural mechanisms of learning and inferring the correct threat context and was based on a prior investigation (Chan et al. 2016). There are 2 phases of the task (see Fig. 1) separated into a learning (day 1) and inference (day 2) phase. During the learning phase, the participant’s objective is to learn a mapping between stimulus associations (CS1, CS2, CS3; 3 different faces whose presentation coterminates with a shock of varying probabilities) and rooms (yellow, green, orange, blue). Participants are first presented with a map of 4 different rooms designated by different colors. Participants are then taken into 1 of the 4 rooms. They are explicitly told which room they are in, and the background color and picture configuration for trials completed in that room is the same as the room they are in (i.e. yellow background with distinct room picture when in yellow room). While in the room, participants complete 21 trials, during which they are presented with 1 of 3 neutral faces (CS1, CS2, CS3) whose presentation coterminates with a shock with varying probabilities depending on which room they are in (see Fig. 1). Each of the 4 rooms is associated with distinct probabilities, and the probability between any 1 stimulus and shock outcomes does not uniquely identify a room. In other words, participants must learn the mapping from room to stimuli based on information about all of the stimuli. Additionally, 2 of the rooms (yellow and green) are considered “low-threat” rooms due to the fact that there is only 1 stimulus in each room associated with a relatively high probability of shock, whereas the other 2 rooms (orange and blue) are considered “high-threat” rooms due to the fact that there are 2 stimuli in each room associated with a relatively high probability of shock. To promote participation engagement and learning, participants are asked on a trial-by-trial basis to indicate whether they think they will receive a shock or not at the end of the stimulus presentation. After completing 21 trials in a given room, participants are explicitly taken to a different room. Participants complete 6 runs (2 rooms per run) and enter each room 3 different times for a total of 252 total trials.

Participants return 24 h later and first complete an additional learn phase run outside the MRI environment (i.e. refresher task). Participants then complete the inference phase in the MRI scanner during which the participant’s objective is to infer which room they are in based on observed relationships between stimuli and shock outcomes. Participants are taken into a room (with a gray background, i.e. not using color cues to denote the correct room) where they complete 21 trials, during which they are presented with the 3 neutral faces (CS1, CS2, CS3) whose presentation coterminates with a shock with varying probabilities (each room has identical stimulus–shock probabilities as the corresponding room from the day 1 learn phase). Participants again are asked on a trial-by-trial basis to indicate whether they think they will receive a shock or not at the end of the stimulus presentation. Additionally, after completing the first 11 trials in a given room, participants are then asked to rate (0 = extremely unlikely; 10 = extremely likely) how likely it is that they are in each of the 4 rooms. Participants are then told they are still in the same room and then complete an additional 10 trials (as previously described), followed by room likelihood ratings (as previously described). After completing 21 trials in a given room, participants are explicitly told they are now taken to a different room. Participants complete 6 runs (2 rooms per run) and enter each room 3 different times for a total of 252 total trials. The order in which participants enter rooms during the infer phase is different than the order in which they enter rooms during the learn phase (see Fig. 1).

MRI procedures and first-level analyses

fMRI data were acquired on a GE MR750 3T scanner using a Nova 32-channel head coil. T1-weighted anatomic images were acquired with an MP-RAGE sequence (matrix = 256 × 256, 156 axial slices, TR/TE/FA = 8.2 ms/3.2 ms/12°, FOV = 25.6 cm, final resolution = 1 × 1 × 1 mm). A multiband sequence, in close approximation to the ABCD protocol, was used to collect the functional images using the following parameters: TR/TE/FA = 800 ms/25 ms/65, FOV = 21.6 cm, matrix = 90 × 90, 54 axial slices, slice thickness = 2.4 mm, acceleration factor = 6.

Image preprocessing followed standard steps and was completed using AFNI software. In the following order, images underwent field map correction, despiking, slice timing correction, deobliquing, motion correction using rigid body alignment, alignment to participant’s normalized anatomical images, spatial smoothing using an 8 mm FWHM Gaussian filter (AFNIs 3dBlurToFWHM that estimates the amount of smoothing to add to each dataset to result in the desired level of final smoothing), detrending, high-frequency (128 s) filtering, and rescaling into percent signal change. Images were normalized using the MNI 152 template brain. Following recommendations (Power et al. 2014), we corrected for head motion–related signal artifacts by using motion regressors derived from Volterra expansion, consisting of [R R2 Rt-1 R2t-1], where R refers to each of the 6 motion parameters, and separate regressors for mean signal in the CSF and WM. This step was implemented directly after motion correction and normalization of the EPI images in the image preprocessing stream. Additionally, we censored TRs from the first-level GLMs based on threshold of framewise displacement (FD) > 0.4. FD refers to the sum of the absolute value of temporal differences across the 6 motion parameters; thus, a cut-off of 0.4 results in censoring TRs where the participant moved, in total across the 6 parameters, more than ~0.4 mm plus the immediately following TR (to account for delayed effects of the motion artifact). Additionally, we censored isolated TRs where the preceding and following TRs were censored, and we censored entire runs if more than 50% of TRs within that run were censored. These TRs flagged based on motion were only censored during the first-level GLMs and were not removed from the data during the ICA analyses (see below). Following preprocessing, images were visually inspected (e.g. to ensure no distortions), and design matrices were created using AFNI’s 3dDeconvolve. The design matrices for the learn and infer phases of the task included columns for all CS stimulus presentations, shock delivery, and rating periods (onset and length).

Electric shock delivery

BIOPAC MP160 Data Acquisition System and BIOPAC STM100C module was used to administer shocks using pregelled electrodes placed on the skin of the fleshy portion of the mediolateral left lower leg directly over the tibialis anterior. Amperage on the stimulation device was set to the maximum (50 mA) to allow the greatest range of intensity selections. Participants were told to select an intensity of a 5/10, corresponding to a stimulus rated as uncomfortable, but not painful.

Independent component analysis

Toward the goal of identifying network-level neurocircuitry processes engaged while learning and inferring threat contexts (Ross and Cisler 2020; Crombie et al. 2021; Heilicher et al. 2022; Cisler et al. 2023b), we used independent component analysis

(ICA; Calhoun et al. 2001), which provides both a spatial map indicating distributed voxels that comprise the network and a time course of activity for the network. We used ICA (using all subjects’ data throughout the entire task, including the learn- and inference-phase imaging data) with a model order of 40 components, which delivered a good balance between component reliability estimated across 50 ICASSO iterations and interpretability of canonical networks. Eight of the 40 components were deemed functional networks of interest (in addition to a control network with peak loadings in the thalamus) after visual inspection (see Fig. 2). Networks of interest were chosen based on their involvement in saliency tracking, cognitive control of memory retrieval and mental reinstatement, and during cognitively demanding tasks (e.g. contextual threat learning) such as the task implemented in this study (Smith et al. 2009; Bressler and Menon 2010; Menon 2011; Scimeca and Badre 2012; Moughrabi et al. 2022; Cisler et al. 2023b). Components arising from artifacts of head motion or CSF and components of noninterest (i.e. motor, sensorimotor, and visual networks), which are not as relevant for the cognitive processes of interest here, were excluded. As a demonstration of specificity for the networks of interest, we also selected a network of noninterest (i.e. a network with peak loadings in the thalamus) to include in the main analyses.

Fig. 2.

Depiction of networks of interest and a control network (thalamus) obtained from ICA and used in subsequent MVPA (see Figs 3 and 4).

Multivariate pattern analyses

We differentiated the hypothesized separate processes of the mental reinstatement of the context vs tracking the shock probabilities unique to each context into 2 distinct multivariate pattern analyses (MVPA) pipelines (see Figs. 3 and 4).

Fig. 3.

Graphical overview of the MVPA for learn- and infer-phase neural data for stimulus feature decoder. An SVM classifier is first trained on the stimulus features of the task to decode shock frequency hyperplanes for each stimulus (CS1, CS2, CS3). Next, SVM classifiers are trained on ICA network patterns of neural activity to determine whether stimulus features of the task could be accurately decoded within ICA networks of interest. The SVM classifiers are then applied to the infer-phase data to determine whether stimulus feature hyperplanes predicted infer phase hyperplanes for the various different ICA networks. In other words, on each trial, a prediction was made for each of the rooms based on shock frequency hyperplanes for each room. This process is repeated until each participant has served as the left-out test participant.

Fig. 4.

Graphical overview of the MVPA for learn- and infer-phase neural data for room decoder. For each ICA network, trial × voxel matrices of beta coefficients are created for all participant’s data (separately for learn and infer data). SVM classifiers are then trained on N-1 subject’s learning data, resulting in separate decoders for each of the rooms within a given stimulus (CS1, CS2, CS3). Next, these decoders are applied to the trial × voxel matrices of beta coefficients for the infer-phase data. This results in a prediction about the degree to which the learned state representations are active during inference. This process is repeated until each participant has served as the left-out test participant.

MVPA of the CS shock probabilities unique to each context

We first sought to test the hypothesis that specific networks may be encoding stimulus features of the context (e.g. unique shock frequencies on a trial × trial basis), and that tracking of these stimulus features could accurately be decoded during the learn phase and would reactivate accordingly during the infer phase. Figure 3 provides an overview of the analytical approach to address this question, which was implemented separately for each of the 8 ICA networks of interest (and 1 control network). The first step of this analytic approach was to demonstrate that stimulus features of the task (i.e. shock frequencies to each of the CS on a trial × trial basis) could be accurately decoded for each subject using support vector machine (SVM) regression. To create a trial-by-trial indicator of the relative likelihood of room based solely on veridical CS shock frequencies, we first trained a simple SVM decoder to predict rooms based only on the trial-by-trial shock frequencies unique to each room during the learn phase. The output of this simple SVM decoder would then be trial-by-trial likelihood of room based on the current CS shock frequencies. The training features for this simple decoder were the veridical trial-by-trial shock frequencies for each face (i.e. a 252 [row] × 3 [column] matrix, where each column is one of the CSs and represents the trial-by-trial true shock probability for that CS). Thus, this simple SVM decoder learns to predict the room based on configurations of shock frequencies for each CS. A separate simple SVM decoder was created for each of the 4 rooms using the libsvm function in MATLAB using options for support vector regression with a radial basis function kernel, the slack “c” parameter = 1, and accounting for the less number of target rooms (e.g. yellow) to other rooms (e.g. blue, orange, green) by weighting the observations accordingly. These 4 simple SVM decoders (i.e. a separate decoder for each room) then provided trial-by-trial hyperplane predictions about the relative likelihood of each room on each trial (i.e. 4 hyperplane predictions on each trial, representing the likelihood of each room based on the current CSs’ shock probabilities).

These hyperplanes from the simple SVM decoder were then used as the training labels for SVM regression decoders applied to the network activity patterns. That is, a second SVM decoder was trained to identify network patterns that activated in accordance with the trial-by-trial room predictions based on the shock frequencies (rather than the true room labels). Each participant’s trial-by-trial activation patterns were characterized using AFNI’s 3dLSS function, which allowed us to derive a beta value for every voxel toward each trial’s CS. The voxel-wise beta values were then used as inputs to these SVMs. The timepoint × voxel matrices were centered within each timepoint to ensure no difference in overall activation levels across trials. We established the accuracy of these decoders using a leave-one out approach, such that the decoders were trained on all participants’ neural activity except for 1 left-out participant, and the resulting decoders were applied to the left-out participant’s data. This process continued until each subject served as the test subject (i.e. the number of iterations was equal to number of participants). This process also allowed us to determine decoder accuracy for each participant, which we could then control for in subsequent analyses (i.e. include decoder accuracy for each subject as covariate). Any network for which the SVM decoder outperformed chance (r- to z-transformed values of relationship between hyperplanes and decoded shock frequencies that were statistically greater than 0 for all 4 rooms as verified via 1-sample t-tests) were deemed accurate.

After establishing accuracy of these stimulus feature decoders, we used the same leave-one-out approach to build a decoder based on data from the learn phase and apply the resulting decoder to the left-out participant’s infer phase data. The simple SVM regression decoder from the learn data was applied to the trial-by-trial observed shock frequencies from the infer phase, resulting in trial-by-trial predictions about the likelihood of each room during the infer phase. The network activity SVM decoder, trained from the learn phase data, was also applied to network patterns during the infer phase, resulting in hyperplane predictions representing the degree to which network patterns during the infer phase reactivated in accordance with predictions for each room based on shock frequencies based on the CS-shock probabilities in each room learned on day 1.

MVPA of mental representations of threat

We also sought to test the hypothesis that mental representations of threat context that were activated in response to CSs during learning were reactivated during presentations of CSs during inference and were related to participant’s guesses about specific threat contexts when given imprecise information. Figure 4 provides an overview of the analytical approach to address this question, which was implemented separately for each of the 8 ICA networks of interest (and 1 control network). The first step of this analytic approach was to demonstrate that network activity patterns in response to CSs to each room during the learning task could accurately be decoded for each subject (e.g. are there unique CS patterns for yellow vs blue rooms?). We implemented a similar methodological process as described above (see Fig. 4), although room labels were used instead of shock frequencies to each of the CSs on a trial × trial basis when building classifiers. Each participant’s trial-by-trial activation patterns were characterized using the 3dLSS function, which allowed us to derive a beta value for every voxel on each trial. The voxel-wise beta values were then used as inputs to the SVMs. The timepoint × voxel matrices were centered within each timepoint to ensure no difference in overall activation levels across trials. SVM, again implemented in MATLAB through libsvm and using a radial basis function kernel with a slack “c” parameter = 1 (Chang and Lin 2011), was used to decode specific rooms for each CS. We trained a separate classifier for each CS to differentiate rooms using binary classification (e.g. for CS1, we trained a yellow vs all other room classifiers, a green vs all other room classifiers), and accounted for imbalanced classes by weighting the classes within the libsvm function. We established the accuracy of the decoders using a leave-one out approach, such that the decoders were trained on all participants’ neural activity except for 1 left-out participant, and the resulting decoders were applied to the left-out participant’s data. This process continued until each subject served as the test subject (i.e. number of iterations was equal to number of participants). This process also allowed us to determine decoder accuracy for each participant, which we could then control for in subsequent analyses (i.e. include classifier accuracy for each subject as covariate). Any network for which the classifiers outperformed chance (cross-validation accuracy significantly greater than 50% for all 4 rooms as verified via 1-sample t-tests) was deemed accurate. This process was repeated separately for each of the 8 functional networks of interest (and 1 control network).

After testing the accuracy of the decoders, the next step was to apply the decoders built during the learning task to the participant’s network patterns of neural activity during the inference phase. Again, 3dLSS was used to define trial-by-trial activation. Applying the decoders to participant’s data during the infer phase resulted in hyperplane distances representing the degree to which the trained multivariate patterns of a given room were active during the presentation of a given CS. For the infer-phase mental representation of threat data, we focused on the last 10 trials from each time a participant entered a room (i.e. trials in the latter half of each room after providing their initial room rating guesses for that room). This process was repeated separately for each ICA network of interest (for which the classifier was accurate), resulting in unique predictions (i.e. hyperplane distances) about threat (i.e. rooms) for each separate network.

Region-of-interest analyses

Additionally, we repeated the aforementioned processes within select regions of interest (ROIs: amygdala, hippocampus, insula, and ventral/medial prefrontal cortex) in order to determine if specific regions not entirely captured in our large-scale neural network analyses (or those for which network activity could not be accurately decoded by classifiers) were also encoding stimulus features of the task and threat context. ROIs were selected from the corresponding regions in the Harvard Oxford atlas and resampled into the normalized space matching the current participant dataset. These specific ROIs were selected based on prior work demonstrating a role for these regions in threat/fear learning, prospective neural replay during decision-making, spatial navigation, and episodic future thinking (Doll et al. 2015; Shadlen and Shohamy 2016; Schacter et al. 2017; Gillespie et al. 2021; Widloski and Foster 2022). We ensured appropriate signal in ROIs where dropout is common (e.g. ventral/medial prefrontal cortex [vmPFC]) by restricting the vmPFC ROI to voxels within a group mask of common voxels.

Experimental design and statistical analyses

Behavioral analyses

In regard to the behavioral ratings during the learn phase, our primary analyses focused on whether participants were able to accurately learn threat probabilities for each of the faces (CS1, CS2, CS3) in each of the rooms (i.e. accurately learned task structure). Based on the structure of the task, the following linear mixed-effects models (LMEMs) were used for the CS1 and CS3 stimuli: shock expectancies for CS1 and CS3 (separate models) ~ threat context contrast 1 (yellow vs green room) × room repetition (z-scored 1st through 4th entry into room) + threat context contrast 2 (orange vs blue) × room repetition + (1 | subject); whereas the following LMEM was used for the CS2 stimuli: shock expectancies for CS2 ~ threat context contrast 1 (yellow vs orange room) × room repetition (z-scored 1st through 4th entry into room) + threat context contrast 2 (green vs blue) × room repetition + (1 | subject). In regard to the behavioral ratings during the infer phase, our primary analyses focused on whether participants were able to correctly infer the actual room they were in and differentiate high- from low-threat rooms. The following LMEMs were used to examine this effect: room rating ~ predicted/inferred room threat (−1 = low threat; 1 = high threat) × actual/true room threat (−1 = low; 1 = high) + discrimination coding (.75 = actual room in; −.25 = other rooms) + (1 | subject).

MVPA of CS shock probabilities unique to each context

In regard to the stimulus feature analyses (i.e. MVPA of the CS shock probabilities unique to each context), our primary interest focused on examining whether network patterns that encoded the shock probabilities during the learn phase were reactivated in accordance with observed shock probabilities during the infer phase. In this case, the main regressor of interest were the hyperplane predictions from the simple SVM decoder of the shock probabilities from the learn phase applied to the shock probabilities of the infer phase. The following LMEM was used to test these effects within each network: hyperplane predictions of network activity patterns for each room (z-scored) ~ stimulus feature hyperplane predictions for each room × room threat level (−1 = low threat; 1 = high threat) + mean number of TRs per run (z-scored) + (1 | subject). In the event that the stimulus features of the task (i.e. shock frequencies to each of the CS) could have been accurately decoded in any of the ROIs, identical analyses would have been implemented, simply replacing ICA networks for ROIs.

MVPA of mental representation of threat context

In regard to the mental representations of threat context analyses, our primary interest focused on examining the degree to which network activity patterns during learning were reactivated during the infer phase. We differentiated between 2 models by which participants might have learned and recalled the context—CS shock probability mappings: (i) a discrimination model (Fig. 5A) and (ii) a threat magnitude model (Fig. 5B). The discrimination model hypothesizes that participants differentiate a specific room from all other rooms, regardless of the level of threat of that room. The threat magnitude model hypothesizes that participants might demonstrate structured errors in recall, such that they confuse the 2 low-threat rooms with each other and/or the 2 high-threat rooms with each other (i.e. predicted/inferred threat × actual/true threat interaction). These 2 models are not mutually exclusive. Our analyses also allowed us to examine whether participants/classifiers were biased toward low- or high-threat rooms regardless of room (i.e. main effect of predicted/inferred threat; see Supplementary Fig. 1) or biased toward a low- or high-threat room when in a low- or high-threat room, respectively (i.e. main effect of actual/true threat; see Supplementary Fig. 1). The following LMEMs were used to test these effects within each network: hyperplane predictions for each room (z-scored within each model for a given room) ~ predicted room threat (−1 = low threat; 1 = high threat) × actual room threat (−1 = low; 1 = high) × mean number of TRs per run (z-scored) + predicted room threat × actual room threat × cross-validation classifier accuracy for each subject (z-scored mean accuracy to each CS in all 4 rooms) + discrimination coding (0.75 = actual room in; −0.25 = other rooms) × mean number of TRs per run × cross-validation classifier accuracy for each subject (z-scored mean accuracy to each CS in all 4 rooms) + (1 | subject). Z-scoring the hyperplane predictions within each model of a given room was important for controlling for any differences in decoder accuracy between the models (e.g. if the model for the green room was better than the model for the yellow room).

Fig. 5.

Prototypical effects of interest based on linear mixed-effects model analyses. Greater hyperplane predictions for a given room when actually in that room depicts participant’s or classifier’s ability to accurately infer the correct room (main effect of discrimination coding; panel A). Greater hyperplane predictions for yellow and green rooms when actually in yellow and green rooms and greater hyperplane predictions for orange and blue rooms when in orange and blue rooms depicts differentiation between low- and high-threat rooms when in low- and high-threat rooms, respectively (true × inferred room interaction; panel B).

Additionally, participants provided ratings throughout the infer phase pertaining to how likely they thought they were in each room based on information gleaned from the trials within a given room (i.e. shock probabilities to each CS). This allowed us to determine whether there was a relationship between participant’s room guesses (i.e. room with the highest ranking) and hyperplane predictions for each of the various rooms, including the corresponding room that participants thought they were in. In these analyses, we removed all trials for which there was not 1 room with higher ratings compared to all other rooms in order to ensure participants were making a clear guess for a unique room. The following LMEM was used to test these effects within each network: hyperplane predictions for each room (z-scored) ~ participant room guess (dummy coded 1 = participant’s guessed room, 0 = participant did not guess this room) × mean number of TRs per run (z-scored) + cross-validation classifier accuracy for each subject (z-scored mean accuracy to each CS in all 4 rooms) × participant room guess + (1 | subject). In the event that the threat context (i.e. rooms) could be accurately decoded in any of the ROIs, identical analyses would be implemented, simply replacing ICA networks for ROIs.

Results

Behavioral ratings during learn phase

Prior to large-scale network analyses, we first tested whether participants could accurately learn threat probabilities for each of the faces (CS1, CS2, CS3) in each of the rooms during the learn phase (i.e. accurately learn the task structure; see Fig. 6A–C). We established that participants accurately learned the task structure as mean shock expectancies during the learn phase mapped onto the true probabilities of each face within each room (see Fig. 6A–C and Supplementary Fig. 2). Specifically, for the CS1 stimuli, there were significant main effects for the yellow vs green (t[598] = −14.79, P < 0.001) and orange vs blue (t[598] = −10.78, P < 0.001) contrasts. For the CS2 stimuli, there were significant main effects for the yellow vs orange (t[598] = −29.98, P < 0.001) and green vs blue (t[598] = −29.31, P < 0.001) contrasts. Additionally, for the CS3 stimuli, there were significant main effects for the yellow vs green (t[598] = 15.38, P < 0.001) and orange vs blue (t[598] = 12.14, P < 0.001) contrasts.

Fig. 6.

True and mean probabilities for each stimulus within each room during learn phase (panels A–C) and mean inference ratings for likelihood of being in a given room throughout the infer phase of the task (panels D and E). Panel A depicts true probabilities for each facial stimulus (CS1, CS2, CS3) within each room. Panel B depicts mean shock expectancies for each facial stimulus within each room at end of learning in MRI environment (block 3; third time in room). Panel C depicts mean shock expectancies for each facial stimulus within each room at end of learning during refresher phase (prior to infer phase) outside MRI environment (block 4; fourth time in room). Participants accurately learned threat probabilities for each of the faces (CS1, CS2, CS3) in each of the rooms during the learn phase (i.e. accurately learned the task structure). Panel D depicts mean inference ratings when a participant is first being asked the likelihood of being in a given room (trials 1 to 11) during block 3 (third time in room; end of inference). Panel E depicts mean inference ratings when a participant is asked for the second time the likelihood of being in a given room (trials 1 to 21) during block 3 (third time in room; end of inference). Participants were able to accurately infer the actual room they were in (main effect of discrimination coding) and differentiate between low- and high-threat rooms (actual/true × predicted/inferred threat interaction; see Fig. 5 for prototypical effects).

Behavioral ratings during infer phase

We next tested whether participants could correctly infer the actual room they were in and differentiate high- from low-threat rooms during the inferential learning phase. Our results revealed that there was a significant main effect of discrimination coding (t[603] = 2.85, P = 0.004), indicating that participants generally provided higher ratings for the actual room they were in (i.e. correctly identified actual room they were in from all other rooms; see Fig. 6D and E and Supplementary Fig. 3), and a significant actual/true threat × predicted/inferred threat interaction (t[603] = 7.62, P < 0.001), indicating that participants provided higher ratings for low-threat rooms and lower ratings for high-threat rooms when actually in a low-threat room, and vice versa when actually in high-threat rooms (i.e. lower ratings for low-threat rooms and higher ratings for high-threat rooms; see Fig. 6D and E and Supplementary Fig. 3).

MVPA

MVPA of the CS shock probabilities unique to each context

We first sought to test the hypothesis that specific networks may be encoding stimulus features of the context (e.g. unique shock frequencies on a trial × trial basis) and that tracking of these stimulus features could accurately be decoded during the learn phase and would reactivate accordingly during the infer phase. Results revealed that stimulus features of the task (i.e. shock frequencies to each of the CS on a trial × trial basis) could be accurately decoded (using SVM regression) for 7 of the 8 networks of interest and not the control network (see Fig. 7A). As a result, we used a Bonferroni correction to control for multiple comparisons across networks in all subsequent analyses (i.e. 0.05/7 = corrected P < 0.0071). The decoders were then applied to the infer-phase data to determine whether stimulus features of the task predicted greater hyperplane distances for the corresponding room (i.e. trial × trial predictions about the likelihood of each room based on the shock probabilities rather than the true room labels). LMEMs revealed that this effect was evident within the left frontoparietal (t[34,427] = 7.30, P < 0.001), right frontoparietal (t[34,427] = 5.46, P < 0.001), posterior default mode (t[34,427] = 6.33, P < 0.001), striatum (t[34,427] = 4.12, P < 0.001), salience (t[34,427] = 8.83, P < 0.001), and inferior frontal gyrus (t[34,427] = 2.80, P = 0.005) networks (see Fig. 8 and Supplementary Fig. 4), but not the ventral visual network (t[34,427] = 0.27, P = 0.789; see Supplementary Fig. 5).

Fig. 7.

SVM stimulus features of the task decoder (panel A) and SVM room decoder (panel B) results for each of the 8 networks of interest (and 1 control network) from the ICA. SVM regression and SVM classification revealed that the classifiers performed better than chance in predicting outcomes for 7 of the 8 networks (i.e. r- to z-transformed values of relationship between hyperplanes and decoded shock frequencies that were statistically greater than 0 for all 4 rooms as verified via 1-sample t-tests for SVM stimulus features of task decoder and cross-validation accuracy greater than 50% for all 4 rooms as verified via 1-sample t-tests for SVM room decoder). The anterior default mode (a network of interest) and the thalamus (a control network) networks did not meet significance for all stimuli and therefore were not included in subsequent analyses.

Fig. 8.

Depiction of MVPA of stimulus features of task results. Figures depict beta-coefficients from models predicting respective network hyperplane predictions for each room from stimulus feature (shock frequency) hyperplane predictions. The omnibus LMEMs revealed this effect was significant within the left frontoparietal (panel A), right frontoparietal (panel B), posterior default mode (panel C), striatum (panel D), salience (panel E), and inferior frontal gyrus (panel F) networks.

MVPA of mental representations of threat context

We also sought to test the hypothesis that mental representations of threat context that were activated in response to CSs during learning were reactivated during presentations of CSs during inference and were related to participant’s guesses about specific threat contexts when given imprecise information. The cross-validation tests demonstrated that threat context (i.e. rooms) could be decoded accurately in 7 of the 8 networks of interest and not in the control network (see Fig. 7B). After testing the accuracy of the decoders (built during learn-phase data), the decoders were then applied to the infer-phase data to determine whether the decoders were accurate in predicting threat context based on network patterns of activity during the presentation of a given CS (i.e. degree to which trained multivariate patterns of a given room were active during presentation of a given CS). LMEMs revealed significant main effects of discrimination coding in the right frontoparietal (t[16,385] = 2.98, P = 0.003), striatum (t[16,385] = 3.77, P < 0.001), and salience (t[16,385] = 3.34, P < 0.001) networks, indicating generally greater hyperplane predictions within the aforementioned networks for the actual room participants were in compared to the other rooms (see Fig. 9 and Supplementary Fig. 6). In other words, participants reinstated the neural patterns in these networks that were engaged during learning when they re-entered the rooms during the infer-phase CS presentations.

Fig. 9.

Mean hyperplane prediction values after applying room decoder (created from learning phase data) to infer phase data for the right frontoparietal (panel A), striatum (panel B), salience (panel C), and left frontoparietal (panel D) networks. Discrimination of the correct room from all other rooms (main effect of discrimination coding) was significant for the right frontoparietal, striatum, and salience networks. The classifier was also able to accurately differentiate between low- and high-threat rooms (true × inferred danger interaction) based on participant’s neural representations within the salience and left frontoparietal networks.

LMEMs also revealed significant predicted/inferred threat × actual/true threat interactions in the salience (t[16,385] = 3.29, P = 0.001) and left frontoparietal (t[16,385] = 2.70, P = 0.006) networks, indicating participants were more likely to engage representations of low-danger rooms when actually in low-danger rooms and high-danger rooms when actually in high-danger rooms within the salience and left frontoparietal networks (see Fig. 9 and Supplementary Fig. 6). In other words, participants reinstated the neural patterns in these networks that were engaged during learning for rooms with similar threat magnitude, but not necessarily the correct room.

In addition to neural activity, participants also provided ratings throughout the infer phase pertaining to how likely it was that were in a given room (i.e. had to use information gleaned from task [shock probabilities to each CS]). As a result, we included the participant’s guessed room (i.e. the room with the highest rating, limited to trials in which there was a unique guessed room) as a contrast variable (guessed room vs other rooms) in an additional LMEM (described above) and found there was a significant effect for the left frontoparietal (t[16,394] = 3.28, P = 0.001), salience (t[16,394] = 3.28, P = 0.001), and ventral visual (t[16,394] = 2.76, P = 0.005) networks, indicating greater hyperplane predictions within the left frontoparietal, salience, and ventral visual networks for rooms that participants believed were most likely, regardless of the actual room they were in (see Fig. 10 and Supplementary Fig. 7).

Fig. 10.

Mean hyperplane prediction values for each room for trials where participants thought they were in a given room (i.e. participant room guess analyses). The room that participants thought they were most likely in resulted in greater hyperplane predictions for respective rooms within the left frontoparietal (panel A), salience (panel B), and ventral visual (panel C) networks.

Additionally, given that our aforementioned analyses revealed that stimulus features of the task predicted greater hyperplane distances for the corresponding room for all networks other than the ventral visual network (see MVPA of the CS shock probabilities unique to each context section), we then hypothesized that perhaps the tracking of stimulus features in the salience network (i.e. because it had the greatest effect size in this analysis) may help bring online the mental representation of a given room within the ventral visual network when an individual is tasked with inferring threat context (i.e. infer phase). In other words, we examined whether hyperplane predictions from the stimulus feature analyses (within the salience network) predicted hyperplane predictions from the participant room guess analyses (within the ventral visual network). Results revealed that salience network hyperplane predictions from the stimulus feature analyses predicted ventral visual hyperplane predictions from the participant room guess analyses (t[16,383] = 12.35, P < 0.001). However, this effect was not specific to the salience network, as the effect was also evidence within the left frontoparietal (t[16,383] = 14.28, P < 0.001), right frontoparietal (t[16,383] = 15.15, P < 0.001), posterior default mode (t[16,383] = 15.68, P < 0.001), striatum (t[16,383] = 21.10, P < 0.001), and inferior frontal gyrus (t[16,383] = 11.83, P < 0.001) networks.

ROI analyses

Additionally, we repeated the aforementioned processes within select ROIs (amygdala, hippocampus, insula, ventral/medial prefrontal cortex, and orbitofrontal cortext) to determine if specific regions not entirely captured in our large-scale neural network analyses (or those for which network activity could not be accurately decoded by classifiers) were also encoding stimulus features of the task and threat context. Results revealed that neither the threat context (i.e. rooms) nor the stimulus features of the task (i.e. shock frequencies to each of the CS) could accurately be decoded in any of the ROIs. As a result, no further analyses were conducted.

Discussion

The current study administered a novel 2-day fMRI task and leveraged ICA and MVPA approaches to probe neural mechanisms of learning and inferring correct threat contexts. In this study, individuals had to determine the threat context by tracking the signal value of several different cues. During the learn phase of the task, individuals were able to learn the level of danger that each cue (face) presented in each specific context, which was unsurprising given that context was explicitly stated (i.e. participants were informed which room they were in). Although this process became much more complex during the infer phase when participants were placed in ambiguous contexts during the day 2 inference phase (i.e. context not explicitly cued and instead must be inferred), participants were still generally able to correctly infer the threat context in which they found themselves in. Our neuroimaging analyses corroborated these findings and suggest that individuals accurately infer threat contexts under ambiguous conditions, in part, due to neural reinstatement (specifically striatum, salience, and frontoparietal networks) of multivariate patterns of brain activity that tracks the signal value of environmental cues (i.e. identify how dangerous each cue is in a given context), which, in turn, allows individuals to reinstate a mental representation of context in a ventral visual network.

More specifically, we differentiated between a discrimination and a threat magnitude model. Analyses revealed support for the discrimination model in the right frontoparietal, striatum, and salience networks (i.e. participants reinstated neural activity patterns during inference that were unique to the specific room from learning), and our threat magnitude model within the salience and left frontoparietal networks (i.e. participants reinstated neural activity patterns during inference in accordance with a room threat magnitude, but not the specific room). It should be noted that support for neural activity reinstatement associated with both discrimination of contexts and discrimination of threat magnitude is consistent with behavioral results, where we observed that participants not only generally correctly identified the specific room but also demonstrated confusion between low- and high-threat contexts. Furthermore, our analyses revealed that if participants thought they were in a given context (i.e. a specific room) during the infer phase, patterns of neural activity for that specific guessed room were reinstated in the left frontoparietal, salience, and ventral visual networks. Finally, we found that the stimulus features of the task (i.e. shock frequencies to each CS on a trial-by-trial basis) could accurately be decoded in the left and right frontoparietal, posterior default mode, striatum, salience, and inferior frontal gyrus networks. Given the role of the salience network in detecting threat-relevant stimuli (Seeley et al. 2007; Menon and Uddin 2010; Uddin 2015; Fullana et al. 2016; Marstaller et al. 2021) (paired with the fact that the ventral visual network did not track stimulus features of the task), we then hypothesized that the salience network tracks the stimulus features of the task, which helps bring online the mental representation of the threat context (room) within the ventral visual network. Although we found support for this, we also found that network activity patterns in accordance with shock probabilities across all other networks similarly predicted network activity reinstatement for a given specific room in the visual network, suggesting a broad correspondence between tracking stimulus features and mental reinstatement (rather than an effect specific to the salience network). Finally, in the context of understanding neural mechanisms of threat, it is also relevant to mention that laboratory analogues of threat (e.g. mild electric shocks) are not necessarily generalizable to real-world threats (e.g. physical assault), though they do enable studying of threat in an ethical way in laboratory settings.

One interpretation of our behavioral and neural results is that participants were able to identify stimuli as occasion setters (Trask et al. 2017). That is, participants were generally able to identify that particular stimuli in certain contexts informed them about the relationship between other stimuli and the accompanying degree of danger that those stimuli present in various different contexts. Our results also suggest that the contextual mental representation (and accompanying neural signature) formed during learning is reinstated as individuals infer threat-related contexts. In other words, our findings add to a growing body of research (Doll et al. 2015; Shadlen and Shohamy 2016; Widloski and Foster 2022) that suggest that the retrieval of contextual mental representations (i.e. context reinstatement) may be a mechanism by which individuals effectively navigate their environment. Although this task involved deciphering threat context, prior neuroimaging research has demonstrated that individuals that make decisions (in pursuit of reward) according to a model-based approach exhibit a neural signature of prospective future paths at the time of decision (Doll et al. 2015). Additionally, we recently administered an approach-avoidance task and used MVPA to demonstrate that decisions to approach reward vs avoid threat were predicted by the relative degree to which prior reward vs threat representations were active at the time of choice (Moughrabi et al. 2022). Relatedly, it has also recently been proposed that the recall of past memories informs prospective reasoning, which aids in value-based decision-making (Shadlen and Shohamy 2016). Although speculative, this may shed light on why individuals with aberrant or biased memory retrieval (e.g. PTSD) often engage in avoidance behaviors (i.e. biased memory retrieval leads to incorrect inference about whether a social invitation, for example, may lead to a rewarding/worthwhile experience), though this hypothesis remains to be tested. Incorporating information from the past to make prospective predictions about future outcomes, including threat, is advantageous for an organism’s survival. Although the task administered as part of this study did not allow individuals to avoid aversive outcomes if they correctly inferred threat context, our results suggest that neural reinstatement of prior experiences (i.e. context reinstatement) allows for identification of threat-related contexts, thereby enabling defensive or otherwise adaptive behaviors.

A contemporary view in the field of memory research is that context may play a central role in episodic memory (Bornstein and Norman 2017; DuBrow et al. 2017; Hennings et al. 2020). In general, in the event that individuals are able to successfully recall information about a given stimuli (for example), then that specific retrieval of information is thought to also trigger the accompanying context associated with the given stimuli. In fact, this retrieval of context can almost be viewed as a sort of prerequisite for episodic memory (i.e. reinstatement of contextual information is what separates an episodic memory from a simple factual experience). The current study found evidence in support of this notion as individuals were able to use information surrounding stimuli/shock contingences to correctly infer threat context. Recent research from the fear-conditioning literature also supports this finding. For instance, using a novel paradigm that embedded contextual tags (nature scenes) between the presentation of conditioned stimuli during extinction, researchers found that individuals without PTSD not only exhibited greater extinction recall compared to individuals with PTSD but also MVPA provided evidence of a neural signature in the dorsal/ventral prefrontal cortex that helped individuals accomplish the successful differentiation of memories for fear and safety (Hennings et al. 2020, 2022). In other words, individuals who exhibited intact extinction recall during tests of fear renewal were reinstating a mental representation of the context (nature scenes) initially presented as extinction memories were formed. Although the task administered in this study was not a traditional fear conditioning or episodic memory task, results from this study similarly suggest that healthy individuals (generally) correctly infer threat and may do so by bringing online a pattern of neural activity (within the left and right frontoparietal, striatum, salience, and ventral visual networks) that is consistent with the pattern of activity present when individuals are explicitly learning stimuli/shock contingences associated with a given threat context.

This study also revealed that the left and right frontoparietal, posterior default mode, striatum, salience, and inferior frontal gyrus networks track the stimulus features of the task on a trial-by-trial basis. Interestingly, this effect was not present in the ventral visual network. One hypothesis explaining the differential neural reinstatement in the visual network is that there are dissociable processes for tracking stimulus features of the environment, which then allows for identification of a threat context and subsequent signal propagation to the ventral visual network to bring online a mental representation of that context. Although we found evidence in support of this hypothesis, the effect was also found in other networks, which limits the degree of specificity that can be assigned to this particular mechanistic hypothesis. Further research is needed to tease apart precisely how individuals are capable of brining online the correct mental representation of a threat context when given imprecise information. It is also worth noting that the decoding of our learning-phase data was arguably not solely capturing perceptual attributes but also metacognition/decision processes, given that participants were also required to predict whether they would receive a shock on a trial-by-trial basis.

Finally, we conducted supplementary ROI analyses given that there were several ROIs (e.g. amygdala, hippocampus, orbitofrontal cortex) that were not entirely captured in our large-scale neural network analyses, but play a role in threat/fear learning, prospective neural replay during decision-making, spatial navigation, and episodic future thinking (Doll et al. 2015; Shadlen and Shohamy 2016; Schacter et al. 2017; Gillespie et al. 2021; Widloski and Foster 2022). Interestingly, we found that neither the stimulus features of the task (i.e. shock frequencies to each of the CS) nor the threat context (rooms) could be decoded in select ROIs, which lends support to contemporary models that suggest the brain processes information across spatially distributed networks. It is important to reiterate that participants completed a relatively complicated, dynamic task in which they are continually learning from their environment on a trial-by-trial basis as opposed to simply reacting to stimuli in a static environment (e.g. tasks assessing amygdala reactivity). Given the complexity of this task, it is unsurprising that we only observed effects when examining temporally and spatially connected brain networks that were responsive to the task, as opposed to discrete, brain regions. Additionally, we tested less than 40 subjects, and as such, it is possible that we were simply underpowered to detect significant effects with our ROI analyses.

Overall, these results provide novel insight into distinct, but overlapping, neural mechanisms by which individuals may utilize prior learning to effectively make decisions about ambiguous threat-related contexts as they navigate the environment. Although additional study is needed among larger and more clinically diverse populations (e.g. PTSD), this study adds to the growing body of knowledge aimed at understanding the neural mechanisms of learning and inferring correct threat contexts.

Author contributions

Kevin M. Crombie (Conceptualization, Data curation, Formal analysis, Investigation, Methodology, Project administration, Software, Validation, Visualization, Writing—original draft), Ameera Azar (Data curation, Formal analysis, Investigation, Project administration, Writing—review & editing), Chloe Botsford (Investigation, Project administration, Writing—review & editing), Mickela Heilicher (Investigation, Project administration, Writing—review & editing), Michael Jaeb (Investigation, Project administration, Writing—review & editing), Tijana Sagorac Gruichich (Investigation, Project administration, Writing—review & editing), Chloe M. Schomaker (Investigation, Project administration, Writing—review & editing), Rachel Williams (Investigation, Project administration, Writing—review & editing), Zachary N. Stowe (Project administration, Resources, Writing—review & editing), Joseph E. Dunsmoor (Conceptualization, Methodology, Visualization, Writing—review & editing), and Josh M. Cisler (Conceptualization, Formal analysis, Funding acquisition, Investigation, Methodology, Project administration, Resources, Software, Supervision, Validation, Visualization, Writing—review & editing)

Funding

This study was supported by the National Institute of Mental Health (R01MH119132) awarded to J.M.C. K.M.C. was supported by a National Institute on Alcohol Abuse and Alcoholism (NIAAA/NIH) training grant (T32AA007471) and is currently supported by the National Institute of Mental Health (K01MH132545). The funding sources had no role in the collection, analysis, and interpretation of data; in writing the report, or in the decision to submit the article for publication.

 

Conflict of interest statement: None declared.

Data availability

The data that support the findings of this study are available from the corresponding author upon reasonable request.