A neurally constrained computational model of context-dependent fear extinction recall and relapse

Abstract

Exposure therapy, a standard treatment for anxiety disorders, relies on fear extinction. However, extinction recall is often limited to the spatial and temporal context in which extinction is learned, leading to fear relapse in new settings or after delays. Animal studies offer insights into fear extinction in humans. Computational models that integrate these findings into a neurally grounded framework, while generating testable hypotheses for humans, can bridge this gap. Current models either focus on neuron-level activity, limiting their scope, or abstract away entirely from neural mechanisms. They also often overlook the distinct contributions of cue and context in fear extinction and recall. To address these gaps, we present ConFER, a neurally constrained model of fear extinction, recall, and relapse. ConFER integrates findings from the neural fear circuit, modeling distinct pathways for cue and context processing. These pathways independently activate positive and/or negative memory engrams in the basolateral amygdala, competing to determine the fear response. ConFER simulates fear renewal and spontaneous recovery across context combinations, while generating novel, testable predictions. Notably, it predicts counterconditioning may better prevent relapse than extinction in new contexts or after delays. By mechanistically modeling fear relapse, ConFER offers insights to improve exposure therapy outcomes.

Fig. 1. ConFER Model Overview.

a A simplified schematic of the circuit of fear in the brain. Cue information travels directly to the basolateral amygdala (BLA), activating either the fear or extinction engram based on cue associations. Context information travels to the hippocampus and then to the BLA via the Infralimbic Cortex to suppress fear or the Prelimbic cortex to activate fear. There is competition between the fear and extinction engrams in the BLA, following which the central amygdala (CeA) has downstream projections to initiate the appropriate behavioral response (b) Computational model architecture. Context and cue inputs are connected to stored engrams in the positive (green) and negative (red) stimulus-responsive BLA populations. Purple circles indicate engrams stored in each BLA population, with the numbers corresponding to engram indices. The filled yellow engram indicates activation because the corresponding US (Shock) is present during an experimental trial. The yellow square indicates a selection of a random subset of BLA-positive neurons to form an extinction engram. The scaled sum of activations of each BLA population is used to determine the final fear level, which is computed as a tanh function.

Fig. 2. Fear acquisition and extinction learning curves.

a, b Model predictions for fear acquisition learning across trials (in blue) and extinction learning across trials (in orange). a Acquisition and Extinction take place in the same context b Extinction takes place in a different context c Acquisition and Extinction curves from the Quirk fear conditioning study in rats. c reproduced with permission from Quirk. Memory for extinction of conditioned fear is long-lasting and persists following spontaneous recovery. Learning & Memory, 9(6), 402–407. Copyright © 2002 Cold Spring Harbor Laboratory Press.

Fig. 3. Fear renewal across different context combinations.

The charts on the left contain ConFER’s predictions, and the charts on the right reflect corresponding empirical findings from Bouton, Bouton and Bolles, and Bouton and Ricker. a, b Level of fear renewed when acquisition and extinction take place in different contexts A and B, respectively. Bars show the level of fear renewed in the acquisition context (ABA Renewal in red), a novel context (ABC Renewal in green), and in the extinction context (ABB Renewal in yellow). c, d Level of fear renewed when acquisition and extinction take place in the same context A. Bars show the level of fear renewed in a novel context (AAB Renewal in yellow) and in the acquisition/extinction context (AAA Renewal in red). Model mechanisms involved in fear acquisition in Context A (e), fear extinction in Context B (f), and fear renewal in Context A (ABA Renewal) (g), Context C (ABC Renewal) (h), and Context B (ABB Renewal) (i). Dotted associations are susceptible to temporal decay while solid associations are not. Thickness of arrows depicts strengths of connection weights. Red arrows show negative BLA associations, while green arrows show positive BLA associations. Grayed out arrows depict existing associations that are inactive for the presented cue-context pair. Fear acquisition involves strengthening cue (CS-1) and acquisition context (Context A, Living Room) associations with the negative US present (US-1) (e), whereas fear extinction only involves strengthening of the extinction context (Context B, Bedroom) association with the associated extinction engram (yellow rectangle) (f). ABA Renewal reactivates CS-1 and Acquisition Context associations with US-1 leading to a full return of fear (g). ABC Renewal works only through the CS-1 association, with neither the positive extinction context association nor the negative acquisition context association present, leading to a slightly decreased return of fear (h). ABB Renewal reactivates the extinction engram through the extinction context pathway which neutralizes the negative CS-1 association with US-1 (i).

Fig. 4. Spontaneous recovery.

a ConFER’s predictions for the amount of fear that returns after prolonged delays after extinction, with the delay measured across 21 simulated days. b Empirical findings on the return of fear following fear extinction in rats across 14 days from Quirk show a similar learning curve over time. Model mechanisms involved in fear acquisition in Context A (c), fear extinction in Context A itself (d), and exponentially increasing spontaneous recovery in Context A following a time delay t (e) and complete spontaneous recovery in Context A after a 21 day simulated delay within ConFER (f). Dotted associations are susceptible to temporal decay while solid associations are permanent. Thickness of arrows depicts strengths of connection weights. Red arrows show negative BLA associations, while green arrows show positive BLA associations. Fear extinction in the acquisition context involves the acquisition context forming new positive associations with the extinction engram to neutralize existing cue and context associations with US-1 (d). With a time delay t, context associations decay as indicated by the thinning dotted lines, while the negative cue association begins to dominate (e).

Fig. 5. Counterconditioning.

a Model predictions of levels of fear during Fear acquisition (in blue) and counterconditioning (in purple) across experimental trials. b Model predictions of levels of fear during Fear acquisition (in blue) and extinction (in orange) across experimental trials. c A comparison of the amount of fear that returns following counterconditioning (purple bars) and extinction (orange bars). The first 3 sets of bars represent the immediate return of fear via fear renewal in the acquisition, extinction, and novel contexts, respectively. The final pair of bars represents the return of fear in the extinction context after a 21-day delay via spontaneous recovery.