A theory-guided model maps four recurrent modes of fear-goal arbitration across learning and relapse in rats

Reviewer #1 (Public review):

Summary:

This study addresses an important question of how recovery from conditioned fear can be characterized beyond the reduction of defensive responses. The authors combine multiple behavioral measures with a theory-guided hidden Markov model to identify recurring behavioral modes and examine how the organization of behavior changes during fear conditioning and extinction. The study provides an interesting and potentially valuable framework for examining recovery from conditioned fear as a broader change in behavioral repertoire rather than simply as a reduction in defensive behavior.

Strengths:

A major strength of the study is the attempt to characterize fear-related behavior as a dynamic organization of multiple behavioral measures rather than relying on a single conventional measure such as freezing. The integration of freezing, ultrasonic vocalizations, and food-reinforced behavior into recurring behavioral modes provides a potentially useful approach for capturing behavioral states that may be missed by analyses of individual responses. The study also presents a theoretically motivated framework for distinguishing defensive, quiescent, exploratory, and goal-directed behavioral modes, which could provide a useful conceptual basis for studying functional recovery following aversive experiences.

Weaknesses:

The main concern is whether the analyses and behavioral definitions provide sufficient support for the interpretation of these behavioral modes and, in particular, for the proposed concept of functional recovery. First, the data-driven model selection indicates substantially better predictive performance for eight states, whereas the authors retain four states primarily to conform to their theoretical framework. A stronger justification for prioritizing the four-state solution would therefore strengthen the interpretation of the results. Second, it is difficult to determine from the current behavioral measures whether the Quiescent state reflects safety and rest or behavioral inhibition associated with residual fear or anxiety. The marked reduction in exploratory and goal-directed behavior could potentially be consistent with either interpretation, making it important to consider alternative explanations for this state.

The definition of functional recovery also warrants further consideration. The authors propose the reinitiation of food-seeking behavior in the previously threatening environment as an important indicator of recovery. From an evolutionary and ethological perspective, however, an animal that has experienced a threat while foraging in a particular location may benefit from shifting its foraging to an alternative location rather than returning to the previously threatening location, even after extinction learning. Thus, the resumption of goal-directed behavior in the previously threatening environment may not necessarily represent a biologically appropriate benchmark for successful fear extinction or functional recovery, as functional recovery may involve substantial changes in motivational processes that cannot be captured by the simple return of a previously suppressed behavior. More broadly, the results may support the value of examining behavioral repertoire during extinction, but they do not yet establish that restoration of the broader repertoire, or specifically the return of goal-directed behavior in the previously threatening environment, is required for functional recovery.

Finally, the use of the no-shock control group could be more informative. The observed reduction in exploratory behavior during conditioning cannot by itself distinguish associative fear learning from habituation to novelty. A direct comparison between shock and no-shock groups would therefore help determine whether the changes in behavioral modes are specifically related to conditioning.

Overall, the study provides a potentially useful framework and an interesting conceptual perspective on recovery from conditioned fear. However, the strength of the evidence supporting the interpretation of the identified behavioral modes and the proposed definition of functional recovery is currently limited. Addressing these interpretational and analytical issues would strengthen the manuscript and help establish the extent to which this framework provides a meaningful advance in the study of fear extinction and recovery.

Reviewer #2 (Public review):

Summary:

The authors aimed to develop and validate a preclinical computational framework for assessing functional recovery after aversive learning. They argue that a reduction in defensive behavior (e.g., freezing) is insufficient to conclude recovery, and that a comprehensive assessment must also track the re-emergence of exploration and goal-directed action. To test this, they adapted a multimodal paradigm in rats, measuring freezing, 22-kHz and 50-kHz ultrasonic vocalizations (USVs), and food-reinforced lever pressing. Using a theory-guided hidden Markov model (HMM) with four states derived from tri-reference-point theory (Fear/Danger, Quiescent, Exploratory, Goal-directed), they decoded behavioral trajectories across fear conditioning, extinction, delay, and contextual relapse experiments.

Strengths:

(1) The study tackles a significant gap in animal models of fear and trauma-related disorders. It correctly identifies that a focus on a single defensive response like freezing may be misleading and that a broader behavioral assessment is necessary to understand the impact of interventions.

(2) The application of a hidden Markov model to combine multiple behavioral channels into a cohesive, theory-driven framework is a strength. This approach moves beyond simple summary statistics to capture the temporal dynamics of behavior and allows for the identification of distinct, recurring states.

(3) The authors demonstrate commendable rigor in their model selection and validation. They employ discrete-aware scoring to address the boundary issues of count data and conduct extensive stability analyses across restarts, variance floors, and studies. The association of the Fear/Danger mode occupancy with fecal boli, an independent autonomic measure, provides a convincing external validation.

(4) The study provides several important findings. For instance, it shows that during the first extinction session, although Fear/Danger occupancy declined, the dominant state was Quiescent, rather than Exploratory or Goal-directed. This directly supports the core premise that the absence of fear does not equal functional recovery.

Weaknesses:

(1) The authors transparently report that their results are not emission-family invariant. The failure of count-native models (e.g., negative binomial) to converge or produce the expected four-state solution is a major limitation. This suggests the theory-guided state structure is tied to the specific Gaussian representation chosen and may not be a robust and universal feature of the behavioral data. This limits the strength of the claim that the data naturally resolve into these four modes.

(2) The exclusive use of male rats from a single laboratory limits the generalizability of the findings to female animals and other experimental settings. The cross-study recurrence analysis is a helpful sensitivity check but does not substitute for true model transfer to an independent, out-of-sample dataset.

(3) Testing K=4 vs. K=8 raises the question of whether the four-state solution is the most appropriate level of abstraction for the questions being asked.

(4) The interpretation of the 50-kHz USV band as an Exploratory state is somewhat broad. 50-kHz calls are emitted in a variety of contexts, including social interaction and reward. While the use of the band is justified, it remains an indirect measure of exploration and should be interpreted with caution. The interpretation and link to the exploratory state should be toned down and discussed.

Author response:

We thank the editors and both reviewers for their careful evaluation and for recognizing the value of examining the wider behavioral repertoire after aversive learning. Our aim is to follow what an individual does after danger and intervention, rather than infer recovery from a decline in one defensive response. The reviews identify where the present computational account needs stronger connections to observed behavior, experimental comparisons, and subsequent outcomes. We are preparing a substantive revision addressing these issues. This provisional response outlines the work planned; it does not report completed revisions or new findings.

Response to Reviewer 1:

(1) Four-state versus eight-state resolution

We agree that the better held-out performance of K=8 must be central to the interpretation. K=4 is a theory-guided coarse descriptive scale, not the predictively optimal model or evidence that behavior contains exactly four natural states. We will explain this distinction more clearly and show what the additional states resolve, particularly within the low-event-output configurations. Comparisons of experimental change at both resolutions will help establish which conclusions depend on the coarse scale and which persist at finer resolution. We will not treat agreement within one emission representation as proof of a universal four-mode organization.

(2) Interpreting Quiescent

The reviewer's concern is important: Quiescent denotes low vocal and lever output, not an absence of freezing, a safe internal state, or successful recovery. Likewise, a decline in Fear/Danger occupancy does not establish that defensive responding has disappeared. We will present raw freezing, vocalizations, and lever pressing alongside state occupancy, with experimental groups distinguished. The higher-resolution subdivision of count-silent observations will make differences in freezing more visible. These comparisons can describe residual defensive behavior, but cannot by themselves determine whether quietness reflects rest, inhibition, anxiety, or another internal condition. We will retain these alternatives explicitly rather than infer safety from silence.

(3) The ethological meaning of functional recovery

We appreciate the distinction between returning to a previously rewarded action and adapting appropriately after danger. Resuming food-directed responding in the original setting is one task-specific outcome, not a necessary or sufficient definition of recovery. Avoiding that location or pursuing resources elsewhere could also be adaptive. The present task does not offer the alternatives needed to distinguish those possibilities. We will therefore frame the measured repertoire as an assay of context-dependent functional change, not a universal standard to which animals should return. The translational question remains whether an intervention changes the broader organization of behavior, but judging successful adaptation also requires evidence about the opportunities and risks available to the individual. Pooled state percentages will not be presented as normal or desirable recovery targets.

(4) Shock versus no-shock comparisons

We will report direct comparisons between shock and no-shock animals for the raw channels and inferred occupancies, including changes over conditioning. This will allow the reader to distinguish changes associated with aversive conditioning from changes shared with unshocked animals, such as familiarization or declining novelty. The revision will also separate extinction from time-matched exposure and present renewal by group rather than pooling treatment histories. Comparisons will use rats as the experimental units, report uncertainty, and make differences in observation coverage explicit.

(5) Connecting post-extinction behavior to later outcomes

We will examine whether the available paired records permit associations between post-extinction configurations and subsequent defensive responding, alongside the corresponding raw behavioral measures. These analyses will respect test chronology and intervening treatment. They will be identified as retrospective and exploratory, not as independent validation of individual prediction: the current shared representation was fitted using the full dataset. We will report limited sample size, missing modalities, or insufficient variation when these prevent a question from being estimated. A failure to establish an association will not be treated as evidence that no biological relationship exists. If these data cannot support a prognostic claim, we will state that limitation and identify the prospective test needed.

Response to Reviewer 2:

(1) Emission-family dependence

We agree that the count-native sensitivities materially limit the interpretation. Alternative emission models also retain temporal dependence; their disagreement with the Gaussian model therefore does not show that temporal modeling uniquely produces the four labels, or that the Gaussian account is biologically privileged. We will explain the count transformation and scaling, and distinguish numerical fit eligibility from recovery of the complete behavioral-label configuration. We will also make clearer which findings are conditional on the Gaussian representation. Stable classification within that representation is useful evidence for a reproducible reference, but is not evidence for an emission-family-invariant taxonomy.

(2) Generalizability and independent transfer

All animals were male and the experiments came from one laboratory. Recurrence across related study datasets, including a retest of the same animals, does not substitute for transfer to independent subjects or settings. We will strengthen the prospective validation roadmap: fix preprocessing and model parameters before testing independent animals, include both sexes and an independent setting, measure individual baselines, and evaluate later outcomes without using those outcomes to fit the representation. Direct exploration measures and tests of instrumental outcome sensitivity would provide independent checks on functional interpretation. This is a proposed validation design, not generalization demonstrated by the present study.

(3) The appropriate level of abstraction

The revision will distinguish predictive resolution from interpretive scale throughout, as described in our response to Reviewer 1. K=8's advantage will remain explicit. The relevant question for K=4 is whether it offers an informative coarse summary of experimental change, not whether it wins model selection. Conclusions that require a finer distinction will be presented at that resolution rather than forced into four categories.

(4) Vocal and instrumental labels

We agree that 50-kHz calls are a context-sensitive vocal measure, not a direct observation of exploration or a unique indicator of positive affect. We will qualify the Exploratory label consistently by its measured emission pattern. Similarly, food-reinforced lever pressing establishes instrumental responding in this task; without outcome-sensitivity tests it does not establish goal-directed rather than habitual control in the technical sense. The labels will describe observed behavioral configurations, not subjective feelings, neural mechanisms, or independently validated motivational states. A simplified behavior-to-model mapping will make these distinctions easier to follow.

Together, these revisions will place the raw behavior and experimental contrasts alongside the model, clarify the limits of the labels and recovery interpretation, and make the next validation steps concrete. The intended contribution is a reproducible framework for describing history-dependent changes in the measured repertoire and testing their functional relevance, without equating a model category with recovery or claiming individualized prognosis that the present evidence has not established.