Distributed neural representations of conditioned threat in the human brain

The study addresses how the human brain represents conditioned threat and safety across acquisition, extinction learning, and extinction memory recall. Traditional work has emphasized a core 'threat circuit' (amygdala, hippocampus, insula, medial prefrontal regions, periaqueductal gray), linking localized activations to defensive responses and learning. However, this circuit alone may not capture the complex sensory and cognitive processes engaged during threat encounters. The authors hypothesize that distributed neural systems beyond the classical circuit contribute substantially to encoding threat vs. safety and that multivariate pattern analysis (MVPA) can yield sensitive, generalizable decoders. They aim to build and validate robust neural decoders across large samples and diverse paradigms, integrating threat-circuit, sensory, and cognitive nodes to better characterize neural representations of conditioned threat and safety.

Prior research in Pavlovian threat conditioning has implicated the amygdala, hippocampus, insula, and medial prefrontal cortex in acquisition, expression, extinction, and contextual processing of conditioned threat. Meta-analyses and cross-species work support specific associations between subregions (e.g., basolateral/centromedial amygdala, anterior cingulate, vmPFC) and behavioral expressions during learning and extinction. Recent frameworks posit broader involvement of distributed systems (sensory processing, attention, cognitive control, default mode) in threat processing. MVPA studies have begun probing fear-related representations but often with small samples and single paradigms, raising concerns about generalizability. The present work extends these findings using large-scale, multi-dataset MVPA to test distributed representations and specificity to conditioned threat relative to other negative affect paradigms.

Participants: Data from 1465 participants across multiple studies. Discovery dataset: n=425 (Milad two-day conditioning/extinction paradigm). External validation datasets 1–2: n=98 and n=127 using the same paradigm. Additional external datasets (3–9) included variant visual/auditory conditioning, subjective fear, negative affect, pain, and social rejection paradigms. Experimental paradigm (discovery and validation 1–2): Two-day fear conditioning/extinction. Day 1: Conditioning with three colored CS (two CS+ partially reinforced at 62.5%, one CS− unreinforced), followed by extinction (CS+ and CS− without shock). Day 2: Extinction memory recall with extinguished CS+ (CS+E), unextinguished CS+ (CS+U), and CS−. Each phase had 32 trials; trials were grouped into four trial-blocks (TB1–TB4; 8 trials/TB). Contexts (office/library) were counterbalanced; CS colors counterbalanced. MRI acquisition and preprocessing: Multiple 3T scanners across sites; preprocessing with fMRIPrep 20.0.2. Structural T1w: bias correction, skull stripping, tissue segmentation, nonlinear normalization to MNI152NLin2009cAsym. Functional: motion correction (FSL mcflirt), slice-timing correction (AFNI 3dTshift), BBR coregistration (FSL flirt), normalization to MNI, resampling to 2 mm isotropic, 6-mm FWHM Gaussian smoothing. Activation estimation: Univariate GLM (Nistats). Trials per CS type split into trial-blocks; boxcar regressors convolved with canonical HRF. Nuisance regressors: 6 motion parameters, high-pass filter (128 s), volume censoring (framewise displacement >0.9 mm), polynomial drift; first-order AR model for noise. For each TB, produced activation maps for CS+ and CS− (and CS+E/CS+U during recall). Analytic approaches: 1) Threat-circuit decoding: Features restricted to masks for BLA, CMA, anterior/posterior hippocampus (aHPC/pHPC), insular subregions (dAI, vAI, PI), dACC, sgACC, vmPFC. Masks derived from literature-based atlases/meta-analytic peaks (e.g., amygdala subnuclei masks; Harvard-Oxford; insula parcellation; Neurosynth spheres). 2) Whole-brain decoding: Features from gray matter excluding the threat-circuit voxels and a 3-voxel-radius neighborhood around them; also examined all gray matter as a completeness check. Classifier and validation: Logistic regression with L2 regularization (scikit-learn). Hyperparameter C chosen from 20 values (0.01–100) based on training folds only. Performance assessed via forced-choice accuracy. Discovery dataset: 5-fold cross-validation repeated 10 times with different splits; mean accuracy reported. External generalization: train on all discovery data, test directly on validation datasets 1–2. Statistical significance via permutation tests (1,000 label shuffles) for CV; binomial tests for some external datasets. Voxel contributions: Model weights transformed to predictive patterns using Haufe transformation; bootstrap averaging (10 resamples). For whole-brain analyses, permutation-derived voxelwise p values (FDR q<0.05, two-sided) identified significant contributors. Extended circuit construction: From whole-brain analyses, identified 14 representative regions (bilateral) contributing across robustly decoding trial-blocks: AG, OFC, SMA, S1, M1, VIS/occipital pole, CER/Crus I, TH, IFG opercularis (Opr), IFG triangularis (Tri), MFG, caudate (Cd), SFG, PCC. Combined with 10 threat-circuit regions to form an extended “threat detection and flexible responding circuit.” Representational similarity analysis (RSA) across regions defined communities with consistent CS+ coding, consistent CS− coding, or flexible coding across phases. External validations (datasets 3–9): Applied conditioning-phase decoders (TB1–TB4) trained on the extended circuit to: multiple visual conditioning datasets (n=299; n=94; n=48), an auditory conditioning dataset (n=68), subjective fear (n=65), picture-induced negative affect (n=182), and pain/social rejection (n=59). Compared accuracies across paradigm categories (conditioned threat; subjective fear; intrinsically salient stimuli) via chi-square tests with FDR correction.

- Threat-circuit decoding (discovery dataset, n=425): Accuracies for CS+ vs. CS− during conditioning were significant across all TBs (TB1–TB4: 70.5%, 74.0%, 65.5%, 68.7%; all p<0.001). During extinction, TB1 was significant (62.4%, p<0.001), with later TBs near chance (TB2 53.5%, p=0.077; TB3 52.9%, p=0.12; TB4 51.2%, p=0.26). During recall, CS+E vs. CS− was 68.0% (p<0.001) and CS+U vs. CS− was 67.4% (p<0.001). Comparable results held when excluding shock-related trials and controlling for autocorrelation concerns. - Generalization (threat-circuit) to validation datasets 1 (n=98) and 2 (n=127): Successful external validation across corresponding phases/trial-blocks. - Whole-brain decoding (excluding threat circuit) showed numerically superior performance: Conditioning TB1–TB4 accuracies: 88.6%, 85.7%, 81.7%, 77.5% (all p<0.001); Extinction TB1: 80.2% (p<0.001); Recall TB1: CS+E vs. CS− 74.6% (p<0.001); CS+U vs. CS− 72.6% (p<0.001). Generalized well to validation datasets 1–2; superiority remained when voxel counts were matched. - Distributed voxel contributions: Significant contributors spanned somatomotor, ventral attention, control, and default-mode networks. Fourteen representative regions identified beyond the classical circuit: AG, OFC, SMA, S1, M1, VIS/occipital pole, CER/Crus I, TH, IFG opercularis, IFG triangularis, MFG, caudate, SFG, PCC. - Extended circuit communities (via RSA): Consistent CS+ coders: dACC, dAI, PCC, IFG opercularis, caudate, thalamus. Consistent CS− coders: PI, aHPC, pHPC, OFC. Flexible coders (phase-dependent CS+/CS−): BLA, CMA, vAI, vmPFC, sgACC, AG, S1, M1, SMA, MFG, VIS, IFG triangularis, cerebellum. - External datasets (extended-circuit decoders, conditioning-phase TBs): • Visual conditioning dataset 3 (n=299): Accuracies TB1–TB4 = 78.3%, 91.0%, 88.6%, 79.6% (all p<0.001). • Visual conditioning dataset 4 (n=94): 69.1%, 90.9%, 89.4%, 88.3% (all p<0.001). • Visual conditioning dataset 5 (n=48): 79.2%, 87.5%, 85.4%, 83.3% (all p<0.001). • Auditory conditioning dataset 6 (n=68): 51.4% (p=0.90), 83.8% (p<0.001), 67.6% (p=0.005), 61.7% (p=0.068). • Subjective fear dataset 7 (n=65): 76.9%, 81.5%, 73.8%, 80.0% (all p<0.001). • Negative affect dataset 8 (n=182): 75.3% (p<0.001), 70.3% (p<0.001), 72.0% (p<0.001), 50.5% (p=0.94). • Pain dataset 9 (n=59), physical pain: 49.2% (p≈1.0), 71.2% (p=0.002), 50.8% (p≈1.0), 57.6% (p=0.30); social rejection: 50.8% (p≈1.0), 61.0% (p=0.12), 61.0% (p=0.12), 57.6% (p=0.30). - Cross-paradigm comparisons: Decoding accuracies for conditioned threat (datasets 3–6) were not significantly different from subjective fear (dataset 7) across TBs, and were significantly higher than intrinsically salient stimuli (datasets 8–9) at TB2–TB4 (chi-square tests, FDR-corrected). This indicates partially overlapping but distinct neural representations for conditioned threat vs. general negative affect. - Spatial patterns within classical regions showed stable CS+ coding (e.g., dACC, dAI) and stable CS− coding (e.g., posterior insula, hippocampus), alongside dynamic coding in amygdala subnuclei and prefrontal/sensory regions across phases.

The findings show that both the classical 'threat circuit' and broader distributed neural systems encode threat vs. safety in a manner that is decodable and generalizes across datasets, scanners, and paradigms. Whole-brain patterns excluding the classical circuit yield higher accuracies, supporting frameworks that threat processing relies on integrated sensory, attentional, memory, and control networks. The extended circuit delineates communities with consistent CS+ or CS− preference, and a set of flexible coders whose phase-dependent shifts may reflect rapid associative learning, extinction, habituation, and adaptive control. Despite paradigm diversity, classifiers trained on one paradigm generalized well to other visual conditioning tasks and to subjective fear, but less so to intrinsically salient negative affect, physical pain, and social rejection—suggesting both shared and distinct coding components. The work advances the construction of sensitive, reproducible neural decoders of conditioned threat and highlights the value of considering multiple large-scale brain systems in threat-related cognition and behavior.

This study establishes robust, generalizable MVPA decoders for conditioned threat and safety that integrate traditional threat circuitry with distributed sensory and cognitive nodes. The extended ‘threat detection and flexible responding circuit’ captures both stable valence coding and flexible, phase-dependent representations supporting adaptive behavior during conditioning, extinction, and recall. The decoders generalize across datasets, scanners, and variant paradigms, and show specificity relative to intrinsically salient negative affect and pain. Future work should dissect paradigm-specific vs. general threat features, examine finer temporal dynamics and plasticity of flexible nodes, and test causal roles of circuit nodes (e.g., via neuromodulation) by measuring changes in decoding and behavior.

- Some threat-circuit masks (vmPFC, sgACC, dACC) were defined as small spheres from meta-analytic peaks and may not fully cover these cortices; voxels outside these masks could have contributed to whole-brain decoding. Sensitivity analyses with larger masks indicated minimal impact on whole-brain results. - The schematic coding-community summary reflects patterns specific to the present design and trial-block structure; coding dynamics may vary with other stimuli, reinforcement schedules, or contexts. - Classifiers trained on visual paradigms showed lower performance on auditory conditioning, indicating some paradigm-specific features in the learned representations. - Application to intrinsically salient stimuli (negative affect, pain, social rejection) yielded lower accuracies, suggesting only partial overlap with conditioned threat representations. - The observational fMRI design precludes causal inference; no direct neural manipulations were performed. Spatial resolution limitations excluded some relevant regions (e.g., PAG, BNST).