Pattern differentiation and tuning shift in human sensory cortex underlie long-term threat memory

Summary

The amygdala-prefrontal-cortex circuit has long occupied the center of the threat system ¹, but new evidence has rapidly amassed to implicate threat processing outside this canonical circuit ^2-4. Through non-human research, the sensory cortex has emerged as a critical substrate for long-term threat memory ^5-9, underpinned by sensory cortical pattern separation/completion ^10,11 and tuning shift ^12,13. In humans, research has begun to associate the human sensory cortex with long-term threat memory ^14,15, but the lack of mechanistic insights obscures a direct linkage. Towards that end, we assessed human olfactory threat conditioning and long-term (9-day) threat memory, combining affective appraisal, olfactory psychophysics, and functional magnetic resonance imaging (fMRI) over a linear odor-morphing continuum (five levels of binary mixtures of the conditioned stimuli/CS+ and CS− odors). Affective ratings and olfactory perceptual discrimination confirmed (explicit) affective and perceptual learning and memory via conditioning. fMRI representational similarity analysis (RSA) and voxel-based tuning analysis further revealed associative plasticity in the human olfactory (piriform) cortex, including immediate and lasting pattern differentiation between CS and neighboring non-CS and late-onset, lasting tuning shift towards the CS. The two plastic processes were especially salient and lasting in anxious individuals, among whom they were further correlated. These findings thus support an evolutionarily conserved, sensory cortical system of long-term threat representation, which can underpin threat perception and memory. Importantly, hyperfunctioning of this sensory mnemonic system of threat in anxiety further implicates a hitherto underappreciated sensory mechanism of anxiety.

Blurb

Rapidly accruing evidence questions amygdala’s dominance in (human) threat processing. Favoring a distributed threat circuitry, You et al. identify long-term threat memory in human sensory (olfactory) cortex (but not amygdala or orbitofrontal cortex). Notably, this sensory cortical memory system hyperfunctions in anxiety.

RESULTS

Behavioral effects

As in previous non-human ^12,16,17 and human ¹⁸ research, we employed a linear morphing continuum of odor mixtures, with the two extreme odor mixtures (i.e., threat CS/CSt and safety CS/CSs) differentially paired with bimodal unconditioned stimuli (i.e., aversive or neutral sounds and images; Figure 1A-C). We examined threat learning and long-term (9-day) threat memory based on affective appraisal and perceptual discrimination (in an odor discrimination task/ODT) of the CS.

Figure 1. Odor stimuli and experimental design.

(A) Stimuli consisted of a continuum of five parametrically-morphed binary odor mixtures of neutral odors (acetophenone and eugenol labeled as Odor A and Odor B). The extreme mixtures (20% A/80% B and 80% A/20% B; to rule out confounds related to pure or mixture odorants, all stimuli consisted of binary mixtures) were differentially conditioned as CSt (threat) or CSs (safety) via paired presentation with aversive or neutral unconditioned stimuli (UCS: bimodal aversive or neutral pictures and sounds). SCR evoked by the aversive (vs. neutral) UCS confirmed their effectiveness (Aversive vs. Neutral: t = 2.88; P = .007). Assignment of CSt/CSs was counterbalanced across participants. The three intermediate mixtures (35% A/65% B, 50% A/50% B, and 65%A/35% B) were non-conditioned stimuli (nCS), representing the odor neighboring CSt (nCSt), the midpoint mixture (nCSm), and the odor neighboring CSs (nCSs).

(B) Two-alternative-forced-choice (2-AFC) odor discrimination task (ODT) accompanied by fMRI and respiration acquisition. Each trial presented an odor mixture pseudo-randomly for 1.8 seconds, to which participants made judgments of “Odor A” or “Odor B” with button pressing.

(C) Experiment schedule. Day 1 consisted of pre-conditioning 2-AFC ODT, conditioning, post-conditioning 2-AFC ODT, and odor risk rating. Day 9 consisted of post-conditioning 2-AFC ODT, odor risk rating, and an olfactory localizer scan.

(D) Regions of interest (ROIs). Anatomical masks of the primary olfactory cortex (anterior piriform cortex/APC and posterior piriform cortex/PPC), the olfactory orbitofrontal cortex (OFC_olf), and the amygdala (AMG) are displayed on 3D T1 sections of one participant. These ROIs were further functionally constrained by the olfactory localizer.

See also Figure S1.

Affective appraisal

Pre-experiment odor valence ratings (on a VAS of 0-100) indicated neutral affective values for the five odors, conforming to a flat neutral baseline over the odor continuum [P = 0.416; Mean (SD) = 50.6 (19.7)]. Risk ratings of the odors (i.e., the likelihood of aversive UCS following a given odor) were acquired post-conditioning on Day 1 and Day 9. Consistent with our hypothesis (Figure 2a), an ANOVA of Odor (five odors) and Time (Day 1/Day 9) demonstrated a strong ascending linear trend over the odor continuum (F_1,30 = 6.99, P = 0.013; Figure 2C). There was no Odor-by-Time interaction (P = 0.908), suggesting equivalent trends for Day 1 and Day 9. Akin to the linear trend, CSt and CSs had maximal and minimal risk ratings, respectively (CSt vs. CSs: t₃₁ = 3.02, P = 0.005), deviating from the neutral level (50%) in opposite directions (CSt: t₃₁ = 2.67, P = 0.012; CSs: t₃₁ = −2.40, P = 0.023). Ratings for the three nCS odors remained neutral on both days (49.3-51.3%; all P values > 0.581) and comparable to each other (F_1,31 = 0.14, P = 0.708) but differed from CSt and CSs in opposing directions (nCSt vs. CSt: t₃₁ = −2.41, P = 0.022; nCSs vs. CSs: t₃₁ = 1.92, P = 0.032 one-tailed). These results thus confirmed that our differential conditioning produced a threat and a safety CS that persisted till Day 9, with limited generalization to the non-CS.

Figure 2. Behavioral effects of olfactory conditioning.

(A) Hypothetical affective space over the odor continuum: the initial neutral baseline (gray line) would change to an ascending safety-to-threat line (black line) after acquiring affect (safety/threat) through differential conditioning. Inset shows changes (Δ) in odor affect over the continuum via conditioning, which conforms to a linear trend.

(B) Hypothetical perceptual (quality) space over the odor continuum: the initial ascending trend (gray line; tracking the linear increase in the proportion of CSt) would be warped after conditioning due to expanded distances (i.e., enhanced perceptual discrimination) between the CS (CSt/CSs) and neighboring nCS (nCSt/nCSs) (black line). Inset illustrates the changes (Post – Pre) in perceived odor quality (solid line), which can be fitted by a cubic trend, anchored by respective increase and decrease in “CSt” rate for nCSt and nCSs (dotted line).

(C) Empirical risk ratings (likelihood of aversive UCS) on both days conformed to the predicted profile of differential conditioning: below-chance risk for CSs and above-chance risk for CSt. Risks for the three intermediate mixtures remained chance-level (50%; indicated by the dotted line).

(D) Empirical 2-AFC ODT performance (“CSt” responses rate) over the CSs-to-CSt continuum conformed to a linear trend before conditioning, which was warped after conditioning. Inset illustrates differential “CSt” rates (Post - Pre) over the odor continuum on Day 1 and Day 9, which largely conformed to the hypothesized cubic trend. Specifically, perceptual distances between the CS and neighboring nCS increased after conditioning, with the nCSt odor less endorsed as “CSt” and the nCSs odor more endorsed as “CSt” (i.e., less as “CSs”). Error bars represent s.e.e. (individually adjusted s.e.m.). *: P < .05.

Perceptual discrimination

Akin to the linear odor morphing continuum, baseline ODT performance conformed to a strong linear trend of increasing endorsement of the dominant odor of the CSt (i.e., “CSt” rate), F_1,31=79.62, P < 0.0001 (Figure 2B & D). We hypothesized that differential conditioning would expand perceptual distances (i.e., enhance perceptual discrimination) between the CS and their neighboring nCS. Across the odor continuum, this expansion would manifest as a cubic pattern of differential “CSt” rates (Post- minus Pre-conditioning), anchored by respective increase and decrease (from Pre- to Post-conditioning) in “CSt” rate for the neighboring odors of CSt and CSs—nCSt and nCSs (Figure 2B inset). Indeed, an ANOVA of Odor (five odors) and Time (Day 1/Day 9) on differential “CSt” rates confirmed this cubic trend (F_1,31 = 3.16, P = 0.043 one-tailed; Figure 2D inset). Like risk ratings, there was no Odor-by-Time interaction (P = 0.405), suggesting equivalent changes for Day 1 and Day 9. The expansion between the CS and neighboring nCS was further ascertained in a follow-up ANOVA (Odor: nCSt/nCSs by Time: Day 1/Day 9 on differential “CSt” rates). We observed an Odor effect (F_1,31 = 5.19, P = 0.030) and specifically, a negative differential “CSt” rate for the nCSt odor (i.e., less CSt endorsement and greater perceptual distance from CSt post-conditioning) and a positive differential “CSt” rate for the nCSs odor (i.e., more CSt endorsement and greater perceptual distance from CSs post-conditioning; Figure 2D inset). Again, this ANOVA showed no Odor-by-Time interaction (P = 0.427), suggesting equivalent effects for both days. Therefore, differential conditioning warped odor quality space, particularly expanding perceptual distances between the CS and neighboring nCS. Overall, results in affective appraisal and perceptual discrimination converged to confirm threat learning and long-term memory in the participants.

Neural effects

Non-human research has evinced plasticity associated with conditioning in the sensory cortex that arises immediately and lasts for days to weeks ^19-21, serving a critical role in the formation ^5,22-24 and storage of long-term memory of conditioning ^22,25-28. In humans, threat conditioning also induces immediate sensory cortical plasticity ^18,29-34. Recently, a link between human sensory cortex and long-term memory of conditioning has begun to emerge, indicated by enduring (15-day-long) plasticity in human visual cortex (i.e., enhanced V1/V2 response to CS)¹⁴ and impairment in delayed (24-hour) conditioned response following inhibitory stimulation of human somatosensory cortex¹⁵.

Non-human research has revealed pattern separation/completion in the sensory (particularly, olfactory) cortex to underpin memory of conditioning ^10,11. The olfactory primary (piriform) cortex is considered as an associative, content-addressable memory system and thus ideally positioned to store long-term memory of conditioning ^8,35,36. Non-human research has further implicated “associative representational plasticity” as a mechanism of long-term memory and sharpened perception of the CS ¹². This plasticity is characterized by sensory cortical tuning shift, i.e., sensory cortical neurons initially tuned to non-CS become preferentially tuned to the CS ^9,13,19,20. Critically, this plasticity would consolidate over time and last for a long time, thereby underpinning stable sensory representation and long-term memory of CS ^7,12,13,37.

We thus interrogated whether threat conditioning would induce long-term pattern differentiation and tuning shift in the human olfactory cortex (anterior and posterior piriform cortices; APC/PPC) using fMRI representational similarity analysis (RSA)²⁹ and voxel-based tuning analysis ^38,39, respectively. To compare the sensory cortex with the canonical amygdala-prefrontal-cortex circuit, supplemental analyses further explored these processes in the amygdala and orbitofrontal cortex (OFC; Figure 1D). Importantly, given that threat learning and memory represents an eminent model of anxiety disorders ^40-42, we examined individual differences in sensory cortical associative plasticity as a function of anxiety, thereby identifying a sensory cortical underpinning of anxiety.

Plasticity in the olfactory cortex

Pattern differentiation

Immediate pattern differentiation for CS has been observed in the human piriform cortex²⁹, which, we hypothesized, would persist to support long-term threat memory. RSA was applied to extract a Pattern Differentiation Index (PDI), reflecting pattern dissimilarity (i.e., pattern differentiation) between the CS and its neighboring nCS (Figure 3). ANOVAs of ROI (APC/PPC) and Time (Day 9/Day 1) on the differential PDI (Post- minus Pre-conditioning) showed a main effect of Time, F_1,30 = 4.30, P = 0.047. This Time effect reflected significant PDI increase in the piriform (APC/PPC) cortex (t₃₀ = 1.91, P = .033 one-tailed) on Day 1, in contrast to no PDI increase on Day 9 (all P values > 0.366). Interestingly, PDI increase in the piriform cortex on Day 9 correlated positively with anxiety (r = 0.40, P = 0.025, FDR P < 0.05; Figure 3C), indicating that olfactory cortical pattern differentiation persisted among anxious individuals. There was no effect of ROI (P = 0.296) or interaction (P = 0.271): as illustrated in Figure 3A & B, APC and PPC showed similar, significant PDI increase on Day 1 and similar, significant correlation between PDI increase on Day 9 and anxiety. Finally, these results were confirmed by non-parametric tests and leave-one-out cross-validation (Supplemental Information; Table S1).

Figure 3. Olfactory cortical pattern differentiation between CS and neighboring nCS odors.

(A) Group-average representational dissimilarity matrices (RDMs) for APC, PPC, OFColf and amygdala (AMG) at each phase. Each cell of the matrix indicates pattern dissimilarity (1-r), reflecting pattern differentiation, for a given odor pair. Cells right off the diagonal indicate pattern differentiation between neighboring odors: CSs and nCSs (d1), nCSs and nCSm (d2), nCSm and nCSt (d3), and nCSt and CSt (d4). Based on that, we derived a Pattern Differentiation Index (PDI) for the CS and the neighboring nCS [PDI = (d1 + d4) – (d2 + d3)].

(B) PDI for each ROI at pre-, Day 1, and Day 9 post-conditioning. Both APC and PPC (but neither amygdala nor OFC) demonstrated increased PDI from pre- to post-conditioning on Day 1, but not on Day 9. Center red line = group mean; red and blue boxes = 95% confidence interval and mean ± 1 SD, respectively.

(C) Correlations between conditioning-induced PDI changes and anxiety. PDI changes on Day 9 (vs. Pre) in the APC and PPC correlated positively with anxiety, indicating persistent pattern differentiation in anxious individuals. *: P < 0.05; +: P < 0.1.

See also Figure S2.

Tuning shift

We then examined tuning shift towards CS in the piriform cortex, i.e., whether voxels initially responded maximally to neighboring odors of the CS (i.e., nCSt and nCSs) became maximally responsive to the CS after conditioning ^12,43. Before conditioning, tuning was evenly distributed across the morphing continuum in the piriform cortex (and the supplemental regions—amygdala and OFC); i.e., equivalent % of voxels tuned to the five odors (all F values < 1.88, P values > 0.125). An ANOVA of ROI (APC/PPC) and Time (Day 9/Day 1) on Tuning Shift Index (TSI; % of nCS voxels towards the neighboring CS vs. the neighboring nCS) revealed a significant ROI-by-Time interaction (F_1,30 = 6.99, P = 0.013) and no main effect of ROI (P = 0.379) or Time (P = 0.612). Specifically, the interaction reflected significant TSI in the PPC on Day 9 (t₃₀ = 3.00, P = 0.005; FDR P < 0.05) but not on Day 1 (P = 0.802) or in the APC on either day (all P values > 0.269; Figure 4A & B). Interestingly, PPC TSI on Day 9 positively correlated with anxiety, suggesting that anxiety potentiated this delayed PPC tuning shift (r = 0.44, P = 0.014; FDR P = 0.056; Figure 4C). Finally, these results were confirmed by non-parametric tests and leave-one-out cross-validation (Supplemental Information; Table S2).

Figure 4. Olfactory cortical tuning shift towards the CS.

(A) Day 1 (dashed lines) and Day 9 (solid lines) post-conditioning tuning profiles of nCSs (green) and nCSt (pink) voxels (i.e., respectively tuned to nCSs and nCSt at the baseline). In PPC on Day 9, the nCS voxels exhibited a strong tuning preference for their respective CS: highest % of nCSs voxels tuned to CSs (shaded in green) and highest % of nCSt voxels tuned CSt (shaded in pink).

(B) Tuning shift index (TSI; % of nCS voxels towards respective CS vs. the middle nCS) on Day 1 and Day 9 post-conditioning. On Day 9, PPC showed significant TSI for both nCSs and nCSt voxels towards their respective CS (CSs and CSt, respectively). The dotted line indicates zero tuning shift (TSI = 0). Center red line = group mean; red and blue boxes = 95% confidence interval and mean ± 1 SD, respectively.

(C) Correlations between anxiety and tuning shift towards CS (collapsed across nCSs and nCSt). Day 9 TSI in the PPC correlated positively with anxiety, indicating amplified tuning shift in anxious individuals. The inset: in anxious (red dots) but not non-anxious (blue dots) participants, Day 9 PPC TSI correlated with Day 9 PPC PDI increase (more details in Supplemental Information, Figure S2). *P < 0.05; **P < 0.01.

See also Figure S2.

Association between pattern differentiation and tuning shift via conditioning

Results above showed that pattern differentiation and tuning shift were both present on Day 9 among anxious participants. We thus explored the inherent association between these plastic processes and found that PDI change and TSI in the PPC on Day 9 were correlated in the high-anxiety group (based on median split; r = .45, P = .039 one-tailed; Figure 4C inset; Figure S2). By contrast, there was no correlation (r = .04, ns) in the low-anxiety group. These results confirm the shared origin of these plastic processes in threat conditioning and anxiety.

Plasticity in the amygdala and OFC

We then explored pattern differentiation and tuning shift in the amygdala and OFC. As for pattern differentiation, the amygdala showed a marginal increase in PDI on Day 1 (t₃₀ = 1.48, P = 0.075 one-tailed) but not on Day 9 (P = 0.413; Figure 3a & b). The PDI scores on neither day were correlated with anxiety (all P values > 0.252; Figure 3c). The OFC showed no PDI increase nor correlations of PDI increase with anxiety (all P values > 0.292). As for tuning shift, TSI in the amygdala and OFC showed no significant tuning shift on either day (All P values > 0.169) nor correlation with anxiety (All P values > 0.144; Figure 4). Therefore, in contrast to the olfactory cortex, the canonical amygdala-OFC circuit failed to exhibit clear pattern differentiation or tuning shift following conditioning.

DISCUSSION

We demonstrated affective and perceptual learning and long-term memory, accompanied by immediate and lasting pattern differentiation as well as late-onset, lasting tuning shift in the human olfactory cortex, especially among anxious individuals. These findings highlight the role of the human sensory cortex in threat memory and advance the extant (human and non-human) literature of anxiety modulation of the sensory cortical system of threat memory. Together, they illuminate hitherto underexplored human sensory mechanisms of threat processing and their contribution to the pathophysiology of anxiety.

Differential conditioning is known to promote divergent conditioned responses to (threat and safety) CS, minimizing conditioning generalization and facilitating CS discrimination (especially from similar stimuli) ^17,29,44,45. Our risk ratings over a parametrically-morphed odor continuum confirmed differential affective learning and memory for CSt and CSs and minimal generalization to the nCS. Our odor discrimination task further demonstrated perceptual (discrimination) learning and memory for CS (vs. neighboring nCS). Together, differential conditioning warped both affective and perceptual spaces over the continuum, expanding the distance between the CS and their neighboring nCS and compressing the distance among the nCS. Such paralleled reorganization of affective and perceptual spaces reiterates that acquisition and generalization/specification of threat response tracks the perceptual distance between the CS and nCS ^46-48.

Neurally, the olfactory (APC and PPC) cortex exhibited immediate and lasting pattern differentiation between the CS and neighboring nCS, especially in anxious individuals. This pattern differentiation resembles conditioning-induced pattern separation in non-human sensory cortex, underpinning CS memory representation ^10,11 as a “perceptual-mnemonic” mechanism ⁴⁹. It is unclear whether this process directly relates to with hippocampal pattern separation characterized by sparse orthogonalized representation ^50,51. The human PPC is a critical site for olfactory sensory representation and underpins odor object encoding ^52,53. The immediate effect in PPC replicates our previous finding, reflecting adapted sensory representation of CS ²⁹. The lasting effect in PPC suggests that this plasticity would persist to support enduring sensory cortical representation of CS as part of the long-term memory of acquired threat/safety. The APC exhibited a comparable effect, replicating decorrelation of APC responses to CS and similar nCS in rodents ^10,17. Given human APC’s role in olfactory attention and arousal⁵³, this APC pattern differentiation could reflect heightened sensory vigilance to CS.

The other mnemonic mechanism—tuning shift that underpins associative representational plasticity—is also confirmed here in humans, particularly in the PPC. Interestingly, this plastic process was observed on Day 9 only. This temporal profile accords with non-human findings: sensory cortical tuning shift is relatively weak in magnitude and specificity immediately after conditioning but becomes stronger over time (days and weeks)¹⁹. It also coincides with our recent finding of delayed (Day 16 but not immediate) plasticity in human primary visual cortex (V1/V2)¹⁴. Therefore, this tuning shift, particularly in the PPC that is critical for olfactory sensory representation, underscores time-dependent associative representational plasticity in human olfactory cortex to support enduring CS representation as part of long-term memory of acquired threat/safety ^7,19.

In comparison, the amygdala and OFC exhibited no clear evidence of pattern differentiation or tuning shift by conditioning. The null findings here highlight the human sensory cortex (outside the canonical threat circuit) as an independent neural substrate for threat memory. That said, we analyzed these processes expressly along a physical dimension (i.e., odor-morphing continuum) to elucidate neural representation of CS sensory input, which does not rule out amygdala/OFC associative plasticity in other, abstract dimensions (e.g., valence or value). In fact, previous research comparing (immediate, appetitive) conditioning effects in the rodent piriform cortex and OFC has revealed sensory-based plasticity in the former and value/rule-based plasticity in the latter (e.g., ⁵⁴). Similarly, human research of (both appetitive and aversive) conditioning has underscored value-based (vs. sensory-based) pattern differentiation in the OFC and amygdala ^29,55,56. In sum, the contrast here highlights a sensory-bound representation system of threat memory (“S-memory”;^57,58) in the sensory cortex.

Finally, leveraging self-report from human participants, we demonstrated that anxiety amplified these threat mnemonic processes in the sensory cortex. Similar to our recent finding in the visual cortex¹⁴, anxiety particularly heightened piriform plasticity on Day 9. This anxiety effect helps to reconcile the seeming temporal dissociation between group-level (average) effects of pattern differentiation (present on Day 1) and tuning shift (present on Day 9). That is, in anxious individuals, the two forms of plasticity were both present on Day 9 and, moreover, correlated with each other, highlighting their shared origin in threat conditioning and anxiety. We caution that our sample could be small for a study of individual differences, warranting replication through future large-scale studies. Nonetheless, findings of sensory-based long-term threat memory in anxiety lend direct credence to anxiety theories centered on hyperactive sensory memory of threat^57,58. They also confer mechanistic insights into intrusive memories (a hallmark symptom) in posttraumatic stress disorder (PTSD) laden with vivid sensory fragments of trauma and readily triggered by simple sensory cues ^59-61.

To conclude, lasting pattern differentiation and tuning shift in the human PPC, paralleling long-term threat memory, provides mechanistic evidence for the human sensory cortex as a key component of the threat circuitry. This long-term threat representation may serve to underpin threat processing in the sensory cortex, even in the initial feedforward sweep^62-64. Importantly, that this sensory cortical mnemonic system of threat is hyperfunctioning in anxiety adds to the growing support for a sensory mechanism—exaggerated sensory cortical representation of threat—in the pathogenic model of anxiety⁶⁵.

STAR METHODS

RESOURCE AVAILABILITY

Lead contact

Further information and requests for resources should be directed to and will be fulfilled by the Lead Contact, Wen Li (wenli@psy.fsu.edu).

Materials availability

All experiment stimuli are publicly available at a repository (https://github.com/LiLabFSU/Threat-memory-in-human-olfactory-cortex). This study did not generate new unique reagents.

Data and code availability

• Anonymized data, including fMRI, behavioral, SCR and respiratory data, have been deposited at a repository (https://github.com/LiLabFSU/Threat-memory-in-human-olfactory-cortex), as listed in the key resource table. They are publicly available as of the date of publication.

• Analysis scripts have been deposited at a repository (https://github.com/LiLabFSU/Threat-memory-in-human-olfactory-cortex), as listed in the key resource table. They are publicly available as of the date of publication.

• Any additional information required to reanalyze the data reported in this paper is available from the Lead Contact upon request.

Key resources table

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Deposited data
Experiment stimuli	This study	https://github.com/LiLabFSU/Threat-memory-in-human-olfactory-cortex
Data (fMRI & behavioral)	This study	https://github.com/LiLabFSU/Threat-memory-in-human-olfactory-cortex
Analysis scripts	This study	https://github.com/LiLabFSU/Threat-memory-in-human-olfactory-cortex
Software and algorithms
MATLAB, v2021a	Mathworks	RRID: SCR_001622
SPSS statistics, v28	IBM	RRID: SCR_019096
SPM, v12	Wellcome Centre for Human Neuroimaging	RRID: SCR_007037
Cogent 2000, v25	Wellcome Centre for Human Neuroimaging	RRID: SCR_015672
AcqKnowledge	Biopac Systems	RRID: SCR_014279

EXPERIMENTAL MODEL AND SUBJECT DETAILS

Participants

Thirty-three individuals (13 males; age 19.9 ± 2.0 years, range 18–25) participated in this two-session fMRI experiment in exchange for course credit or monetary compensation. All participants were right-handed, with normal olfaction and normal or corrected-to-normal vision. Participants were screened to exclude acute nasal infections or allergies affecting olfaction, any history of severe head injury, psychological/neurological disorders, or current use of psychotropic medication. All participants provided informed consent to participate in the study, which was approved by the University of Wisconsin-Madison Institutional Review Board. One participant who failed to provide risk ratings on Day 1 and another who failed to follow the ODT task instruction were excluded from the corresponding analyses. Two participants were excluded from fMRI analysis due to metal artefact and excessive movement.

METHOD DETAILS

Anxiety Assessment

We used the Behavioral Inhibition Scale (BIS) to measure trait anxiety⁶⁶. The BIS is a 7-item self-report questionnaire (score range: 7-28) measuring the strength of the behavioral inhibition system and threat sensitivity, known to reflect trait anxiety. This scale is neurobiologically motivated, with high reliability and strong predictive validity of anxiety ^67,68, and recommended by the National Institute of Mental Health (NIMH) to measure the construct of “potential threat (anxiety)”.

Stimuli

We included two neutral odorants, acetophenone (5% l/l; diluted in mineral oil) and eugenol (18% l/l). These odors have received similar ratings on valence, intensity, familiarity, and pungency and been used as neutral odors in previous research ^69,70. They were labeled as odors “A” and “B” to the participants and were parametrically mixed into five mixtures to create a linear morphing continuum: 80% A/20% B, 65% A/35% B, 50% A/50% B, 35% A/65% B, and 20% A/80% B (Figure 1A). The two extreme mixtures (20% A/80% B and 80% A/20% B) served as conditioned stimuli (CS), differentially conditioned as threat CS (CSt) and safety CS (CSs), counterbalanced across participants, via pairings with threat and neutral unconditioned stimuli (UCS), respectively. The three intermediate mixtures were non-conditioned stimuli (nCS) and denoted as nCSt (neighboring odor of the CSt), nCSm (midpoint of the continuum), and nCSs (neighboring odor of the CSs), respectively. The UCS were bimodal (visuo-auditory) stimuli, including 7 pairs of disgust images (three depicting dirty toilets and four vomits) and disgust sounds (i.e., vomiting) and 7 pairs of neutral images (household objects) and neutral sounds. Images were chosen from the International Affective Picture Set (IAPS)⁷¹ and internet sources⁶⁴. Disgust sounds were from the disgust subset of human affective vocalizations ⁷², and neutral sounds were pure tones (300, 500, and 800 Hz). Skin conductance response (SCR) data confirmed effectiveness of the aversive (vs. neutral) UCS (Figure 1A).

Odor stimuli were delivered at room temperature using an MRI-compatible sixteen-channel computer-controlled olfactometer (airflow set at 1.5 L/min), which permits rapid odor delivery in the absence of tactile, thermal or auditory confounds ^69,73,74. Stimulus presentation and collection of responses were controlled using Cogent2000 software (Wellcome Department of Imaging Neuroscience, London, UK) as implemented in Matlab (Mathworks, Natick, MA).

Unconditioned skin conductance response (SCR)

SCR was acquired with BioPac MP150 (BIOPAC systems, Goleta, CA) from two MRI-compatible Ag/AgCl electrodes placed on the middle phalanx of the second and third digits of the non-dominant (left) hand at a sampling rate of 1000 Hz. A low-pass filter (0.5 Hz) was applied offline to eliminate MRI scanning artifacts. For each trial, evoked SCR response was defined by the magnitude of trough-to-peak SCR deflection during the interval between 0.5 s pre- and 7 s post-UCS onset (ITI = 14.1 s), with a minimal evoked deflection of 0.02 μS. We compared SCR evoked by the aversive and neutral UCS during conditioning. In support of its effectiveness, the aversive UCS produced significantly greater SCR than did the neutral UCS (t = 2.88, p = .007; Figure 1A & Figure S1).

Odor discrimination task (ODT)

During the two-alternative forced-choice odor discrimination task (2-AFC ODT), each trial began with a visual “Get Ready” cue, followed by a 3-2-1 countdown and a sniffing cue, upon which participants were to take a steady and consistent sniff and respond whether the odor smelled like Odor A or B by button pressing (Figure 1B). Each of the five odor mixtures was presented 15 times, in a pseudo-random order without repetition over two consecutive trials. Seven additional trials with a central, blank rectangle on the screen (no response required) were randomly intermixed with the odor trials to help minimize olfactory fatigue and establish a non-odor fMRI baseline. Trials recurred with a stimulus onset asynchrony of 14.1 s.

Experiment procedure

Pre-experiment screening

Approximately a week before the experiment, participants visited the lab to be screened for normal olfactory perception. They were also introduced to acetophenone and eugenol as Odors “A” and “B” and practiced on a 2-AFC ODT between the two odors. They also provided ratings on the five odor mixtures. We performed analyses on odor ratings to exclude confounds related to inherent odor stimulus differences. Baseline ratings for all five odor mixtures on valence, intensity, familiarity, and pungency were submitted to separate repeated-measures ANOVAs, which revealed no significant difference among five odor mixtures on any of the scales (all F values < 1.61, all P values > 0.182).

Experiment Day 1

Participants first performed the 2-AFC ODT, then underwent differential conditioning, and then repeated the 2-AFC ODT (Figure 1C). During differential conditioning, CSt and CSs odors were presented (seven trials each, randomly intermixed; ITI = 12 s) for 1.8 s while the aversive or neutral UCS were presented respectively for 1.5 s at 1 s after CS odor onset, with 100% contingency. To prevent extinction by the repeated unreinforced CS presentation during the post-conditioning 2-AFC ODT (on both Day 1 and Day 9), five extra trials of CSt paired with the aversive UCS were randomly inserted ^14,29,30,75. Data from these trials were excluded from analysis. After the post-conditioning ODT, the five odor mixtures were presented (three trials per odor mixture, randomly intermixed), to which participants performed risk rating (likelihood of an aversive UCS to follow the odor) on a visual analog scale (VAS; 0-100%).

Experiment Day 9

Participants repeated the 2-AFC ODT and risk rating. After that, participants underwent an independent olfactory localizer scan involving a simple odor detection task, from which functional ROIs were extracted. Four additional odorants (α-ionone, citronellol, methyl cedryl ketone, 2-methoxy-4-methylphenol), neutral in valence and matched for intensity, were presented in this task (15 trials/odor), pseudo-randomly intermixed with 30 air-only trials.

Respiratory monitoring

Respiration measurements were acquired (1000 Hz) during the ODT, using a BioPac MP150 (AcqKnowledge software) with a breathing belt affixed to the participant’s chest to record abdominal or thoracic contraction and expansion. For each odor trial, a sniff waveform was extracted from a 6 s window post sniff onset and was baseline-corrected by subtracting the mean activity within 1 s preceding sniff onset. Sniff parameters (inspiratory volume, peak amplitude, and peak latency) were generated by averaging across all 15 trials per odor. We examined respiration parameters during the 2-AFC ODT, including peak amplitude, peak latency, and sniff inspiratory volume. ANOVAs (Odor X Time) on these sniff parameters revealed no effects of odor or odor-by-time interactions (all P values > 0.095). These results thus ruled out variations in sniffing as potential confounds.

Imaging acquisition and preprocessing

Gradient-echo T2 weighted echoplanar images (EPI) were acquired with blood-oxygen-level-dependent (BOLD) contrast and sagittal acquisition on a 3T GE MR750 MRI scanner. Imaging parameters were TR/TE = 2350/20 ms; flip angle = 60°, field of view = 220 mm, slice thickness = 2 mm, gap = 1 mm; in-plane resolution/voxel size = 1.72×1.72 mm; matrix size = 128×128. A field map was acquired with a gradient echo sequence, which was coregistered with EPI images to correct EPI distortions due to susceptibility. A high-resolution (1×1×1mm³) T1-weighted anatomical scan was acquired. Five scan runs, including pre-conditioning, conditioning, Day 1 post-conditioning, Day 9 post-conditioning, and odor localizer, were acquired. Six “dummy” scans from the beginning of each scan run were discarded to allow stabilization of longitudinal magnetization. Imaging data were preprocessed in SPM12 (www.fil.ion.ucl.ac.uk/spm), where EPI images were slice-time corrected, realigned, and field-map corrected. Images collected on both Day 1 and Day 9 sessions were spatially realigned to the first image of the first scan run on Day 1, while the high-resolution T1-weighted scan was co-registered to the averaged EPI of both scan sessions. All multivariate pattern analyses were conducted on EPI data that were neither normalized nor smoothed to preserve signal information at the level of individual voxels, scans, and participants.

A general linear model (GLM) was computed on pre-conditioning ODT, conditioning, Day 1 post-conditioning ODT, and Day 9 post-conditioning ODT scans. Applying the Least Squares All (LSA) algorithm, we set each odor trial as a separate regressor, convolved with a canonical hemodynamic response function ⁷⁶. Six movement-related regressors (derived from spatial realignment) were included to regress out motion-related variance. For the odor localizer scan, we applied a GLM with odor and no odor conditions as regressors, convolved with a canonical hemodynamic response function and the temporal and dispersion derivatives, besides the six motion regressors of no interest. A high-pass filter (cut-off, 128 s) was applied to remove low-frequency drifts and an autoregressive model (AR1) was applied to account for temporal nonsphericity.

ROI definition

All four ROIs (APC, PPC, OFC, and amygdala) were manually drawn on each participant’s T1 image in MRIcro ⁷⁷ (Figure 1D). The olfactory OFC (OFC_olf) was defined by a meta-analysis ⁷⁸ and a prior study ⁷⁹, and the other ROIs were defined by a human brain atlas ⁸⁰. Left and right hemisphere counterparts were merged into a single ROI. Functional constraints were applied to these anatomical ROIs based on the odor-no-odor contrast of the independent odor localizer scan for each participant, with a liberal threshold at P < 0.5 uncorrected ²⁹.

QUANTIFICATION AND STATISTICAL ANALYSIS

fMRI analysis

Representational similarity analysis (RSA)

The RSA uses correlations across multivoxel response patterns to indicate the degree of similarity in response patterns ^81,82 and thus presents an effective test of pattern differentiation ²⁹. For each participant and every ODT session, trial-wise beta values were extracted for all voxels within a functionally constrained ROI, which were then averaged across all 15 trials for each odor mixture, resulting in an odor-specific linear vector of beta values across a given ROI. Pearson’s correlation (r) was computed between all pairs of pattern vectors at each session, resulting in a 5 x 5 correlation matrix—the representational similarity matrix—for each session. To directly represent pattern differentiation, this matrix was converted into a representational dissimilarity matrix (RDM) by replacing the r values with dissimilarity scores (1 – r)⁸³. To assess pattern differentiation, we computed a pattern differentiation index (PDI) based on the RDM matrix (dissimilarity/distance = 1-r), following Fisher’s Z transformation: PDI = [(d1+ d4) – (d2 + d3)], reflecting the dissimilarity/distance of CSt and CSs from their neighboring nCS odors (nCSt and nCSs, d1 and d4 respectively), controlled by the dissimilarity/distance between the midpoint odor (nCSm) and its neighbors (d2 and d3).

Tuning analysis

We adopted a voxel-based tuning analysis used for visual sensory encoding ^38,39 to assess olfactory cortical tuning. Trial-wise beta values (5 odors × 15 trials) for each voxel were normalized (by z-scoring) across trials after removing the trial-wise mean beta across the ROI, from which we calculated mutual information (MI) conveyed by each voxel about each odor (see below). As low MI values (i.e., minimal mutual dependence between the distribution of responses and odor) reflect indiscriminant or random responses to all odors, voxels with bottom 10% MI values in a given ROI were excluded ³⁸. Voxel-based tuning was defined by the odor mixture eliciting the largest beta (i.e., optimal odor). As such, each of the remaining voxels was classified into one of five odor classes. In line with animal tuning analysis ^12,43, we examined the voxels tuned to the neighboring odors (nCSs and nCSt) of the CS before conditioning and measured their tuning shift to the CS (relative to the neighboring nCS odor/nCSm) after conditioning. Accordingly, we derived a tuning shift index (TSI) for Day 1 and Day 9 post-conditioning: TSI = (% CSs – % nCSm) + (% CSt – % nCSm), reflecting the % of initially nCSs/nCSt voxels that became tuned to the neighboring CSs/CSt, respectively, relative to the % of initially nCSs/nCSt voxels that became tuned to the other neighboring odor—nCSm.

In terms of MI calculation, first, we converted the beta values into a discrete variable (B) by dividing the range of betas into a set of equidistant bins (b). The size of the bins was determined by Freedman-Diaconis’ rule [bin size = (max(B) – min(B))/2*IQR*n^−1/3], where n is the number of trials (n = 75). We selected the median bin size of all voxels within an ROI based on the pre-conditioning data and held it constant for the post-conditioning sessions (Day 1 and Day 9). Next, we computed for each voxel the entropy of (discretized) responses (B) as follows:

$$H(B)=-\sum _{b\in B}p(b)lo{g}_{2}p(b)$$

where p(b) is the proportion of trials whose responses fall into bin b. Then, we computed conditional entropy H(B∣o), the entropy of responses given knowledge of the odor condition, as follows:

$$H(B\mid o)=-\sum _{o\in O}p(o)\sum _{b\in B}p(b\mid o)lo{g}_{2}p(b\mid o)$$

where p(b∣o) is the proportion of trials falling into bin b when responding to a certain odor (o). The index of MI(B; O), i.e., the amount of information a voxel conveys for an odor, was calculated as the reduction in entropy of responses given knowledge of the odor condition:

$$MI(B;O)=H(B)-H(B\mid o)$$

Statistical analysis

Using analyses of variance (ANOVAs) of Odor (five mixtures) and Time (Day 1 Post and Day 9 Post), we performed trend analysis over the odor continuum on risk ratings and ODT response to capture the warping of affective and perceptual spaces by conditioning. We hypothesized that affective learning via conditioning would change the baseline neutral trend to an ascending safety-to-threat trend (Figure 2a). We further hypothesized that differential conditioning would enhance perceptual discrimination of the CS, expanding odor quality distances between the CS and their neighboring odors; resulting changes in odor quality space (i.e., differential CS endorsement rates; Post - Pre) would conform to a cubic trend, anchored by respective increase and decrease in “CSt” rate for the neighboring odors of CSt and CSs—nCSt and nCSs (Figure 2B). As for the neural mechanisms, i.e., enhanced pattern differentiation between the CS and similar (neighboring) nCS and tuning shift towards the CS, we conducted ANOVAs of ROI (APC/PPC) and Time (Day 1 Post and Day 9 Post) on differential PDI scores and TSI scores, respectively. Finally, we examined modulatory effects of anxiety using Pearson’s correlation of BIS scores with behavioral and neural effects of conditioning. Significance threshold was set at P < 0.05. Given the clear a priori hypotheses, one-tailed tests were accepted and are explicitly noted in the Results (two-tailed tests are not explicitly noted). To protect for Type I error, only significant effects in the ANOVAs were followed up with hypothesis testing. Correlational analysis with anxiety involved multiple tests, which were corrected using the false discovery rate (FDR) criterion (i.e., FDR P < 0.05).

Highlights.

• Threat conditioning produces long-term affective and perceptual memory

• Associative plasticity emerges in human primary olfactory (piriform) cortex

• The piriform cortex exhibits long-term pattern differentiation and tuning shift

• These sensory cortical underpinnings of threat memory are hyperactive in anxiety