Abstract
Emotionally salient experiences are encoded and remembered more strongly, an effect that can be amplified by hormones like cortisol. Such memories can in turn profoundly influence later behavior. However, little is known about the link between amplified salience encoding and subsequent behavior. This pathway may be particularly important for risky alcohol drinking, which has been linked to sensitized salience responses, memory, and cortisol. To test this possibility, we integrated pharmacology using a double-blind cross-over design with fMRI, cognitive, and motivation assays across a range of healthy male and female social drinkers. As anticipated, cortisol enhanced memory for salient alcohol-related events; critically, this bias was in turn associated with later alcohol motivation. Increased alcohol motivation was particularly pronounced in more susceptible risky drinkers, for whom cortisol enhanced brain salience responses to alcohol. These sensitized salience responses predicted both memory biases and alcohol motivation. Together, these findings reveal maladaptive consequences of enhanced salience encoding.
Significance Statement
Our memories for salient experiences can strongly influence what we choose to do. Despite the importance of such memories for motivated behaviors like alcohol drinking, the mechanisms by which salient memories are strengthened and influence later behavior are unclear. In the current study, we integrated findings from human research (glucocorticoids enhance salient memory encoding) with findings from rodent models (glucocorticoids enhance alcohol motivation, especially with a history of ethanol exposure) to examine whether these processes are linked in humans. We found evidence for this pathway: hydrocortisone amplified brain signatures of salience (particularly for participants who drank more), which predicted enhanced memory. Both brain salience and enhanced memory predicted increased alcohol motivation. This translational pathway reveals downsides to amplifying salient memories.
Introduction
Emotionally salient experiences hold a distinct place in our lives and our memories. The formation, or encoding, of such salient memories has been shown to engage specific salience-related neural mechanisms (Hamann, 2001; Phelps, 2004; LaBar and Cabeza, 2006), and can be selectively enhanced by agents like glucocorticoids (cortisol in humans; Joëls et al., 2006; de Quervain et al., 2009; Finsterwald and Alberini, 2014; Schwabe et al., 2022). Such targeted benefits may be driven by interactions between glucocorticoids and the noradrenergic system (Roozendaal et al., 2006; Segal et al., 2014). These memory advantages have been theorized to allow such meaningful experiences to guide future behavior (McGaugh, 2015). However, little is known about the behavioral consequences of enhancing salient encoding.
Amplifying salient encoding may have particularly important consequences for behaviors like risky drinking. First, memory for salient (here, drug-related) experiences plays a crucial role in later motivation to use addictive drugs like alcohol (White, 1996; Shaham et al., 2003; Hyman et al., 2006; Koob and Volkow, 2016; Bornstein and Pickard, 2020). These findings extend to laboratory measures of memory. For example, we found that patients with alcohol use disorder (AUD) had stronger memory for personally salient alcohol-related contexts, and that these memories predicted drinking behavior over the next month (Goldfarb et al., 2020b). Thus, enhancing encoding of salient alcohol events might promote drinking upon cue-induced retrieval of those memories (Goldfarb and Sinha, 2018). Second, cortisol can potentiate the salience of drug-related information. Although cortisol is often considered in the context of stress and aversive learning, glucocorticoids can be appetitive (Piazza and Le Moal, 1997) and are released in response to substances like alcohol (Ellis, 1966; Armario, 2010) or even alcohol-related cues. For example, rodents bred to show greater incentive salience for drug-related cues have higher glucocorticoid responses (Flagel et al., 2014) and sign-tracking (or a focus on cues that predict drug outcomes) has been linked to glucocorticoid release (Tomie et al., 2008; Flagel et al., 2009). In humans, glucocorticoids can potentiate craving in the presence of alcohol cues (Soravia et al., 2021). Thus, any enhancing effects of cortisol on salience encoding may be particularly relevant for alcohol motivation. Third, glucocorticoids can change the brain circuits supporting successful encoding, favoring striatal-dependent representations (Goldfarb and Phelps, 2017; Schwabe et al., 2022), and may promote the formation of memories that would drive later maladaptive drinking behavior (Goodman and Packard, 2016; Goldfarb and Sinha, 2018). Finally, cortisol effects on brain and behavior can differ based on drinking history (Blaine and Sinha, 2017). For example, blocking the glucocorticoid system decreased alcohol self-administration in rodents (Koenig and Olive, 2004; McGinn et al., 2021; Farokhnia et al., 2022) and reduced drinking in humans (Donoghue et al., 2020; Farokhnia et al., 2022)—particularly for those who already had a tendency to engage in risky drinking (Vendruscolo et al., 2012; Repunte-Canonigo et al., 2015; Vendruscolo et al., 2015; Savarese et al., 2020). Thus, cortisol effects on salience encoding, and any downstream consequences of such effects, may particularly impact individuals who are more susceptible due to engaging in risky drinking behavior. Based on these findings, we anticipate that cortisol will lead to stronger responses in salience-related brain regions (Van Marle et al., 2010), which would potentiate encoding, particularly for arousing alcohol-related information (Schwabe et al., 2022) and for participants with more susceptible drinking profiles.
To investigate the potential consequences of cortisol amplifying salient encoding in humans, we combined pharmacology targeting the glucocorticoid system, functional MRI, and behavioral assays of subsequent memory and alcohol motivation. Using a randomized, placebo-controlled, cross-over design (Fig. 1A), we recruited healthy participants (N = 27 social drinkers) to complete two rounds of associative encoding after taking 20 mg hydrocortisone or a visually identical placebo tablet (order counterbalanced; 54 scans total). To assess divergent consequences based on drinking history, participants were recruited to display a range from light to risky drinking patterns per NIAAA criteria as assessed by the Alcohol Use Disorders Identification Test screening tool (Fig. 2 inset). During encoding, participants formed associations between trial-unique object (50% alcoholic beverages)/scene pairs while undergoing fMRI. To assess the downstream consequences of altered encoding, we measured memory and alcohol motivation the next day. We first asked participants to remember encoded object/scene pairs and then performed an in-laboratory evaluation of alcohol motivation. This design enabled us to assess, across a range of social drinking profiles, the potential consequences of cortisol for encoding, memory, and subsequent behavior.
Consequences of cortisol for later memory and alcohol motivation. A, Experimental design. Each participant completed all four sessions: on one week, they received 20 mg hydrocortisone before encoding object/scene pairs; the other, they received an identical placebo tablet. The day after encoding, they were asked to remember which scene was paired with each object at encoding. Then, they were presented with three beers in an alcohol taste test (ATT), and given 10 min to decide whether the beers were the same or different. Amount of beer they consumed (mL) and blood alcohol concentration (BAC) were used to quantify alcohol motivation. B, Hydrocortisone led to significant elevations in salivary cortisol that were maintained throughout encoding. Light lines = individual participants; black stars = differences between pills per timepoint; pink stars = differences between timepoints post-hydrocortisone. Error bars = 95% CIs. C, Cortisol induces bias toward remembering salient alcohol-paired contexts the next day (chance = 25%). Left; dots = memory per participant within each image category. **p < .01, *p < .05. Error bars = SEM. Right; trial-level analysis, dots = memory and salience scores for individual trials per participant. Shaded regions = 95% CI. D, Bias toward remembering salient alcohol-paired contexts was associated with greater overall alcohol motivation. Shaded regions = 95% CI. Dots = average memory and alcohol motivation per participant.
Susceptible individuals show distinct cortisol effects on alcohol motivation and brain responses. A, Behavior on ATT. More susceptible individuals (those with higher AUDIT scores) showed greater alcohol motivation after retrieving memories encoded under cortisol. Inset shows distribution of raw AUDIT scores. As these were not normally distributed, log-transformed data was used in all analyses. Left = mL (amount of beer consumed); right = BAC (blood alcohol concentration). Dots = drinking behavior per participant after retrieving memories encoded under cortisol (pink) or placebo (turquoise). Shaded areas = 95% CI. ***p < .001 B, SALcort network identified as responding more to hydrocortisone and alcohol-related events for more susceptible individuals (see also Fig. 3). C, Illustrations of three-way interactions for example identified clusters, driven by higher responses to alcohol after hydrocortisone among more susceptible individuals. STG = superior temporal gyrus. Shaded areas = 95% CI.
Materials and Methods
Participants
Twenty-seven individuals (16 males; M = 27.6 years old) were recruited from the greater New Haven community. All provided written informed consent and all procedures were approved by the Yale Medical School Institutional Review Board. A power analysis performed on pilot data, with an α = 0.05 and power = 0.85, indicated 25 participants were needed to observe cortisol effects on emotional memory (G*Power 3.1.9.3). All participants completed all five sessions; however, one participant’s placebo memory data were excluded due to a technical error resulting in mismatched stimuli between encoding and retrieval. Participants were beer-drinking right-handed adults, fluent in English, BMI 18–30, had normal or corrected-to-normal color vision, and did not meet criteria for any substance use disorder including AUD. Participants were recruited to show a range of drinking behaviors. Exclusion criteria included current use of medications that interfere with the HPA axis response (e.g., SSRIs, beta-blockers, or corticosteroids); peri and post-menopausal females, pregnant or lactating females, and those with hysterectomies. Reported contraceptives included IUDs (N = 3, 27% of female participants) and oral contraceptives (N = 3).
Procedure
The study included 5 sessions in a double-blind, placebo-controlled, crossover design (Fig. 1A). At intake, participants completed the Structured Clinical Interview for the DSM-5 (SCID; First et al., 2015) with a trained interviewer to determine whether they met criteria for current/lifetime AUD (N = 3 ineligible). To classify drinking behavior, participants completed the AUDIT (Bush et al., 1998) and Alcohol Severity Index (Mäkelä, 2004).
Eligible participants completed two rounds of encoding while undergoing functional magnetic resonance imaging (fMRI; Days 1 & 3) and retrieval (Days 2 & 4, 24 h post-encoding), a minimum of one week apart. At the start of one encoding session, participants consumed 20 mg of hydrocortisone; in the other, they received a visually identical placebo pill (order counterbalanced). Study procedures were registered on clinicaltrials.gov (NCT04896489). Participants were instructed not to consume alcohol for 24 h prior to fMRI sessions, and all sessions began with urine (drug/pregnancy) and breathalyzer (sobriety) testing. Sessions occurred between 12:00 and 6:00 pm to control for circadian variation in cortisol levels (Lupien et al., 2007). Immediately upon completing memory retrieval tasks, participants performed an adapted version of the Alcohol Taste Task (ATT).
Hydrocortisone administration and cortisol measurement
Hydrocortisone randomization and administration procedures were described previously (Sherman et al., 2023).
Salivary cortisol
Participants provided 6 saliva samples throughout the fMRI sessions (baseline, immediately pre-encoding, and post-scan) to assess salivary cortisol levels. Samples were collected using Starstedt Salivate Tubes and were analyzed using radioimmunoassay (RIA) by the Yale Center for Clinical Investigation (YCCI).
Affect
Shortly after arrival at each visit, participants completed the Positive and Negative Affect Scale (PANAS; Watson et al., 1988). On scan days, participants also completed the questionnaire upon finishing the scan.
fMRI procedure
BOLD fMRI data were acquired while participants performed the encoding task. Procedures are described in detail in Sherman et al. (2023), and scans that are the focus of current analyses are detailed below.
MRI acquisition parameters
Data were acquired on Siemens 3T Prisma scanners using a 64-channel coil at the Magnetic Resonance Research Center at Yale University as described (Sherman et al., 2023). Data were acquired across three scanners (N = 4 on scanner A, N = 2 on scanner B, N = 21 on scanner C). Parameters were the same across all scanners. Every participant completed both MRI sessions in the same scanner. Follow-up analyses confirmed that the same clusters were identified in the group-level model if scanner was included as a covariate, and that AUDIT scores, memory, and drinking behavior did not differ significantly based on which scanner participants used.
Functional images were collected using an echoplanar imaging (EPI) sequence (TR = 1,000 ms, TE = 30 ms, 75 axial slices, voxel size = 2 × 2 × 2 mm, flip angle = 55°, MF = 5, FOV = 220 × 220). Anatomical data were acquired using a T1-weighted 3D MPRAGE sequence (TR = 2,400 ms, TE = 1.22 ms, 208 sagittal slices, voxel size = 1 × 1 × 1 mm, flip angle = 8°, FOV: 256 × 256).
Memory tasks
Participants performed two rounds of encoding and retrieval (adapted from Goldfarb et al., 2020b), once receiving hydrocortisone and once receiving placebo prior to encoding. No drugs were administered prior to retrieval.
Encoding
During each encoding trial, participants viewed images of objects and scenes (5 s) and were instructed to vividly imagine the object as a part of the scene. They then used a button box to indicate how they felt when imagining the object/scene pair (2 s each) including valence (unhappy, happy, or neutral), arousal (1 not at all intense—4 extremely intense), and alcohol craving (1 not at all—4 a lot). To maximize distinct trial-level BOLD signal, the trials were separated by a jittered inter-trial interval (M = 2 s). Stimuli were presented and responses were acquired using MATLAB with Psychophysics Toolbox (Brainard, 1997; Pelli, 1997).
During each encoding session, participants completed two blocks (order randomized), each with 40 object/scene pairs. One block contained images of neutral, handheld objects, whereas the other block contained alcohol-related objects. Participants were informed that their memory for object/scene pairs would be tested the next day.
Stimuli were presented in separate blocks to facilitate comparison with prior experiments (Goldfarb et al., 2020b) and limit potential confounds of heightened attention allocation to alcohol-related stimuli with heavier drinking (Townshend and Duka, 2001) and stress (Field and Powell, 2007), as such processes could interfere with immediate encoding of neutral stimuli if they are presented in an interleaved fashion (Talmi and McGarry, 2012; Barnacle et al., 2016). This approach was also intended to promote stronger alcohol-related salience responses as in the past investigations of cue reactivity.
Using neutral handheld objects for comparison maximized differences between stimulus sets and created a baseline that would be relatively stable across different drinking levels and pill conditions. For example, although some cue reactivity studies use nonalcoholic beverages matched on emotional salience and perceptual detail as a control condition (e.g., Dager et al., 2014), emotion ratings for purportedly matched stimuli vary with drinking history (Wrase et al., 2002), sex (Pulido et al., 2010), and IQ (van Duijvenbode et al., 2012). Thus, using common household objects would provide an affectively neutral, perceptually distinct baseline against which to assess biases in alcohol-related memory.
Retrieval
Twenty-four hours after each encoding session, participants returned for memory retrieval. Tests were conducted separately for each stimulus type and followed the same order as encoding (e.g., if alcohol/scene pairs were encoded first, then those events were retrieved first). All memory tests were described in detail in Sherman et al. (2023). In brief, we first tested participants’ memory for individual objects, presenting 80 old images (viewed the day before) and 80 new images and asking participants to indicate whether each image was old or new. Then, we presented all old objects, and asked participants to indicate whether they had been previously paired with an indoor or outdoor scene, followed by selecting the exact scene from a list. As this context, memory representation is the focus of the current manuscript, we describe these trials in more detail below. Finally, participants were presented with the original object/scene pairs and asked to recall the feelings that they experienced when viewing these pairs during encoding on the previous day.
Retrieval
To assess context memory, participants were presented with each object and asked to choose which of four presented scenes was paired with the object (4 s). Scenes included the original scene pairmate from encoding and another scene viewed during encoding (familiarity control), each with a matched perceptual lure. They were instructed to make their best guess if they were not sure. Selection of the original scene pairmate was coded as correct context memory.
Alcohol motivation
After the retrieval assessments, participants completed a modified version of the alcohol taste test (ATT; Marlatt et al., 1973). They were presented with three identical cups filled with cold beer (1050 ml total). Participants were instructed to taste each beer and determine whether they were the same or different. They were told to drink as much as they needed to make their decision and that they would receive a $10 bonus for correct responses. The three cups were all filled with Bud Light (Anheuser–Busch; 4.2% alcohol).
After 10 min, the experimenter returned, removed the three cups, and measured how much was consumed (due to experimenter error, the amount consumed during two sessions [1 cortisol, 1 placebo] was not recorded). Participants were not aware that the amount they drank was being measured. After drinking water to rinse out any alcohol residue, participants’ BAC was measured with a breathalyzer. After another 10 min and a snack, the second BAC reading was taken. If BAC was >20 mg/dL, participants were offered another snack and told the next reading would be performed in 10 min. If BAC was <20 mg/dL, the session was over, and participants were allowed to leave. For safety, participants were not allowed to drive after these sessions.
Analysis
Analyses of fMRI data were conducted in FSL (6.0.1) and AFNI (AFNI_23.0.07). Analyses of behavioral data were conducted in R (4.3.0). As AUDIT scores were not normally distributed (Shapiro-Wilk normality test W = .88, p = .004), these were log-transformed prior to all analyses. Behavioral analyses largely used linear mixed-effects models (LME, nlme package, Pinheiro and Bates, 2024) with follow-up tests conducted using emmeans (Lenth, 2023). Effect sizes (
Pill efficacy
Salivary cortisol was analyzed using LME as a function of time (baseline, post-drug, post-scan), pill, log-AUDIT, and their interactions, with session as a covariate and subject as a random effect. Changes in positive and negative affect were assessed by modeling change per PANAS subscale as a function of pill, log-AUDIT, and their interaction, with session as a covariate and subject as a random effect.
Drinking
Drinking behavior during the ATT (both amount consumed and BAC) were each analyzed using LMEs as a function of pill, log-AUDIT, and their interaction, with session as a covariate and subject as a random effect.
Memory
As noted above, participants encoded 40 alcohol/scene pairs per encoding session. These alcoholic images portrayed an equal number of wine, beer, liquor, or mixed drinks to help ensure that participants viewed images that they would find personally salient (N = 10 each). To determine which type of alcohol participants perceived as most salient, we computed a weighted sum of craving and arousal ratings (plus a penalty for rating alcohol/scene pairs as neutral) averaged across sessions per beverage type and participant. This resulted in a separation of the alcohol/scene pairs participants found most salient (N = 10) from alcohol/scene pairs featuring other types of beverages (N = 30; as described in Goldfarb et al., 2020b). Then, average context memory was computed per participant, session, and ranking (most salient alcohol, other alcohol, object). Note that, because this method uses rank-ordering, any cortisol-induced changes in ratings would not impact the number of trials assigned to each category. Context memory was analyzed as a function of pill, log-AUDIT, and ranking (with all interactions), including session and stimulus block order as covariates and subject as a random effect. As the secondary analysis, we computed a salience score per trial using the same weighted sum described above. We then ran a general linear model to predict trial-level memory as a function of pill and salience score, including trial number and session as covariates.
Relationship between brain responses and salient memory
Using the brain network identified from the group model below, we computed mean trial-evoked responses per participant and pill condition. We then modeled the bias toward remembering salient alcohol contexts (salient alcohol/scene memory—other alcohol/scene memory) as a function of brain responses and pill, with session and log-AUDIT as covariates and subject as a random effect. As a follow-up analysis, we ran a binomial general linear model predicting whether each alcohol trial was remembered or forgotten as a function of trial-level brain responses, salience (most salient vs other alcohol), pill, and log-AUDIT, including trial number and session as covariates.
Relationship between salient memory and drinking
For each participant, we computed the bias toward remembering salient alcohol contexts (salient alcohol/scene memory—other alcohol/scene memory) and their drinking behavior (BAC). We then modeled BAC as a function of pill, log-AUDIT, and memory bias (with all interactions), including session as a covariate and subject as a random effect. As a follow-up analysis, we used mean salience scores for alcohol trials per participant as a predictor in the same model.
fMRI preprocessing
Encoding runs were preprocessed as described in Sherman et al. (2023). In brief, after verifying that all runs had below-threshold (<1.5 mm absolute mean frame-to-frame displacement; Jenkinson et al., 2002), data were skull-stripped (BET; Smith, 2002), pre-whitened (FILM; Woolrich et al., 2001), and high-pass filtered at 0.01 Hz. Confounds including motion (6 DOF, plus temporal derivatives), white matter time series (plus temporal derivatives), and nonlinear motion outliers were regressed out from the data in a GLM (FSL’s FEAT; http://www.fmrib.ox.ac.uk/fsl). Model residuals were aligned to a reference functional scan, followed by boundary-based registration to the participant’s high-resolution anatomical scan (Greve and Fischl, 2009) and linear registration (12 DOF) to standard (MNI 152) space (Jenkinson et al., 2002).
fMRI analysis
To identify cortisol-modulated brain responses to alcohol that varied for individuals engaged in more risky drinking, first-level models estimating trial-evoked responses were run in FSL, and group-level models identifying differences in these responses across participants were run in AFNI.
First-level model
After regressing out motion and other confounds (see above), residuals per run were entered into a general linear model (GLM) with 40 regressors, one per trial. Each trial was modeled using a boxcar (initiating at picture onset and terminating at picture offset) convolved with a canonical double-gamma HRF. Separate first-level GLMs were conducted for each run in native space. For each run, these trial-level beta estimates were averaged (yielding one estimate per run) and aligned to MNI space. Mean beta estimates were smoothed with a 6 mm FWHM Gaussian kernel prior to group-level analysis.
Group-level model
Mean beta estimates per participant and run were entered into a linear mixed-effects model (3dLMEr) with responses modeled as a function of pill, stimulus type, and log-AUDIT, with subject as a random effect. Voxels showing a significant main effect of pill and stimulus type, as well as the three-way interaction, were identified using a voxelwise threshold of FDR-corrected q < .001.
Overlap with the existing networks
To help interpret the brain regions identified in the group-level model, we compared significant voxels to a set of previously identified intrinsic networks (Laird et al., 2011). For each of these 20 previously defined networks, we computed Sørensen-Dice similarity coefficients to our identified regions.
Relationship to subjective salience
The clusters identified from the three-way interaction in the group-level model were used to mask each participants’ trial-level beta estimates. We then computed the mean response within this network per trial, resulting in Ntrials network responses per participant and run.
To assess whether these network responses tracked subjective salience, we separated trials based on encoding ratings and computed mean beta estimates for each rating in each run (e.g., for trials rated as different levels of arousal). These mean beta estimates were then analyzed as a function of rating, stimulus type, and pill (and all interactions), including log-AUDIT and session as covariates and subject as a random effect. For between-subjects analyses, we computed overall mean beta estimates per run (similar to inputs for the predictive model below) and mean ratings per run (e.g., averaged craving rating throughout the run). We then modeled beta estimates as a function of stimulus type, pill, and ratings (with all interactions), with session as a covariate and subject as a random effect.
Predictive modeling
The following models were used to determine whether the clusters identified as showing a significant three-way interaction were also predictive of the subsequent drinking.
Feature selection
The feature selection process is similar to that used previously in connectome-based predictive modeling (Shen et al., 2017). On every leave-one-out (LOO) fold, one participant’s data (from both Cortisol and Placebo sessions) was omitted, creating a separate set of training data. Within the training data, beta values from each cluster were separately entered into Equations 1 or 2 above. For Equation 1, if the cluster resulted in a two-way interaction (beta * Pill) that was statistically significant (p < .05), that cluster was selected. For Equation 2, if the cluster resulted in a three-way interaction (beta * AUDIT * Pill) that was statistically significant (p < .05), that cluster was selected. If no clusters met this criteria, the top X most predictive clusters were selected until the largest drop in predictive power (see Goldfarb et al., 2020a). The features, or clusters, that were selected on every LOO fold were taken as the clusters that were predictive of cortisol-induced changes in drinking (Eq. 1) for risky drinkers (Eq. 2).
Cross-validation
To determine whether these clusters were significantly related to behavior across participants, we computed the average beta value across all selected features and determined parameters to fit Equation 1 or 2. We then took the left-out subject’s data, computed their mean beta value across the same features, and entered their beta, Pill (Eq. 1) and AUDIT (Eq. 2) with parameters determined by the training set, and recorded the estimated BAC level. After repeating this LOO process for all participants, we ran Pearson correlations relating these predicted to observed BAC values.
Statistical significance was determined by randomly permuting the input data 1000x, re-running the predictive modeling procedure, and counting the number of instances when the random data yielded a predicted × observed correlation above what was obtained from the true data.
Results
Hydrocortisone increases salivary cortisol
Administration of 20 mg hydrocortisone led to significant and sustained elevations in salivary cortisol throughout encoding (Fig. 1B; main effect pill:
Cortisol at encoding amplifies salient memory with consequences for alcohol motivation
We first assessed whether cortisol at encoding promoted memory for salient events. In an earlier study with a similar design, we found that patients with AUD had stronger memory for alcohol/scene pairs that they considered to be highly salient (Goldfarb et al., 2020b). Using the same approach, we incorporated trial-level ratings of valence, arousal, and craving to identify alcohol events that were personally salient for each individual participant.
Consistent with our hypothesis, we found that cortisol at encoding had distinct effects on salient alcohol memories (pill × cue ratings:
What does this memory bias mean for later behavior? We examined whether memory biases (i.e., stronger memory for personally salient vs other types of alcohol) were related to alcohol motivation, as measured immediately following memory retrieval in an alcohol taste test (Fig. 1A). Indeed, cortisol-induced memory biases had behavioral consequences: participants who had stronger memory biases also went on to show greater alcohol motivation overall, achieving a higher BAC (main effect memory bias on BAC:
Susceptible individuals show stronger cortisol-induced increases in alcohol motivation and brain salience responses
As cortisol effects vary with alcohol history (Vendruscolo et al., 2012; Repunte-Canonigo et al., 2015; Vendruscolo et al., 2015; Savarese et al., 2020), we recruited healthy adults showing a range of drinking behaviors (AUDIT scores shown in Fig. 2 inset; note that AUDIT scores were also significantly correlated with other measures of alcohol consumption, including average drinks per week [r = .76, p < .001] and number of drinking days in the past month [r = .48, p = .01]). We recruited this diverse sample in order to test how drinking severity modulated cortisol effects on salient encoding and downstream behavior. We did not find significant effects of AUDIT on memory biases (all p > .5) or the relationship between memory biases and drinking (all p > .8), with all results reported above remaining statistically significant even when accounting for AUDIT. These findings indicate that cortisol-modulated memory effects were not specific to risky drinkers.
Notably, riskier drinkers showed distinct effects of cortisol on later alcohol motivation. Participants with higher AUDIT scores showed stronger alcohol motivation after retrieving memories encoded under cortisol (Fig. 2A; AUDIT x Pill–mL:
As cortisol had stronger downstream consequences for susceptible individuals, we asked whether these individuals also experienced distinct cortisol effects on neural responses during encoding. To test this, we performed a whole-brain analysis modeling trial-evoked brain responses based on pill (hydrocortisone v placebo), which images they were encoding (alcohol v neutral objects), and AUDIT scores. This analysis revealed widespread increases in brain responses to cortisol, including in visual, sensorimotor, and cerebellar regions (Fig. 3A), and increased responses to alcohol-related events in ventromedial prefrontal cortex (vmPFC), posterior cingulate cortex, and amygdala (Fig. 3B). Critically, we identified a network that showed distinct hydrocortisone responses based on AUDIT scores when encoding alcohol-related images (three-way interaction; Fig. 3C, also 2A; all clusters listed in Table 1). These included regions associated with the salience network, including insula, putamen, amygdala, anterior cingulate cortex (ACC), midbrain, and inferior temporal gyrus (Seeley et al., 2007; Schwabe et al., 2022) as well as those associated with default mode network, including hippocampus and medial prefrontal cortex (Schwabe et al., 2022). Most identified regions showed that riskier drinkers had larger responses to alcohol images under cortisol (illustration in Fig. 2B). Going forward, we refer to the clusters identified through this three-way interaction as the SALcort network (Susceptible individuals’ encoding of Alcohol under Cortisol) for parsimony.
Brain networks responsive to cortisol, alcohol, and drinking susceptibility. Each panel shows Sørensen-Dice similarity of identified regions from linear mixed-effects model to previously defined intrinsic connectivity networks Laird et al., 2011; bar plots) and visualization of whole-brain results on slices. All results were voxelwise FDR-corrected p < .001. Panels portray main effects of pill type A, stimulus type B, and three-way interaction between pill, image type, and AUDIT scores C.
Regions showing distinct responses based on a three-way interaction of Pill (Cortisol v Placebo), Stimulus (Alcohol v Object), and AUDIT scores
To characterize these networks, we computed voxel overlap between regions sensitive to cortisol, alcohol, and those identified through the three-way interaction as part of SALcort, with a previously defined set of intrinsic connectivity networks Laird et al., 2011; Fig. 3). Brain regions enhanced by cortisol overlapped most with motor/visuospatial networks, whereas those enhanced when encoding alcohol events overlapped with default mode network as well as networks associated with interoception (gustation) and emotion. The SALcort network overlapped most with networks associated with emotion and interoception, including traditional “salience” networks (for which classic seed regions are those included in ICN 4, with additional regions in ICN 1, 3, and 5). That is, participants with higher AUDIT scores showed greater cortisol-induced increases in responses within brain salience networks.
SALcort network tracks subjective salience and later memory
The SALcort reflects the specific effects of cortisol on alcohol processing for those with greater history of alcohol intake. To further validate that the SALcort network was associated with salience, we assessed whether responses were associated with the subjective experience of salience reported by our participants. We found that the SALcort network was indeed sensitive to salience, with stronger responses to alcohol events for participants experiencing higher arousal and craving. For example, individuals who reported higher alcohol craving also had larger SALcort responses to alcohol stimuli under cortisol (including AUDIT and session as covariates; pill × craving interaction
Sensitized neural salience tracks subjective salience and memory. Greater activation of SALcort (Fig. 2B) under cortisol is associated with higher craving (alcohol only; A) and stronger biases toward remembering salient alcohol contexts the next day B. Shaded areas = 95% CI. Each dot = mean behavior and mean trial-evoked response per participant.
Given this sensitivity to subjective salience, we asked whether SALcort network responses were associated with an increased tendency to remember salient alcohol events. We modeled memory biases (salient alcohol—other alcohol) as a function of brain responses during encoding, pill, AUDIT, and session. We found that participants who had stronger SALcort responses under cortisol were more likely to remember salient alcohol the next day (brain response × pill:
SALcort network predicts subsequent alcohol motivation
The above results uncovered a salience-related brain network that was amplified by cortisol exposure in more susceptible individuals. We also found that these individuals also showed stronger downstream alcohol motivation. However, it was unclear whether these two effects were related: specifically, whether amplifying salient encoding could explain these changes in later behavior. To test this possibility, we took an unbiased data-driven approach by (1) extracting responses to alcohol events per cluster in our SALcort network and (2) using leave-one-subject-out cross-validation to determine whether these responses could significantly predict later behavior (similar to Goldfarb et al., 2020a; Fig. 5A). Such approaches typically explain behavior as linearly related to brain responses (e.g.,
Sensitized neural salience under cortisol predicts behavioral consequences. A, Illustration of predictive modeling analysis, assessing which SALcort clusters (green, center) could be used to predict consequences for alcohol motivation overall (left) or particularly in more susceptible individuals (right). (B–C) Results of predictive modeling analyses. Models used cross-validation to identify clusters for which activity would differentially predict alcohol motivation the day after hydrocortisone vs placebo B, and clusters for which activity would differentially predict alcohol motivation depending on both pill type and participant susceptibility (C). Both panels show clusters identified on every leave-one-subject-out fold (top); predictive success of model (middle, with inset density plots showing distribution of predictions from randomly shuffled data with true predictions shown as vertical lines); and visualization of relationship between cluster responses and the subsequent alcohol motivation (bottom). AUDIT visualized categorically in C, but was treated continuously in analyses. Shaded areas = 95% CI. *p < .05 based on permutation tests.
Discussion
Emotionally salient experiences are strongly encoded and remembered. Although cross-species evidence indicates that agents like cortisol can potentiate this process, the functional consequences of amplifying salient encoding are unclear. Here, we investigated the consequences of cortisol for salience encoding, later memory for salient experiences, and the subsequent behavior. As anticipated, we found that oral hydrocortisone at encoding led to stronger memory for salient alcohol-related events. Strikingly, this memory bias was associated with later alcohol motivation. We also identified particular consequences of cortisol for more susceptible individuals (that is, those who engaged in risky drinking outside the laboratory and have been shown to be more sensitive to cortisol in other contexts). For these individuals, cortisol potentiated neural salience encoding and alcohol motivation. Our machine learning analyses revealed a link between these effects, with cortisol-induced enhancements in salience network responses at encoding significantly predicting cortisol-induced increases in motivation. Together, these results highlight novel downsides to forming strong memories.
Why would it be maladaptive to encode memories more strongly? There are many examples from the addiction literature that patients with substance use disorders attend more and show distinct salience-related brain responses to drug-relevant cues (Jasinska et al., 2014; Koban et al., 2023; Addiction Cue-Reactivity Initiative (ACRI) Network et al., 2024), and increasing evidence that they also form stronger memories for drug-related events (Franken et al., 2003; Klein et al., 2013; Fridrici et al., 2014; Goldfarb et al., 2020b). This is particularly notable given robust findings that chronic drug use impairs memory (e.g., Manning et al., 2017)—although it is important to highlight that these findings are often based on neuropsychological batteries that probe memory for neutral, not drug-relevant, information. Having a bias toward remembering drug-related experiences (perhaps at the expense of alternative events) may then result in increased motivation to consume drugs (Goldfarb and Sinha, 2018; Bornstein and Pickard, 2020).
Strikingly, we were able to recapitulate the memory alterations that we previously observed in AUD (Goldfarb et al., 2020b) by administering hydrocortisone to healthy social drinkers. Cortisol has been shown to mimic other aspects of encoding that are enhanced in populations with substance use disorders, including amplifying responses to relevant cues (Zorawski and Killcross, 2003; Pool et al., 2015), facilitating goal-directed approach behavior (Putman and Roelofs, 2011), and promoting consolidation of salient conditioned associations (Miranda et al., 2008; Wichmann et al., 2012). These findings are consistent with the possibility that higher levels of basal cortisol in AUD, as well as acute alcohol-induced cortisol increases (Wemm and Sinha, 2019), may play key roles in explaining why alcohol-related memories are formed differently within riskier populations (Goldfarb and Sinha, 2018). Further work is needed to determine the causal role of this system in potentiating memory within populations with substance use disorders. It is also worth noting that the type of memory that was enhanced—namely, the integration between different elements of an event—is one that is typically impaired when one of the elements is emotionally salient (Bisby et al., 2016; Bisby and Burgess, 2017). We replicate this typical pattern under placebo conditions. Critically, cortisol rescued memory integration for salient events relative to placebo. Even within the cortisol condition alone, we found that stronger salience was associated with enhanced memory integration. Our design precludes the ability to assess whether this benefit would extend to other (non-alcohol-related) salient stimuli. Nevertheless, it is notable that this finding is consistent with past work showing that acute stress and cortisol promoted memory integration for other salient memoranda (van Ast et al., 2013, 2014; Goldfarb et al., 2019) and may be related to interactions between glucocorticoids and adrenergic responses to salient events (Roozendaal et al., 2006; Segal et al., 2014). This ability of cortisol to induce a memory flip, enhancing rather than impairing integration for salient events, is consistent with emerging findings from the stress and alcohol fields and underscores the need for further work evaluating the organization of emotional memories (Loetscher and Goldfarb, 2024).
In addition to cross-species studies showing that cortisol can promote formation of emotional memories, there is also reason to anticipate that cortisol would change later behavior—particularly in more susceptible individuals. For example, cortisol can increase alcohol palatability (Söderpalm and Hansen, 1999) and intake (Fahlke et al., 1996), particularly for rodents who already preferred alcohol (Fahlke et al., 1994). On the other hand, blockade of this system decreased alcohol intake, specifically for animals who already had a tendency to drink more (Savarese et al., 2020) or who had chronically been exposed to ethanol vapor, known to induce dependent-like drinking patterns (Vendruscolo et al., 2012; Repunte-Canonigo et al., 2015). Previous human studies blocking the glucocorticoid system have had mixed effects on drinking behavior (Vendruscolo et al., 2015), perhaps due to the severity of drinking or treatment-seeking status of their population (Haass-Koffler et al., 2023). Our finding that cortisol-induced increases in alcohol motivation were stronger among riskier drinkers is consistent with these preclinical findings and provides several key extensions. First, our effects were at a much longer timescale: cortisol increased alcohol motivation 24 h later (well past the half-life of this dosage; Roger et al., 1982), highlighting the possibility that altered memory encoding can extend the opportunities for cortisol to potentiate behavior. Second, we could directly link these behavioral consequences to amplified neural salience encoding via machine learning.
Using a whole-brain approach, we identified regions for which cortisol altered neural responses during encoding. Cortisol led to widespread elevations in sensorimotor networks during encoding, consistent with past findings (Buades-Rotger et al., 2016; Harrewijn et al., 2020) and the idea that cortisol enhances processing of task-relevant content (Putman and Roelofs, 2011). Critically, we also identified a set of regions for which more susceptible individuals (who also went on to show higher alcohol motivation the next day) showed cortisol-induced increases during alcohol encoding. Based on voxelwise comparisons to extant networks (Laird et al., 2011) as well as relationships to subjective ratings and later salient memory biases, we determined that these regions were involved in salience processing. Notably, the same dose of cortisol administered in this study has been shown to modulate salience network engagement in clinical populations (Soravia et al., 2018). Showing that brain salience responses to alcohol images are exacerbated by cortisol, and go on to predict later substance use within a non-clinical population, extends the central role of salience in maladaptive substance use (Kwako et al., 2019; Padula et al., 2022) and accords with recent findings that risky drinking can alter the structure and function of salience and emotion-related regions (Pérez-García et al., 2022). These findings also suggest that cortisol may alter the function of salience networks and contribute to the overattribution of salience to drug-related content (Goldstein and Volkow, 2011). Our machine learning approach enabled us to pinpoint the parts of this network that explained consequences for later behavior (see also Goldfarb et al., 2020a). We found distinct clusters that predicted cortisol-induced increases for all participants and those that were specific to susceptible individuals. This susceptible set included key salience-related nodes like insula, striatum, ACC, and OFC (Seeley et al., 2007; Uddin, 201). These regions have previously been implicated in AUD (e.g., stress increased insula responses to alcohol cues, which tracked craving and the subsequent alcohol consumption; Bach et al., 2023), maladaptive drinking patterns (e.g., higher ACC and insular responses to uncertain threats predicting greater risk for subsequent binge drinking; Gorka et al., 2023, and binge drinkers showed enhanced connectivity within a circuit including striatum, ACC, and OFC; Arienzo et al., 2020), and alcohol seeking (associated with striatum; Hyman, 2005; Chen et al., 2011; Corbit et al., 2012; Everitt and Robbins, 2016; Radoman et al., 2021). Notably, these regions are also implicated in salient encoding. For example, the striatum is known to play an important role in reward processing (Delgado, 2007; Izuma et al., 2008; Lega et al., 2011), and insula has been shown to contribute to long-term retention of drug context memory in rodents (Contreras et al., 2012). Together, our findings highlight ways in which neural salience encoding, perhaps a latent response in these healthy individuals who are at higher risk for AUD, may be amplified by cortisol with deleterious consequences for later behavior.
In summary, the present results uncover a downside to neural salience encoding. Salient or emotional events are remembered more strongly in part so that they can guide future behavior (McGaugh, 2015). Here, boosting neural salience encoding with cortisol, while promoting lasting representations of salient events, also led to maladaptive motivation to consume alcohol. These findings underscore the importance of cortisol modulating the encoding process itself by demonstrating that neural alterations during encoding were predictive of these downstream behaviors. The mechanism explored here also has public health relevance: alcohol consumption is highly prevalent, with 85% of US adults having tried alcohol (SAMHSA, 2021), and tendencies to engage in risky patterns of drinking have steadily increased (Grant et al., 2017; Patel and Balasanova, 2021). Strengthening salient alcohol encoding (and potentially making alternative neutral memories less available; see Bornstein and Pickard, 2020) may be one underexplored pathway by which cortisol drives these risky behaviors. More broadly, these findings highlight the far-reaching implications of the way we form and retain our emotional memories.
Footnotes
This work was supported by a Yale Center for Clinical Investigation Junior Faculty Scholar Award (NIH/NCATS KL2 TR001862 to E.V.G.) and the National Institutes of Health (K01 AA027832 to E.V.G.). We thank Nia Fogelman for assisting with randomization and maintaining the double-blind; and Gretchen Hermes for performing physical exams.
The authors declare no competing financial interests.
- Correspondence should be addressed to Elizabeth V. Goldfarb at elizabeth.goldfarb{at}yale.edu.
