“Expectancy-Mood Neural Dynamics Predict Mechanisms of Short- and Long-Term Antidepressant Placebo Effects”

Abstract

Objective:

Acute experimental models of antidepressant placebo effects suggest that expectancies, encoded within the salience network (SN), are reinforced by sensory evidence and mood fluctuations. However, whether these dynamics extend to longer timescales remains unknown. To answer this question, we investigated how SN and default mode network (DMN) functional connectivity during the processing of antidepressant expectancies facilitates the shift from salience attribution to contextual cues in the SN to belief-induced mood responses in the DMN, both acutely and long-term.

Methods:

Sixty psychotropic-free patients with major depressive disorder (MDD) completed an acute antidepressant placebo fMRI experiment manipulating placebo-associated expectancies and their reinforcement while assessing trial-by-trial mood improvement, before entering an 8-week double-blind, randomized, placebo-controlled trial (RCT) of a selective serotonin reuptake inhibitor (SSRI) or placebo.

Results:

Learned antidepressant expectancies predicted by a reinforcement learning model modulated SN-DMN connectivity. Acutely, greater modulation predicted higher effects of expectancy and reinforcement manipulations on reported expectancies and mood. Over 8 weeks, no significant drug effects on mood improvement were observed. However, participants who believed they were receiving an antidepressant exhibited significantly greater mood improvement, irrespective of the actual treatment received. Moreover, increased SN-DMN connectivity predicted mood improvement, especially in placebo-treated participants who believed they received an SSRI.

Conclusion:

SN-DMN interactions may play a critical role in the evolution of antidepressant response expectancies, drug-assignment beliefs, and their effects on mood.

INTRODUCTION

Placebo responses elicited by antidepressants frequently overshadow their pharmacodynamic effects, persisting for weeks without any evident physiological reinforcement. Nonetheless, the mechanisms driving these long-lasting placebo effects over time remain elusive.

Classical theories of the placebo effect, primarily derived from acute placebo analgesia experiments, postulate dual expectancy and conditioning mechanisms (1). While the former theory attributes how placebo effects to positive expectancies, typically elicited by verbal instructions, conditioning theories attribute placebo responses to the learned association between a conditioned (e.g., a placebo pill) and an unconditioned stimulus (e.g., an active drug), leading to a conditioned response even in the absence of an active drug. More recently, reinforcement learning (RL) frameworks (2) have provided a computational account of these effects, wherein expectancy learning starts with a prior (e.g., from verbal instructions) and depends not merely on the contiguity between the conditioned and unconditioned stimuli, but on reward prediction errors (PE)– the discrepancy between what it is expected and what it is experienced. These PE are key to understanding how humans learn from outcomes to guide their choices and have been used to predict placebo analgesia. In depression, our recent study using RL computational theories suggests that model-predicted expectancies—encoded in the salience network (SN)—trigger mood alterations perceived as reward signals (3). These mood reward signals subsequently reinforced antidepressant expectancies through the formation of expectancy-mood loop dynamics (3). However, whether RL mechanisms extend beyond short-term manipulations to influence long-term placebo effects, as those frequently observed in clinical trials, remains unexplored.

Persistent internal cognitive processes (e.g., long-term memory, self-referential thought), facilitated by the default-mode network (DMN) (4,5), are known to influence mood over extended periods, even when removed from perceptual interface (6). From a conceptual standpoint, if learned placebo expectancies encoded within the SN are to exert influence on mood both short-term and over periods spanning days or weeks, they must undergo a transformation into internal representations akin to those characteristics of the DMN, specifically within its anterior hub. Several lines of evidence indirectly support this proposition. First, SN-DMN resting state functional connectivity (rsFC) is linked to antidepressant drug and placebo effects, persisting over 10 weeks of treatment (7,8). Secondly, the activity within the rostral anterior cingulate cortex (ACC), a key node of the anterior DMN, consistently emerges as a robust predictor of mood response to both active antidepressants (9–13) and placebos (7,13), implying its involvement in nonspecific treatment response mechanisms. Third, the rACC has been reliably involved in placebo effects across different clinical conditions (7,14,15), consistent with its role in prediction, prospection, and valuation (16). Collectively, these findings suggest that the processing of placebo expectancies and their associated mood responses derive from SN to DMN connectivity, particularly its anterior hub, and that the strength of these connections plays a pivotal role in the formation of long-term effects of antidepressant placebos.

More broadly, interactions between the SN and DMN, defined here by the Yeo 7-networks 2011 atlas (17), are thought to be critical for cognitive control, including how stimuli are evaluated, responded to, and experienced emotionally. While the SN responds to external events that are behaviorally salient (18) (e.g., treatment cues), the DMN shows high activity when subjects have an internal focus of attention (e.g., internally directed thought). Although initially identified during resting-state (19), recent evidence suggests that the DMN is consistently activated during tasks involving affective, social, and introspective processes (4,5). Furthermore, key regions of the DMN are prominently activated when individuals are asked to self-generate both positive and negative emotional responses by recalling or imagining events and situations. In particular, self-generated positive emotions activate key nodes of the DMN, including the ventromedial prefrontal cortex (vmPFC), and connected subcortical reward areas such as the ventral striatum (16).

To further elucidate the role of SN-DMN coupling in antidepressant placebo effects, we build on previous research on rsFC (7,13), by examining SN and DMN dynamics using a generalized psychophysiological interaction (gPPI) analysis of task-based FC (tbFC). Specifically, we used these networks, as defined by the Yeo et al. 7-network atlas (17), as anatomical masks to estimate how their connectivity responds to experimental manipulations—namely, the placebo task in our fMRI paradigm. We then analyzed FC indices between the SN and the anterior DMN to assess their relationship with 1) acute expectancy-mood dynamics, and 2) subsequent drug-assignment beliefs and mood trajectories in an 8-week double-blind, randomized, placebo-controlled trial (RCT). We hypothesized that a more pronounced modulation of SN-DMN connectivity by acute learned placebo expectancies would not only (i) scale with concurrent acute antidepressant placebo response but also (ii) predict placebo response over time.

METHODS

Participants and study design

As previously reported (3), we recruited 60 participants diagnosed with Major Depressive Disorder (MDD) from July 2017 to January 2021 through the University of Pittsburgh’s recruitment website (https://pittplusme.org). Participants self-identified with the following race categories: 31 Caucasian/White (52%), 14 African/Black (23%), 10 Asian (17%), 5 Other or Mixed (8%). In addition, 3 participants reported being Hispanic. Enrolled participants were unmedicated, right-handed, fluent in English, and provided written and signed informed consent, approved by the University of Pittsburgh Institutional Review Board. The exclusion criteria and the use of authorized deception – common in placebo research (20) –are described in the Supplemental Methods.

All research participants had a diagnosis of non-psychotic MDD with or without anxiety disorders as diagnosed by the Mini-International Neuropsychiatric Interview (M.I.N.I) (21). Participants had at least moderate depression, as determined by a Hamilton Depression Rating Scale (HDRS-21) (22) score of ≥ 16 at screening. All participants were antidepressant medication-free for at least 21 days prior to the collection of imaging data (five weeks for fluoxetine). Only one participant had received an antidepressant during the current episode, per the MGH Antidepressant Treatment Responses Questionnaire (23), but discontinued before study enrollment.

The Antidepressant Placebo fMRI and RCT

Enrolled participants completed the Antidepressant Placebo fMRI experiment (Fig. 1A), which includes a deceptive narrative involving the experimental manipulation of two components of the placebo effect: expectancies and their reinforcement, each followed by the rating of the subjects’ expectancies and mood responses. Briefly, the deceptive narrative describes an experimental manipulation aimed at investigating the brain effects of a “fast-acting antidepressant” compared to a “conventional antidepressant” while recording “participants’ brain activity” and providing neurofeedback. Participants are told that, following each infusion, they will be presented with neurofeedback of positive or baseline signs reflecting acute mood changes in response to the infusions. Participants are also informed that in addition to the infusion periods, there would be periods of “equipment calibration” – the study control condition – where no drug would be administered, although the neurofeedback signal would still be recorded and displayed on the monitor (Figure 1). The experimental expectancy condition involves an “antidepressant infusion” cue and a “no-infusion” cue (“for equipment calibration”). During each “antidepressant infusion” cue (4s), a bar is filled at four 1s-periods representing 0%, 33%, 66%, and 100% of the dose administered. During the “calibration no-infusion” cue (4s) the bar remains empty. For the reinforcement condition (10s), sham neurofeedback acts as a secondary reinforcer of the “antidepressant” effects. In the high-reinforcement condition, sham neurofeedback is positive 75% of the trials (vs. 25% baseline), whereas the opposite pattern is observed in the low-reinforcement condition. Participants rate their expected and actual change in mood (YES/NO) in response to each infusion/neurofeedback signal, respectively, by using a keypad and their index fingers (primary outcome measures for acute placebo effects). Inter-stimulus interval duration was randomly sampled from an exponential distribution bounded between 0.33 and 2 s. Inter-trial intervals were sampled from the uniform distribution with bounds 4-6 s. Each of the four runs included 32 trials. Task credibility and debriefing procedures are described in the Supplemental Methods.

Figure 1:

A) Antidepressant Placebo fMRI Experiment. B) Study Design: Sixty psychotropic-free patients with major depressive disorder (MDD) completed a session of the antidepressant placebo fMRI experiment before (N=60) during an 8-week double-blind randomized controlled trial of SSRI or placebo. Participants received 10 mg of escitalopram or placebo for the first week and increased to 20 mg at week 2 for the remaining 8-week period (except for 4 subjects with significant improvement at 10mg). One subject returned to 10 mg after week 6. Three participants received fluoxetine instead of escitalopram due to a previous lack of response to escitalopram. C) CONSORT Diagram.

After completing baseline procedures and a functional MRI session during the Antidepressant Placebo fMRI experiment, participants were enrolled in an 8-week double-blind, randomized placebo-controlled trial and were assigned to receive either a selective serotonin reuptake inhibitor (SSRI) (93.5 % escitalopram) or a placebo pill (Fig.1B). The patient and the study team were blinded to the drug assignment and the randomization was generated by an unblinded independent pharmacist using an in-house MATLAB software to assign participants (See CONSORT, Fig. 1C).

Assessments

In addition to completing the baseline HDRS-21, participants completed the Montgomery-Asberg Depression Rating Scale (MADRS) (24), and the self-reported Quick Inventory of Depressive Symptomatology (QIDS-SR-16) (25). Both, the QIDS-SR-16 and the MADRS were administered at each study visit (at baseline and weeks 1, 2, 3, 4, 6, and 8, Fig. 1B). As in our prior study of long-term neuroimaging predictors of antidepressant placebo effects (26), the QIDS-SR16 was the primary outcome of long-term placebo effects. This self-report measure was deemed to contain more relevant information about subjective placebo responses than the clinician-administered MADRS. Results using the MADRS as a secondary outcome are reported in Supplemental Table 2. Baseline HDRS-21 scores served only as an entry criterion, facilitating comparisons with older studies and mitigating regression to the mean of QIDS-SR-16 scores.

At every visit post-randomization, participants were asked: “Do you believe that you are currently on the drug or the placebo?”. In addition, they reported the certainty of their drug-assignment belief by answering the question: “On a scale of 0-100, how certain are you about your drug-assignment belief?”.

Summary of previous relevant findings

In a previous cross-sectional analysis of baseline behavioral and fMRI data of this same sample (3), we tested alternative computational RL models of acute antidepressant placebo effects. Briefly, in our basic model, model-predicted expectancies for each of the four trial conditions were updated for each presentation of the “antidepressant” versus “calibration” infusion cue and “positive” versus “baseline” neurofeedback outcome, based on the basic delta learning rule:

$${\text{Q}}_{\text{t+1}}(\text{s})=Q_{\text{t}}(\text{s})+\alpha ({\text{δ}}_{\text{t}})$$

where ${\text{Q}}_{\text{t}}(\text{s})$ is the model-predicted expectancy for stimulus s at trial $t,\alpha$ is a learning rate, and ${\delta }_{\text{t}}$ is the difference between the actual and expected outcome (prediction error, PE) at trial t. The prediction error ($\delta$) was calculated with:

$${\delta }_{\text{t}}={\tau }_{\text{t}}-{\text{Q}}_{\text{t}}(\text{s})$$

where, ${\tau }_{\text{t}}$ is the reinforcement (positive or baseline sham neurofeedback). The sigmoid choice rule included two free parameters $\beta$ (stochasticity) and K (choice bias).

Our dominant model tested the possibility that biased learning and self-reinforcement contributed to antidepressant expectancies. This model included 2 learning rates, depending on whether participants are updating expectancies for the placebo ${\alpha }_{\text{plac}}$ or calibration cue ${\alpha }_{\text{cal}}$, and an augmented reward if mood was rated as improved ${\text{(u}}_{\text{t}}=\text{r}+{\text{mood}}_{\text{t}})$. Models were inverted based on participants’ expectancy and mood ratings using a Bayesian procedure implemented in the Variational Bayes Approach (VBA) toolbox (27) MATLAB (version R2021b, MathWorks). Model comparisons employed a Bayesian model selection algorithm(28).

The superiority of this model [after correction for Bayesian omnibus risk (<0.001) and protected exceedance probability = 97%], compared to all others tested, indicated that expectancy-mood loop dynamics contribute to the evolution of acute antidepressant expectancies and mood responses, providing a mechanism through which expectancies evolve over time.

To further understand the neural dynamics underlying this learning process, we mapped trial-level learned placebo expectancies resulting from the dominant model to whole-brain neural networks. Higher model-predicted expectancies elicited enhanced SN encoding, demonstrating the central role of the SN in the formation of learned antidepressant expectancies. Acutely, SN responses to learned expectancies interacted positively with the high-reinforcement condition to predict expectancy ratings and scaled positively with depression severity (3).

Seed-Based Functional Connectivity (FC) Analysis

MRI data acquisition and processing are described in the Supplemental Methods. To examine the SN FC during expectancy processing we used a generalized psychophysiological interaction (gPPI) analysis (29). The original RL model-based analysis included four event regressors: infusion, expectancy rating, neurofeedback, and mood rating events, and two continuous parametric modulators: model-predicted expectancy and a mood reward signal obtained from the dominant model. For the gPPI analysis, we evaluated interactions between the SN time-course averaged across the SN mask and the model-predicted expectancy parametric modulator aligned to the infusion event. Group-level voxel-wise FC maps were generated using FSL randomise (non-parametric one-sample t-test) (30) with Threshold Free Cluster Enhancement (TFCE) (1- P > 0.95) (31).

As previously reported (3), the anatomical localization of the SN seed was defined as the SN network from Yeo et al. 7-network atlas (17).

To describe the whole-brain pattern of SN connectivity, our voxel-wise connectivity analysis was referenced against the Yeo’s 7-network and Schaefer’s 400 parcellation atlas (32) supplemented with subcortical structures for anatomical location (Supplemental Methods and Supplemental Fig. 1).

Individual regression coefficients (“betas”) from the voxel-wise connectivity map that overlayed with the anterior DMN (41 parcels), as defined by Yeo’s 7-network and Schaefer’s 400 parcellation atlas (32), were averaged across voxels and extracted for brain-to-behavior analyses as a continuous subject-level moderator, excluding parcels with <50% coverage to avoid overlap (Supplemental Figure 1).

Statistical Analysis

We used multi-level logistic regression for predicting expectancy and mood ratings during the Antidepressant Placebo fMRI experiment using the lme4 (33) R (version 4.2.2) package. The model estimated the fixed effects and interactions of experiment conditions (expectancy and reinforcement trial-level predictors coded as 0 and 1) with SN-learned expectancy-related activity or SN-DMN functional connectivity (continuous subject-level predictors). To avoid any potential circularity, Q values from the RL model were not included in the brain-to-behavior analysis. Subject and run (clustering within-participant) intercepts were taken to be random. Significant predictors were identified using the likelihood ratio test (LRT; car:Anova) (34), after refitting the model using full ML (instead of REML).

In our analysis of longitudinal data, we used a multi-level logistic regression model to predict drug-assignment beliefs and linear regression to predict depression severity throughout the RCT. Predictors of drug-assignment beliefs included the drug-assignment group (SSRI vs. placebo coded as 0 and 1) and time. Predictors of depression severity included the drug-assignment group (SSRI vs. placebo coded as 0 and 1) OR the drug-assignment belief (drug vs. placebo belief coded as 0 and 1), and time. We compared models with linear, quadratic, negative inverse-transformed, and completely general (unordered) effects of time, and found no significant differences in model fit. Models reported here use negative inverse-transformed time, which affords the best visual fit.

For the purpose of graphing mixed-effect model interactions, the interaction terms were extracted using the effects package in R (35), which computes estimated marginal means and standard errors from the mixed-effects model. These values were then plotted using ggplot2. For the plotting of the continuous SN-DMN FC in Fig. 2B and 2C, the effects function splits the variable into five intervals, and the 2^nd and 4^th quintiles were plotted. For the plotting of the continuous SN-DMN FC in Fig. 3C and 3D, high and low SN-DMN FC were defined as two standard deviations above or below the sample median, respectively. All models were controlled for baseline depression severity and included a subject-level random intercept.

Figure 2: SN seed-based Psychophysiological interaction (PPI) analysis and brain-to-behavior analysis:

A) RL-informed SN-seed connectivity analysis during the processing of learned antidepressant expectancies revealed increased connectivity with key nodes of the Default Mode Network (DMN), among other regions. SN-DMN connectivity moderated the effect of expectancy and reinforcement conditions (B) expectancy and (C) mood ratings, such that high connectivity was associated with significantly higher effects of the experiment conditions on expectancy and mood ratings. For the plotting of the continuous SN-DMN FC variable, only the 2nd and 4th quintiles were plotted ± the standard errors (SE). Abbrev.: SN, salience network; DMN: Default Mode Network; NF: Neurofeedback; RL: Reinforcement Learning.

Figure 3: Longitudinal predictors of Antidepressant Placebo Effects.

Panels A-D display estimated marginal means with standard errors (SE) for the significant interaction terms derived from the mixed-effects model predicting QIDS-16SR scores over 8 weeks (Model 2B, 3B). Panel A shows significant reductions in depression severity over time across all participants, independent of drug-assignment group (Placebo vs. SSRI). Panel B highlights a significant effect of drug-assignment belief (Placebo Belief vs. SSRI Belief) on depression severity over time. Panel C compares depression severity for participants with high versus low SN-DMN functional connectivity (FC), defined as two standard deviations above or below the sample median FC. Panel D further explores the interaction between drug-assignment belief (Placebo Belief vs. SSRI Belief) and SN-DMN FC (High vs. Low). Abbrev.: SN, salience network; DMN: Default Mode Network; FC: Functional Connectivity.

RESULTS

Sample and participant flow

Our sample comprised 60 psychotropic-free participants with MDD (mean age: 24.5 [SD=6.0], 85% female). All had at least moderate depression, as indicated by HDRS-21 score ≥16 (mean = 21.55, S.D.=4.2). Before randomization, 6/60 participants were lost to follow-up.

Of the 54 randomized participants, 25 received an SSRI, with 22 completing the study, and 29 participants received a placebo, with 24 completing the study (CONSORT diagram, Fig. 1C). Baseline depression severity did not differ between the SSRI and placebo arms (Table 1). All participants endorsed the credibility of the placebo manipulations (Supplementary Fig. 2). One subject was excluded from analysis after randomization due to low-quality imaging data. 63 % of the participants had received at least one antidepressant, while 24% were classified as antidepressant naïve by self-report data at screening. Data on antidepressant history was missing in 6 participants.

Table 1:

Demographics.

Characteristic	Drug (N=25)		Placebo (N=29)		Analysis
	N	%	N	%	χ2	p
Gender					0.37	0.54
Male	5	20.00%	3	10.34%
Female	20	80.00%	26	89.66%
Race					3.96	0.14
White	14	56.00%	15	51.72%
African American	2	8.00%	8	27.59%
Other	9	36.00%	6	20.69%
Employment Status					2.31	0.32
Employed	13	52.00%	10	34.48%
Unemployed	12	48.00%	18	62.07%
Unknown	0	0.00%	1	3.45%
Correct Belief a					1.60	0.21
Correct	16	64.00%	12	41.38%
Incorrect	9	36.00%	16	55.17%
	Mean	SD	Mean	SD	t	p
Age (years)	25.44	6.33	23.45	5.65	1.21	0.23
Education (years)	14.96	2.17	13.90	1.90	1.90	0.06
HDRS score	21.36	4.14	21.45	4.51	−0.07	0.94
QIDS-SR-16 score	14.96	3.83	13.83	4.01	1.06	0.29
MADRS score	25.20	5.77	23.93	6.20	0.78	0.44

^aOne participant had no belief measurements

SN seed-based psychophysiological interaction (PPI) analysis

We first sought to ascertain whether SN-DMN connectivity was indeed modulated by evolving acute placebo expectancies. When RL model-predicted learned placebo expectancies were high, functional connectivity from the SN seed was increased throughout the key nodes of the DMN, including the rostral ACC and the posterior cingulate cortex, along with the thalamus and brainstem (Fig. 2, Supplemental Table 1).

Individual regression coefficients from the voxel-wise SN-DMN FC map extracted from the anterior DMN, a robust predictor of antidepressants (9–13) and placebo response (7,13), did not scale with depression severity at baseline (HDRS: r=0.22, p=0.109, MADRS: r=0.12, p=0.378, QIDS: r=0.19, p=0.178, Supplemental Results). Also, SN-DMN FC indices did not differ between subjects with prior antidepressant history and treatment naïve (mean diff.=0.002, t = 0.43, p = 0.668).

SN-DMN FC and acute antidepressant placebo effects (Model 1)

We then sought to understand whether expectancy-related SN-DMN connectivity predicted behavioral indices of acute placebo response using multi-level models predicting expectancy and mood ratings. SN-DMN FC positively moderated the effect of expectancy and reinforcement manipulations on expectancy ratings (χ²=6.27, b=18.82, S.E.= 7.51, z=2.51, p=0.012, Table 2, Model 1A, Fig. 2B) and mood ratings (χ²=5.55, b=14.11, S.E.= 5.98, z=2.35, p=0.018, Table 2, Model 1B, Fig. 2C), such that the effect of the antidepressant cue on expectancy and mood ratings during positive reinforcement was greater among subjects with high SN-DMN connectivity.

Table 2:

Mixed-effects models for the prediction of acute (expectancy and feedback ratings) and long-term (drug-assignment belief and mood responses, using QIDS-SR-16 as the outcome measure) antidepressant placebo effects, with SN-DMN FC during the processing of learned expectancies as a moderator. Results using the MADRS as an outcome measure were consistent, although less significant, suggesting that placebo effects might be more prominent when using self-report questionnaires (Supplemental Table 2).

ACUTE EFFECTS	A. Expectancy Ratings					B. Feedback Ratings
Model 1	χ2	Estimate	S.E.	z value	p	χ2	Estimate	S.E.	z value	p
Expectancy Condition	159.91	1.14	0.09	12.59	<0.001	6.35	0.23	0.09	2.52	0.012
Reinforcement Condition	6.84	0.24	0.09	2.62	0.009	236.35	1.40	0.09	15.38	0.000
SN-DMN Connectivity	0.03	2.62	15.67	0.17	0.867	17.95	−39.55	9.07	−4.36	0.000
Baseline QIDS-16SR	3.76	−0.12	0.06	−1.94	0.050	3.07	−0.10	0.06	−1.76	0.080
Expectancy C.*Reinforcement C.	9.04	0.38	0.12	3.01	0.003	3.08	0.22	0.13	1.77	0.079
Expectancy*SN-DMN	23.01	−27.56	5.75	−4.80	<0.001	4.65	−10.44	4.84	−2.16	0.030
Reinforcement*SN-DMN	0.07	−1.43	5.38	−0.27	0.790	0.03	−0.77	4.79	−0.16	0.871
Expectancy*Reinforcement*SN-DMN	6.27	18.82	7.51	2.51	0.012	5.55	14.11	5.98	2.35	0.018
Drug/Belief LONGITUDINAL EFFECTS	A. BELIEF					B. QIDS-SR16 SCORE
Model 2	χ2	Estimate	S.E.	z value	p	χ2	Estimate	S.E.	z value	p
Time	0.49	0.20	0.55	0.36	0.716	15.80	−2.19	0.57	−3.79	0.000
Drug	2.09	1.33	0.74	1.78	0.075	0.51	−1.04	1.05	−0.99	0.326
Belief						15.57	0.62	0.70	0.89	0.377
Baseline QIDS-16SR	0.01	0.01	0.09	0.11	0.910	10.95	0.41	0.12	3.31	0.002
Time*Drug	1.25	−0.82	0.74	−1.12	0.263	0.00	−0.14	0.74	−0.20	0.846
Time*Belief						0.21	1.99	0.79	2.50	0.013
SN-DMN LONGITUDINAL EFFECTS	A. BELIEF					B. QIDS-SR16 SCORE
Model 3	χ2	Estimate	S.E.	z value	p	χ2	Estimate	S.E.	z value	p
Time	0.37	−0.05	0.65	−0.08	0.940	19.25	−1.00	0.77	−1.30	0.195
SN-DMN Connectivity	3.95	−68.60	47.57	−1.44	0.149	1.39	−17.08	68.78	−0.25	0.805
Belief						15.74	1.21	0.82	1.47	0.143
Drug	1.16	1.07	0.80	1.33	0.185	1.26	−0.40	1.21	−0.33	0.743
Baseline QIDS-16SR	0.15	0.14	0.36	0.40	0.691	12.69	0.48	0.13	3.56	0.001
Time*SN-DMN	0.11	51.20	48.49	1.06	0.291	1.28	−145.74	55.64	−2.62	0.009
Time*Belief						3.68	0.46	1.27	0.36	0.716
Time*Drug	0.88	−0.44	0.86	−0.52	0.604	0.44	−1.00	1.13	−0.88	0.378
SN-DMN*Belief						4.76	−79.09	65.07	−1.22	0.225
SN-DMN*Drug	0.24	0.82	64.93	0.02	0.99	0.00	−91.87	103.85	−0.89	0.379
Belief*Drug						0.19	−0.65	1.10	−0.59	0.557
Time*SN-DMN*Belief						4.93	222.12	101.50	2.19	0.030
Time*SN-DMN*Drug	1.48	−87.42	71.79	−1.22	0.223	0.00	58.67	107.52	0.55	0.586
Time*Drug*Belief						0.06	0.82	1.70	0.48	0.631
SN-DMN*Belief*Drug						4.35	194.38	93.19	2.09	0.038
SN LONGITUDINAL EFFECTS	A. BELIEF					B. QIDS-SR16 SCORE
Model 4	χ2	Estimate	S.E.	z value	p	χ2	Estimate	S.E.	z value	p
Time	0.30	−0.52	0.73	−0.71	0.478	17.69	−2.58	0.78	−3.33	0.001
SN BOLD	0.40	−19.45	16.50	−1.18	0.239	0.05	12.40	20.53	0.60	0.548
Belief						14.05	0.59	0.86	0.69	0.489
Drug	1.07	0.52	1.01	0.51	0.607	1.05	−0.66	1.41	−0.47	0.640
Baseline QIDS-16SR	0.02	−0.06	0.38	−0.15	0.883	5.61	1.54	0.51	3.02	0.004
Time*SN	2.58	37.78	18.88	2.00	0.045	0.05	19.21	14.99	1.28	0.201
Time*Belief						5.61	1.96	1.28	1.53	0.127
Time*Drug	1.79	−0.35	1.06	−0.34	0.738	0.31	0.50	1.14	0.44	0.658
SN*Belief						0.19	−10.59	18.48	−0.57	0.567
SN*Drug	0.73	29.58	21.37	1.38	0.166	1.09	−17.35	27.71	−0.63	0.533
Belief*Drug						0.77	0.52	1.20	0.43	0.668
Time*SN*Belief						0.00	−0.75	20.33	−0.04	0.971
Time*SN*Drug	1.71	−30.37	23.25	−1.31	0.192	2.76	−30.97	18.64	−1.66	0.098
Time*Drug*Belief						0.00	−0.01	1.70	−0.01	0.995
SN*Belief*Drug						0.21	10.31	22.27	0.46	0.644

Abbrev.: SN, salience network; DMN: Default Mode Network.

Drug-assignment, drug-assignment belief, and depression severity (Model 2).

Models predicting participants’ drug-assignment beliefs revealed no significant effect of drug or time-by-drug interaction (Table 2, Model 2A). Similarly, models predicting depression severity found no significant effect of drug or time-by-drug interaction (Table 2, Model 2B, Figure 3A), indicating no differences between SSRI and placebo. Instead, we found a significant effect of drug assignment belief, such that participants who thought they were receiving an SSRI experienced greater depression improvement over time (belief*time: χ²=0.21, b=1.99, S.E.= 0.79, z=2.50, p=0.013, Table 2, Model 2B, Fig. 3B). Interestingly, participants guessed their assignment no better than expected by chance (correct guesses: 16/25 in the drug arm (64%) and 12/28 in the placebo arm (41%); χ²=2.4, p=0.123; missing one participant lost to follow-up after the first visit).

SN-DMN FC and long-term antidepressant placebo effects (Model 3).

We then examined whether expectancy-related SN-DMN FC predicted mood improvement over the 8-week trial and found a positive relationship (FC*time: χ²=1.28, B=−145.74, S.E.= 55.64, z=−2.62, p=0.009, Table 2, Model 3B, Fig. 3C). Interestingly, mood improvement was limited to participants who believed they were receiving an SSRI (time*FC*belief: χ²=4.93, b=222.12, S.E.= 101.5, z=2.19, p=0.030, Model 3B, Fig. 3D), particularly when they were receiving placebo (FC*belief*drug: χ²=4.35, b=194.38, S.E.= 93.19, z=2.09, p=0.038, Model 4B). Contrary, SN-DMN connectivity did not moderate the effects of drug assignment on beliefs (Table 2, Model 3A).

SN expectancy responses and long-term antidepressant placebo effects (Model 4).

Expectancy-related SN activity increases during the Antidepressant Placebo fMRI experiment predicted drug-assignment beliefs over 8 weeks (SN*time: χ²=2.58, B=37.78, S.E.= 18.88, z=2, p=0.045, Table 2, Model 4A), but not mood.

DISCUSSION

We tested the hypothesis that short-term antidepressant placebo expectancies encoded in the SN would extend into the DMN, supporting the formation of acute and long-term placebo effects. Indeed, learned antidepressant expectancies increased SN-DMN connectivity, scaling with acute placebo expectancies and mood. Over 8 weeks, expectancy-related SN-DMN connectivity, but not SN activity, predicted mood improvement, especially when participants believed they were receiving an SSRI as opposed to placebo. Interestingly, SN activity during the task predicted drug-assignment beliefs over the course of the clinical trial but did not predict mood improvement.

Our results support the hypothesis that placebo expectancies, encoded in the SN, contribute to the formation of long-term antidepressant placebo effects through interactions with the DMN. Increased SN activation during the processing of learned expectancies recruited the DMN extensively, including the rostral anterior cingulate and the posterior cingulate cortex, along with the thalamus and brainstem. These results are consistent with the role of the SN as a dynamic switch between the external world, through connections with the dorsal attention network (DAN), and the self, through connections with the DMN, which is known to play a role in internally-generated cognition, mind-wandering (5), self-related thought (36), autobiographical memory recollection (37), and self-referential judgments (36). Furthermore, increased SN-DMN coupling during the processing of learned expectancies is consistent with the theory that the DMN plays a central role in the formation of schemas and cognitive maps in response to latent or inferred rules from which to guide learning, decision-making, and valuation (38). It has been argued that the DMN plays a key role in the representation of the treatment context and generation of predictions across many forms of expectancy-mediated placebo (16). Furthermore, the DMN can integrate diverse streams of information, form new concepts, and integrate conceptual knowledge with predictions arising from associative learning circuits (16). Finally, heightened DMN activity, specifically in its anterior hub, is the best-replicated predictor of treatment response to a range of antidepressants (9–13), potentially reflecting the placebo effect inherent in the administration of any treatment.

Most importantly, our results provide evidence for the biological underpinnings of expectancy-mood interactions over time. Higher baseline SN-DMN connectivity was associated with both higher SSRI-assignment belief and greater effect of drug-assignment belief on mood improvement, especially when receiving placebo. These results provide a mechanistic understanding of the contribution of SN-DMN coupling to treatment response. Consistently, prior research using rsFC and ICA demonstrated that increased baseline rACC rsFC within the SN was associated with greater improvement following one week of placebo and ten weeks of open-label antidepressant therapy, reinforcing SN-DMN coupling as a biomarker for antidepressant placebo response. Furthermore, machine learning analyses indicated that increased salience network rsFC, primarily centered in the rACC, robustly predicted individual placebo responses (7). Similarly, findings from the Establishing Moderators and Biosignatures of Antidepressant Response in Clinical Care (EMBARC) study corroborate our observations. Using low-resolution electromagnetic tomography, Whitton and colleagues found that stronger baseline theta-band rACC–rAI (salience network hub) connectivity predicted greater depression improvement across 8 weeks of treatment for both treatment arms.

Of note, whereas we previously reported that baseline SN activity significantly scaled with depression severity (3), SN-DMN connectivity was not significantly correlated with depression severity. Together, these findings suggest that more severely depressed individuals might experience higher expectancies, yet a lack of subjective improvement (39), possibly due to alterations in SN-DMN functional connectivity.

The lack of significant drug effects on either beliefs or depressive symptoms in our moderately sized sample is not unexpected as meta-analyses find small and heterogenous SSRI-placebo differences in depression severity scores (i.e., <3 points in the HDRS) (40,41). These small effect sizes suggest that a large proportion of the clinical response in antidepressant trials is likely driven by non-specific factors, such as patient expectations, therapeutic context, and the natural course of the illness. This variability highlights the importance of identifying biomarkers that can predict who will respond to pharmacotherapy versus placebo, ultimately leading to more personalized and targeted treatment approaches.

This study has several limitations. First, the small sample size limits the examination of SSRI effects interacting with placebo beliefs, calling for larger clinical samples. Second, while baseline findings suggest moderation effects, further research is needed to confirm if SN-DMN connectivity mediates the link between beliefs and mood changes and to clarify directionality. Additionally, the lack of a “no treatment” control limits ruling out non-specific symptom improvement unrelated to beliefs. Nonetheless, our computationally defined trial-by-trial manipulation of antidepressant expectancies and their reinforcement aimed to target neural responses during expectancy processing rather than broader neural responses that might be seen during resting state or EEG. Finally, our results may not be applicable to other treatment therapies such as cognitive therapy or transcranial magnetic stimulation (TMS), or to patients undergoing antidepressant medication. A psychotropic-free sample was needed to avoid the potential confounding effects of other psychotropic medications.

In conclusion, our results inform an initial understanding of the cortical mechanisms involved in the persistence of antidepressant placebo effects. Taken together with our earlier findings (3), they reveal a temporospatial gradient from early processing of placebo-related cues in the associative sensory regions (e.g. DAN) to the integration of these cues with evolving internal experience in the SN, and finally the transfer of resulting expectancy representations to the DMN, which encodes persistent placebo-related beliefs. Our study is a step toward identifying mechanistic predictors of individual placebo response. If validated, these predictors could guide pharmacotherapy decisions, such as reserving more aggressive treatments for placebo non-responders, excluding high placebo responders from Phase II-III trials, or informing new therapeutic targets.