Participants
131 were assessed for eligibility, and 42 experiencing a current MD episode as determined by Structured Clinical Interview for DSM-538 and a CRP > 3mg/L were randomized (see Figure S1 for consort diagram). Randomized participants were free from all psychotropic and anti-inflammatory medications for at least 4 weeks and were without evidence of chronic infection, autoimmune or inflammatory disorders, or unstable medical illnesses as determined by medical history, physical exam and laboratory testing (full details of eligibility criteria and assessment are included in the Supplemental Materials). No patient was removed from psychotropic treatment for the purposes of the study. Of the randomized participants, 38 had available self-report measures, inflammatory markers, and behavioral data from the EBDM task, and 37 had available neuroimaging data. Two additional participants exhibited significant motion (3mm-6mm) during task-based fMRI; one additional subject exhibited a low calibration response at 14 days. These subjects were included in analyses, but sensitivity analyses were performed to assess their impact on reported results (see Supplemental Materials). Finally, two participants had an insufficient number of trials for division into training and test datasets, and were therefore excluded from the multivariate analysis.
Demographic and clinical data of randomized participants are presented in Table 1 and a full consort diagram is provided in Supplemental Figure S1. Written informed consent was obtained from all participants. The Emory Institutional Review Board granted study approval (IRB00087941). There were no serious adverse events associated with this study. A full list of adverse events is included in the Supplemental Materials.
Study Design
The study utilized a randomized, placebo-controlled, clinical trial design to examine the effects of a single-dose of the TNF antagonist infliximab on behavioral and brain measures of motivational anhedonia. Baseline blood, behavioral, self-report and neuroimaging assessments were followed by an infusion of either a 5mg/kg of infliximab or saline (placebo), administered over ~ 40 minutes from saline bags matched for color and consistency. Blood samples and behavioral and self-report measures were repeated at 3 and 14 days, and neuroimaging assessments including resting-state fMRI and an effort-based decision-making task11 were repeated at 14 days (see Fig. 1A). All study personnel were blinded to group assignment. The randomization and blind key were tracked by the Emory Investigational Drug Service.
Effort Based Decision-Making Task (EBDM)
The EBDM task is an fMRI-adapted effort-based decision-making task 11, 39 that measures neural responses to effort and reward magnitude. During each trial, participants were given the choice between High Effort and No Effort options. The High Effort option requires more effort (as measured by button presses) than the No Effort option. The reward obtained from the No Effort option was always $1 while rewards from the High Effort option varied between $1.00 and $5.75. The magnitude of effort required in the High Effort option consisted of 20%, 50%, 80%, and 100% of the participant’s maximal effort. The task shows good internal consistency, with a split-half reliability r = .94. Participants made choices in the scanner with the effort completed as soon as the scanning session was concluded. As such, the task can be viewed as measuring the choice to commit to effort expenditure for reward in the near future (see Supplemental Materials).
Clinical Assessments: Measures of motivational anhedonia included the reduced motivation (RM) subscale of the Multidimensional Fatigue Inventory (MFI40), a composite of items from the Motivation and Pleasure Scale-Self Report focused on effort41, and the anhedonia subscale from the Inventory of Depressive Symptomatology-Self Reported (IDS-SR42). Scales were collected prior to infusion at either screening, baseline MRI or infusion visits, and then at three days and 14 days post-infusion. These scales were pre-registered (https://osf.io/r6m49/) as clinical measures of motivational anhedonia given prior associations in inflammation and striatal function31, 43.
Biological Assays
Whole blood was collected into EDTA-containing vacutainer tubes through indwelling catheters after 30 minutes of rest to limit effects of stress. Plasma was isolated and stored at -80 until batched assay. Customized Fluorokine MAP Multiplex Human Biomarker Panels (R&D Systems, Minneapolis, MN) were used to measure plasma soluble tumor necrosis factor receptor 2 (sTNFR2) and other inflammatory markers (see Supplemental Materials). Inter- and intra-assay coefficients of variation were reliably less than 10%, and no values were below limits of detection. Plasma CRP was measured using a high sensitivity turbidimetric assay in the CLIA-certified Emory Medical Laboratory.
Neuroimaging Data Acquistion: Functional and structural neuroimaging data were acquired on a Siemens 3T Tim Trio using a 32-channel phased-array head coil. Trial presentations were synchronized to initial volume acquisition. Functional (T2* weighted) images were aquired using a multiband sequence with the following sequence parameters: 3-mm3 isotropic voxels, repetition time (TR) = 1.0 s, echo time (TE) = 30 ms, flip angle (FA) = 65°, 52 interleaved axial slices, with slice orientation tilted 18° relative to the AC/PC plane to improve the temporal signal-to-noise ratio (tSNR) and minimize signal dropout of ventromedial prefrontal cortex. At the start of the imaging session, a high-resolution structural volume was also collected, with the following sequence parameters: 2-mm × 1-mm × 1-mm voxels, TR = 1.9 s, TE = 2.27 ms, FA = 9°.
Power analysis and Impact of the Covid-19 Pandemic
Data in this study were drawn from a clinical trial (NCT03006393) focused on examining the effects of a single-dose of infliximab on brain function. The study was actively recruiting at the outset of the COVID-19 pandemic, at which point the investigators deemed continuation of the study to be unsafe, given the immune-suppressive effects of infliximab. Consequently, we were unable to achieve our originally proposed recruitment target of n = 80. As a result, the current study retains adequate power to detect large effect-sizes, but does not have adequate power for small or medium effect sizes, which may have increased type II error for some analyses. We note that prior pharmacologic studies of effort-based decision-making tasks have suggested large effect sizes (d > 1.0)44, 45, though these studies did not use infliximab.
Pre-registration: A pre-registered analysis plan that details hypotheses, key dependent variables, and primary methods can be found at https://osf.io/r6m49/. The current paper focuses on a subset of these data. Specifically, given the focus of the current paper on effects of infliximab on motivational circuitry, we used the secondary behavioral endpoint due to its temporal proximity to change in neuroimaging. Analyses not pre-registered are designated as “exploratory”. Table S1 in the Supplemental Materials summarizes our primary, secondary and exploratory analyses.
General Statistical Methods
Multiple linear regression was used to examine associations between change from baseline to endpoint across choice data, clinical measures, sTNFR2 and fMRI data extracted from regions-of-interest (ROIs). To assess the effects of infliximab on inflammatory markers, multivariate analysis of variance (MANOVA) was used. Of note, sTNFR2 was the inflammatory cytokine that exhibited the greatest change following infliximab versus placebo and was thus used as the primary immune endpoint as well as a proxy for TNF signaling in the statistical analyses (see Supplemental Materials). For longitudinal analyses, difference score distributions of sTNFR2 appeared parametric, and raw values were used for computation of difference scores unless otherwise noted. To assess change in anhedonia symptoms or effort discounting (k), an ANCOVA was used with drug condition as a fixed factor and baseline values along with demographic variables as covariates. For choice behavior on the EBDM task, a repeated measures ANOVAs was used in lieu of an ANCOVA to model the interactions between time, drug condition and individual effort levels. Greenhouse-Geisser corrections are reported in cases where the sphericity assumption was violated. All statistical analyses included sex and age as covariates unless otherwise specified. All statistical analyses are two-tailed unless pre-registered as a one-tailed test or otherwise noted.
Neuroimaging Preprocessing and Data Quality Evaluation
For all neuroimaging preprocessing and first-level GLM analysis, we used SPM12 (Wellcome Department of Imaging Neuroscience, Institute of Neurology, London, UK). SPM12 preprocessing included realignment estimation and implementation, co-registration to the individual’s high resolution structural scan, normalization to MNI space, and spatial smoothing using a Gaussian kernel (6mm FWHM). To control for motion and other artifacts, data were visually inspected, and 6 realignment parameters were included as covariates for all GLM analysis. Visual inspection revealed 2 participants showing evidence of significant motion (> 3mm) that were subsequently examined as potential high-influence data points. Additionally, the GLM contrasts used in univariate and multivariate analyses were evaluated for multivariate outliers using the Mahalanobis distance. The impact of subjects or sessions flagged as potential outliers using this criterion were evaluated in a series of sensitivity analyses that included participants with poor motion and/or multivariate outliers as covariates and they were not found to alter the significance of any reported neuroimaging associations (see Supplemental Materials for Sensitivity Analyses).
Neuroimaging Analysis – First Level General Linear Models (GLM): For all first-level GLMs, the canonical HRF was used, and event durations were modeled based on the duration of each cue for each trial with SPM default orthogonalization. Based on our pre-registered analyses and our prior work11, 39, we examined change (14 day vs. baseline scans) across 7 parametric modulator contrasts: effort level at Cue 1, reward magnitude at Cue 1, predicted subjective value at Cue 1 (SVpredicted), choice difficulty at Cue 2, subjective value of the chosen option (SVchosen) at Cue 2, a subjective value prediction error (SVPE) and choice “shifts” at Cue 2 (Fig. 1B).
The parametric modulator contrasts for Cue 1 were defined as follows:
-
Contrast 1: Effort magnitude. The amount of effort required for a given trial (20%, 50%, etc.).
-
Contrast 2: Reward magnitude. The amount of reward offered for a given trial ($1.00~$5.00).
-
Contrast 3: SV predicted . As in our prior work11, 39, SVpredicted was defined using a sliding window analysis of previously-experienced subjective values of the same trial type. Therefore, the SVpredicted for a trial that began by presenting 20% effort at Cue 1 would be calculated as the running average of the most recent SV values for trials that included 20% effort.
The parametric modulator contrasts for Cue 2 were defined as follows:
-
Contrast 4: SV chosen . The SVchosen was calculated as the subjective value of the option
-
Contrast 5: SVPE. The subjective value prediction error (SVPE) regressor was estimated by calculating the absolute value of the difference between the SVchosen, and the SVpredicted.
-
Contrast 6: Choice Difficulty. A parametric modulator calculated as the difference between the subjective value of the effortful and non-effortful options. Choices for which this difference was small indicated greater choice difficulty (because of similar values for both options).
-
Contrast 7: Choice Shift. A parametric modulator coding a “1” if the choice on the prior trial was the same as the choice made on the current trial, and “0” if not.
A full description of results from all a priori contrasts is included in the Supplemental Materials.
In addition to the parametric modulators described above, we estimated single-trial models for the purpose of multivariate, beta-series and brain signature analyses. For these analyses, a trial specific beta-weight was estimated for the Cue1 onset, the Cue2 onset and the decision-phase onset of each trial. Analyses using single trial models focused on the Cue1 timepoint.
Neuroimaging Analysis – Second Level Univariate GLMs
For comparison of Day 14 and baseline scans, the two task runs from each timepoint were concatenated into a single first-level GLM. A second-level contrast of [-1 -1 1 1] was then used to compare parametric modulators during the two baseline timepoint runs to the two 14 day timepoint runs. To examine the effects of infliximab, change in acceptance of 100% effort trials, change in effort discounting (k) and change in sTNFR2 levels, each of these variables were separately entered into second-level GLM that also included sex and age as covariates. For whole-brain analysis, correction for multiple comparisons was obtained using cluster-correction as implemented in SPM12, with an uncorrected height threshold of p < 0.001. For ROI mean betaweights from all voxels in each ROI was averaged and analyzed.
Neuroimaging Analysis – Univariate ROI Analysis: Our analysis plan identified five a priori regions of interest: the dmPFC, bilateral insula, nucleus accumbens (NAcc) and ventromedial prefrontal cortex (vmPFC). Masks for these ROIs were drawn from a prior functional parcellation of medial prefrontal cortex46, the Glasser atlas47 and the Harvard-Oxford atlas48. For ROI analyses, the first eigenvariate of mean betaweights from all voxels within each mask was extracted for each participant and used in subsequent analysis.
Neuroimaging Analysis – Multivariate ROI Analysis
As an exploratory analysis, change in striatal sub-regions for the SVpredicted and SVchosen was used to classify drug condition assignment. A cross-validated 3-fold classification with partial least squares regression was employed to estimate the area under the receiver operating characteristic (AUROC) for drug classification based on activity in bilateral ventral striatum, putamen and caudate as defined by the Pauli basal-ganglia atlas49. Only neuroimaging data was used (i.e., age and sex covariates were not included). The k-fold procedure was estimated 10 times for each region, with the mean AUROC across each fold and iteration used to estimate predictive performance for each region. For statistical inference, we estimated the mean squared error (mse) of classification for each ROI and compared it to a null distribution of 5,000 mse values generated by a randomly permuted neuroimaging data.
Neuroimaging Analysis – Beta Series Correlation (BSC)
To understand how infliximab-induced changes in dmPFC responses during EBDM may drive network-level changes within corticostriatal circuits, we examined task-based functional connectivity (beta-series correlation; BSC). To assess changes in BSC, we first isolated the time series of beta-weights at Cue 1 for each ROI for each participant. We focused on Cue 1 for this analysis as it would detect changes prior to decision-outcome, and thereby could reveal network-level changes related to processing of partial information that would not be confounded by differences in the proportion of effortful options accepted between the two drug conditions. We then estimated the Pearson correlation between dmPFC and target striatal regions (VS, putamen, caudate). Resulting correlations were Fisher transformed to create a difference score in the BSC between each pair of nodes (ΔBSC). These ΔBSC scores were used as dependent variables in Ordinary Least Squares (OLS) regression analyses and bootstrapped-mediation analyses as described above with drug condition or ΔsTNFR2 used as predictor variables.
Neuroimaging Analysis – Reward Signature Comparison The goal of this analysis was to evaluate the similarity between change in neural activity during EBDM (14 day – baseline) and a pre-trained ‘reward signature’ developed to classify monetary wins during losses during gambling and monetary incentive delay tasks50. We first calculated the change in cosine similarity between each participant’s 14day-baseline mean contrast image and a neural signature. Regression was used to assess the relationship between the resulting cosine similarity and drug condition, ΔsTNFR2, Δk, and Δ100% Effort as well as self-report measures of anhedonia.