The data reported here were acquired in the context of a larger study, components of which have been  and will be reported separately. More precisely, we conducted this randomized, single-centre, double-blind, placebo-controlled, parallel group, prospective clinical trial from March 7, 2018, through March 18, 2020, in the Geneva University Hospitals (Switzerland), i.e. before the COVID-19 pandemic – we can reasonably exclude COVID-19 loss of smell and taste in the patients recruited. The trial was carried out in accordance with the protocol and with principles enunciated in the current version of the Declaration of Helsinki, the guidelines of Good Clinical Practice (GCP) issued by ICH, and the Swiss Law and Swiss regulatory authority’s requirements. The trial was registered on ClinicalTrials.gov (NCT03347890) and the protocol and any subsequent amendments were approved by local ethical committee (Commission Cantonale d’Ethique de la Recherche, CCER, Genève) and by Swiss Agency for Therapeutic Products (Swissmedic). Safety of the participants was monitored throughout the trial by an independent data monitoring committee and was monitored from screening to week 16. Participants were informed about the aims of the study and gave their written consent before the initiation of any trial-related procedures. The trial overview is presented in Fig. 1.
Multidisciplinary weight loss program: Participants with obesity (Body Mass Index, BMI ≥ 30 kg/m2 and < 45 kg/m2) were recruited in the Division of endocrinology, diabetology, nutrition and therapeutic patient education in Geneva University Hospitals among patients addressed for a structured and multidisciplinary patient educational weight loss program. This weight loss program based on lifestyle counseling (combining group and individual approach) includes a cognitive behavioral therapy coupled with a diet and physical activity support, as described in detail in Pataky et al. . Patients attended individual and group counseling sessions during the 16-week period, delivered by qualified health care professionals (registered dietician, nurse and physicians specialized in obesity treatment and patient education). The weight loss program was consequently tailored to each patient’s needs.
Inclusion and exclusion criteria: The inclusion criteria were defined as age between 18 and 75 years, stable body weight (< 5% reported change within 3 months before screening), right-handedness and being currently non-smoker. The exclusion criteria were based on contraindications to Liraglutide treatment, any drugs interfering with the body weight control (e.g., Orlistat, Phentermine and Topiramate, Buproprion and Naltrexone) or any centrally acting medication, history of any psychiatric disease, heart failure (NYHA II-IV), type 1 and type 2 diabetes mellitus and deficits of smell and taste. The complete list of eligibility criteria is listed in the supplementary information. All participants gave written informed consent and received 200 Swiss Francs (the equivalent of 200 USD$) for their participation in the two sessions.
Trial population: A total of 73 participants with obesity (OB) were screened. After being checked for the study eligibility criteria, 66 participants were included in the trial. Among them, 32 were randomized to liraglutide 3.0 mg combined with lifestyle counseling and 34 to placebo combined with lifestyle counseling (see Fig. 1 in Supplementary Information). Baseline characteristics of the studied population are described in Table 1, Supplementary Information. In total, 22 participants were excluded from the analysis (10 participants did not complete the second testing session and 12 participants had missing or corrupt MRI data). We report data on the 44 remaining participants (liraglutide group: age 37.4 years ± 11.18, BMI 35.89 kg/m2 ± 3.01, n = 20; placebo group: age 40.04 years ± 14.1, BMI 34.88 ± 2.87 kg/m2, n = 24).
Taste stimuli and presentation
We used two types of stimuli in this experiment: a milkshake and a tasteless solution.
We prepared the milkshake by mixing a chocolate flavored milk drink (300 g) with a fior di latte flavored ice cream (60 g) for a total of 71 kcal/100 g.
We prepared a tasteless solution as close as possible to artificial saliva in three steps. First, we diluted potassium chloride (KCl, 1.8g) and sodium bicarbonate (NaHCO3, 0.21g) in 1 L of distilled water. Second, we created three less concentrated versions of this solution. Thus, there were 4 different tasteless concentrated solutions (1/1, ¾, ½ and ¼). Third, patients were invited to taste the 4 solutions. We picked the one that tasted the most neutral to them (i.e., closest to 50 on a scale from 0 to 100) as their tasteless solution. We believe it was better to use one of these 4 tasteless solutions as the control stimulus rather than water because water has an inherent taste .
The milkshake and the tasteless solution were kept in the fridge. We took them out simultaneously, 30 minutes before the experiment so that they were delivered at ambient temperature.
The apparatus used to deliver the liquids in the scanner has been described in Muñoz-Tord et al. . In a nutshell, a 3D-printed pacifier-shape fMRI mouthpiece paired with a gustometer was used to deliver liquids while participants were lying down. As demonstrated in Muñoz-Tord et al. , this allows a precise, reliable and comfortable liquid delivery. The collection of the responses was controlled by a computer running MATLAB (version R2015a; MathWorks, Natick, USA). The presentation of the visual stimuli was implemented using Psychtoolbox (version 3.0) .
Participants who fulfilled the randomization criteria were randomly assigned, in a 1:1 ratio, to receive liraglutide 3.0 mg or placebo, after a dose escalation period starting at 0.6 mg q.d., with weekly increments of 0.6 mg, administered subcutaneously by pen injectors. The placebo pen injector was strictly identical to the liraglutide pen injector. Participants were followed-up on a weekly basis during the dose escalation period of 5 weeks and monthly afterwards.
Blood samples for study purposes were collected both at baseline and 16-week follow-up. Plasma fasting blood glucose, insulin, plasma lipids and HbA1c were measured by routine biochemistry in fasting conditions.
The experiment consisted of three separate testing days (see Fig. 1). The first time participants came to the laboratory for a pretest (see description below). The second time participants came for a test session before the beginning of the intervention (i.e., baseline testing session). This session took place in the morning. All participants were asked to fast overnight. Afterwards, participants followed the intervention (placebo + counseling vs liraglutide + counseling) and came back to the laboratory a third time at the end of the intervention for a second test session (i.e., 16 weeks follow-up testing session). Both test sessions followed the exact same procedure.
Please note that during the test sessions participants performed multiple experimental tasks, but, here, we only report the results for the hedonic reactivity task.
Pretest. Participants chose the most neutral tasteless solution to them. The 4 solutions were presented to them in shot glasses (= 1dl). Participants self-reported current hunger level and pleasantness, intensity and familiarity levels for their selected tasteless solution and for the milkshake (see Table 2, supplementary information). They also underwent 10–20 minutes of structural scans in the MRI. This small fMRI session allowed them to be more confident and comfortable for the longer functional scans taking place during the test sessions.
Test session. We administered a taste reactivity task while participants were lying in the scanner. The task consisted in the evaluation of the perceived pleasantness, intensity and familiarity of the two different stimuli: the milkshake and the tasteless solution. Participants were instructed to assess the solutions focusing on their current perception of them. During each trial, 1 mL of the solution was administered, and the delivery order of the two conditions was randomized within each participant. Participants were visually guided through the task with on-screen indications. First, they saw a 3 seconds countdown before the solution was delivered, followed by an asterisk presented for 4 seconds and indicating them to keep the solution on their tongue until they saw the indication to swallow: “swallow please” (see Fig. 2). We asked them to wait 4 seconds before swallowing to avoid adding movement noise to the Blood-Oxygen-Level-Dependent (BOLD) response. Since they were lying down, the mouthpiece was placed in such a way that the solution was delivered at the center of the participant’s tongue. We expected that the solution would slide down to the back of their tongue during the 4 seconds period in which they had the solution on it. The experimental trials were intertwined with rinse trials to cleanse the participants’ palates with 1 mL of water. All 40 evaluations (20 per solution) were done on visual analog scales displayed on a computer screen. Participants had to answer through a button-box placed in their hand. The visual scales ranged from “not perceived” to “extremely intense” for the intensity ratings; from “extremely unpleasant” to “extremely pleasant” for the liking ratings; and from “extremely familiar” to “extremely unfamiliar” for the familiarity ratings.
Behavioral and metabolic Data
We analyzed the behavioral and metabolic data with R (version 4.0; R Core Team, 2019).
We build two statistical models. The first model aimed at testing the relationship between weight loss (measured by subtracting the BMI after the intervention from the BMI before the intervention). We entered (1) intervention: placebo or liraglutide as a fixed effect and (2) age and (3) gender as control factors. As a random effect we entered intercepts for participants. We build the model as follows:
$$weight loss \tilde intervention + gender + age + \left(1\right| id)$$
The second model aimed at testing the relationship between the perceived pleasantness of taste and the intervention. We entered (1) the taste stimulus: milkshake or tasteless, (2) session: pre- or post-intervention, (3) intervention: placebo or liraglutide, and (4) a linear decreasing contrast over trials to account for satiation. As random effects we entered intercepts for participants as well as by-participant random slopes for the effect of the interaction between taste stimulus session and trial. We build the model as follows:
$$pleasantness \tilde intervention \times stimulus \times session \times satiation + (stimulus \times session \times satiation | id)$$
We used the lmer4 package  and the LmerTest package . We extracted Bayse factors through linear mixed bayesian analysis using brms , cmdstanr  and baysestestsR  packages. The models were estimated using Markov chain Monte Carlo (MCMC) sampling with 4 chains of 5000 iterations and a warmup of 1000. The dependent variables were scaled before being entered in the model. For the first model (weight loss) prior parameters were set as normal (mean = 0.00, SD = 1.00) distributions. For the second model (perceived pleasantness) prior parameters were set as normal (mean = 0.00, SD = 1.00) distributions. The Bayes factors reported for the main effects compared the model with the main effect in question versus the null model, while Bayes factors reported for the interaction effects compared the model including the interaction term to the model including all the other effects but the interaction term. Evidence in favor of the model of interest was considered anecdotal (1 < BF10 < 3), substantial (3 < BF10 < 10), strong (10 < BF10 < 30), very strong (30 < BF10 < 100) or decisive (BF10 > 100). Similarly, evidence in favor of the null model could also be qualified as anecdotal (0.33 < BF10 < 1), substantial (0.1 < BF10 < 0.33), strong (0.033 < BF10 < 0.1), very strong (0.01 < BF10 < 0.033) or decisive (BF10 < 0.01).
Acquisition Parameters. Acquisition parameters were identical to the ones described in Muñoz-Tord et al. . The neuroimaging data were acquired on a 3-Tesla MRI system (Magnetom Tim Trio, Siemens Medical Solutions) supplied with a 32-channel head coil following a gradient echo (GRE) sequence to record data acquisition BOLD signal. We recorded forty echo-planar imaging (EPI) slices per scan with an isotropic voxel size of 3 mm. The scanner parameters were set at: echo time (TE) = 20 ms, repetition time (TR) = 2000 ms, field of view (FOV) = 210×210×144 mm, matrix size = 70×70 voxels, flip angle = 85°, 0.6 mm gap between slices. Structural whole brain T1-weighted (T1w) images (isotropic voxel size = 1.0 mm) were acquired, as well as dual gradient B0 field maps (Fmaps) for each participant to correct for inhomogeneity distortions in the static-field.
Preprocessing. As in Muñoz-Tord et al. , we created a pipeline optimized for the preprocessing of our neuroimaging data. More specifically, we combined the Functional Magnetic Resonance Imaging of the Brain (FMRIB) Software Library (FSL, version 4.1)  with the Advanced Normalization Tools (ANTS, version 2.1) . The BOLD signal is highly prone to movement artifacts. This characteristic makes our experimental setting particularly challenging to analyze because our participants swallowed solutions in the scanner, which produces major deglutition artifacts. To offset this loss of signal- to-noise ratio (SNR), we followed Griffanti et al. ’s protocol. This protocol uses an fMRI independent component analysis (ICA) to remove artifacts. The multivariate exploratory linear optimized decomposition tool (MELODIC)  decomposes the raw BOLD signal into independent components (IC). We chose this ICA-based strategy for motion artifact removal because it is more reliable to remove motion-induced signal variations than regressions from motion parameters . Two researchers from our laboratory independently hand classified a sample of 20 participants’ IC into two categories: ‘potential signal’ or ‘clear artifact’ (e.g., motion/deglutition, blood flow in arteries). The two researchers’ categorizations were then compared, and each discrepancy was discussed until an agreement was reached (inter-rater reliability = 93%). This process allowed manually classified components. These components were then used to train a classifier using a random forest machine learning algorithm . We used leave-one-out testing, i.e. we iteratively left one participant out of the training data and tested the classifier accuracy on the left-out participant. Leave-one-out testing at the optimal sensitivity (threshold = 5) resulted in a median 94% true positive rate (i.e., the percentage of ‘true signal’ accurately classified). We consequently applied the FMRIB’s ICA-based X-noiseifier (FIX) to automatize the denoising of our BOLD signal . We then applied field maps to correct geometric distortions. We used ANTS for a diffeomorphic co-registration of the preprocessed functional and structural images in the Montreal Neurological Institute (MNI) space, using the nearest-neighbor interpolation and leaving the functional images in their native resolution. Finally, we applied a spatial smoothing of 8 mm full width half maximum (FWHM).
Statistical analysis. We used the Statistical Parametric Mapping software (SPM, version 12)  to perform a random-effects univariate analysis on the voxels of the image times series following a two-stage approach.
For the first-level, we specified a general linear model (GLM) for each participant. We used a high-pass filter cutoff of 1/128 Hz to eliminate possible low-frequency confounds. Each regressor of interest was derived from the onsets of the stimuli and convoluted using a canonical hemodynamic response function (HRF) into the GLM to obtain weighted parameter estimates. The GLM consisted of seven regressors: (1) the onsets of the trial, (2) the onsets of the reception of a taste stimulus modulated by (3) the presence of the milkshake (4) the trial-by-trial ratings of the perceived pleasantness (5) the onsets of the question about pleasantness, (7) the onset of the question about intensity and (8) the onset of the question about familiarity. We extracted the contrast of the taste delivery modulated by the perceived pleasantness for each participant at each session (43 participants x 2 sessions = 86).
For the second-level, we entered the first level contrasts in a mixed measures 2 (session: pre or post) by 2 (treatment: placebo or liraglutide) ANOVA using the multivariate and repeated measures toolbox (MRM) . The MRM toolbox is a MATLAB toolbox allowing to perform mass multivariate group models of neuroimaging data using the summary statistic approach by selecting the correct error term . We extracted F contrasts with a voxel-wise significance threshold set at p < 0.001 FDR corrected for multiple comparisons. For display purposes we plotted non-masked and uncorrected statistical p-maps of our group results overlaid on a high-resolution template (CIT 168) in MNI space.
Code and Data accessibility
The computer code used to preprocess and analyze the data is available in a publicly hosted software repository (for preprocessing of the fMRI data: https://github.com/munoztd0/Mouthpiecegusto/tree/main/preprocessing; for data analysis : https://github.com/evapool/GLP1_Pleasure).