Ethics statement
This trial has been registered at ClinicalTrials.gov (NCT04226911), was approved by national ethical committees (protocols and amendments) and was conducted in accordance with the Declaration of Helsinki. The study was monitored for Good Clinical Practice compliance by the European Clinical Research Infrastructure Network, as described previously18. All participants provided written informed consent.
Trial design
The SWEET trial was a two-armed parallel group RCT conducted at four intervention sites: Athens (Harokopio University of Athens, Greece), Copenhagen (University of Copenhagen, Denmark), Maastricht (Maastricht University, the Netherlands) and Pamplona (University of Navarra, Spain), effectively covering northern, central, southern and south-eastern Europe.
The 1-year RCT consisted of an initial 2-month weight loss period followed by a 10-month, randomized, two-armed parallel WM period. The full 1-year trial on which the majority of outcomes are reported (months M0–M12) comprises the combined periods and is thus termed WM. For adults, the goal was first to achieve a weight loss of ≥5% of the initial weight and second, to maintain their new body weight. For children, the first goal was to achieve weight stability and second, to maintain their BMI-for-age z-score. CIDs were carried out at baseline (M0), after weight loss and weight stability (M2) and twice during WM (M6 and M12).
Participants
In total, 341 adults and 38 children were included in the trial (Fig. 1). The analysis of gut microbiota composition was done in a subgroup of 137 adult completers. Participants were enrolled between June 2020 and October 2021. The recruitment procedure and all inclusion and exclusion criteria are described elsewhere18. In brief, 18–65-year-old men and women (self-reported sex) with BMI ≥ 25 kg m−2 and 6–12-year-old boys and girls (self-reported sex) with a BMI-for-age of >85th percentile were included. Children were included in a family setting with at least one recruited parent. Participants were required to have a regular consumption of sugar-containing or sugar-sweetened products. Adult participants were excluded at screening if, for example, they had been surgically treated for obesity, were taking medication affecting body weight, had been diagnosed with diabetes, had a fasting glucose >7.0 mmol l−1 or had systolic blood pressure of >160 mmHg and/or diastolic blood pressure of >100 mmHg. Children were excluded if, for example, they performed >10 h of intensive physical training per week, had self-reported eating disorders, were diagnosed with diabetes or used medication that affected their body weight. All adults received low-energy diet products free of charge. Participants in Copenhagen, Pamplona and Athens did not receive reimbursement for their participation; travel expenses and financial compensation were provided for participants in Maastricht.
Intervention
The trial lasted 1 year for each participant. The first CID (M0) was on 24 August 2020, and the last participant’s final visit (M12) was on 6 October 2022. During the initial 2-month period, adults—regardless of randomization—received the low-energy diet (Cambridge Weight Plan). For children, weight stability should be achieved by following the dietary recommendations of the American Academy of Pediatrics on the prevention, assessment and treatment of overweight and obesity18. All participants were randomly allocated to one of two diet groups in a 1:1 ratio by a site-specific, computerized randomization list created by a person in Copenhagen not involved in the RCT. Stratification was done by sex, age (<40 or ≥40 years) and BMI (<30 or ≥30 kg m−2) in blocks of four. Each household, including children, was randomized to the same intervention group determined by the oldest member of the household. If there was more than one eligible child per household, it was still the randomization group of the oldest adult participant that determined the child allocation.
Although randomization was done after inclusion (at screening), it was not revealed to the participants before completion of the 2-month weight loss or weight stability period. The two ad libitum intervention diets were a healthy diet with <10 E% added sugar, allowing foods and drinks with all types of S&SE products commercially available (S&SEs group), and a healthy diet with <10 E% added sugar, not allowing S&SE products (sugar group). The maximum allowed sugar intake was calculated individually at M2 and recalculated at M6 and then converted to a simple trial-specific unit system (one unit = 10 g sugar). The participants received their maximum unit intake and lists with sugar-rich products, including the unit content. Lists were divided into different categories; for example, drinks, breakfast or desserts. The lists also provided a corresponding product with S&SEs (similar in weight or volume; details described elsewhere18). For the S&SEs group, the aim was to replace as many sugar-containing products as possible with S&SEs products, whereas S&SEs products were not allowed in the sugar group. S&SEs included high-potency sweeteners (for example, aspartame, acesulfame-K, saccharin, thaumatin, neotame, stevia glycosides), polyols (for example, erythritol, sorbitol, mannitol, isomalt, maltitol, lactitol, xylitol), slowly digestible carbohydrates (for example, sucromalt, isomaltulose) and sweet fibres or oligosaccharides (inulin-type oligosaccharides). Owing to the characteristics of the trial, blinding was not possible, but all efforts to blind trial staff taking measurements and doing statistical analyses were made. During the intervention period, participants were supervised by dieticians at least every third month. The goal for adults was to maintain weight loss, and for children, the goal was to maintain BMI-for-age z-score. Further reduction in body weight or BMI-for-age z-score was allowed if the participant was compliant with the intervention.
Data collection and outcomes
Data were collected according to common standard operating procedures, which were used in all intervention sites. Most data were collected at the CIDs after ≥10 h of overnight fast. All data were stored in a central data hub in Copenhagen, from which pseudo-anonymized data can be requested until 2032 through a data-sharing contract. As of 2032, fully anonymized data can be transferred.
Primary outcomes
Body weight
Body weight was measured to the nearest 0.1 kg on a digital scale, with participants wearing underwear or light clothes. Fasting body weight was measured at screening and CIDs, but fasting was not required at other visits.
Gut microbiota composition
Participants were selected based on whether they completed the full intervention and whether faecal spot samples were available at all time points. The participants were stratified based on age, sex and centre. Faecal spot samples were collected and immediately frozen (−20 °C) at home by the participants before all CIDs. At CIDs, the faecal spot samples were stored at −80 °C at the intervention sites. Barcoded amplicons from the V3–V4 region of 16S rRNA genes (341 F, 5′-CCTACGGGNGGCWGCAG-3′; 785 R, 5′-GACTACHVGGGTATCTAATCC-3′) were generated using the Illumina two-step PCR protocol, and sequencing on an Illumina MiSeq with the paired-end (2×) 300 bp protocol (Nextera XT index kit). After each PCR step, the products were purified (QIAquick PCR Purification Kit), the size of the PCR products was checked on a fragment analyzer (Advanced Analytical), and concentration was quantified by fluorometric analysis (Qubit dsDNA HS Assay Kit). For genomic DNA isolation, QIAamp Fast DNA Stool Mini Kits (Qiagen) were used. All outcomes were determined at M0, M2, M6 and M12. Stool consistency was assessed using BSS for each sample41.
Secondary outcomes
Secondary outcomes included changes in anthropometry and body composition, risk factors for T2D and CVD, IHL content, the occurrence of (serious) AEs, gastrointestinal symptoms and use of concomitant medication in adults with overweight or obesity.
Anthropometry and body composition
The methods for measuring anthropometry (BMI, waist and hip circumference) and body composition (fat percentage, fat mass, fat-free mass and computed visceral adipose tissue (CoreScan software; Encore v.17.0) using Dual-energy X-ray absorptiometry) have been extensively described elsewhere18.
Risk factors for T2D and CVD
Risk factors for T2D include elevated levels of glucose, insulin and HbA1c; for CVD, risk factors include high cholesterol, elevated triglycerides and high blood pressure. During CIDs, fasting venous blood samples were drawn. Whole blood samples were collected in EDTA tubes for HbA1c analysis and immediately stored locally at −80 °C. Blood samples were collected in serum separator tubes for analyses of lipids (triglycerides, total cholesterol, LDL-C and HDL-C) and insulin. Whole blood samples with fluoride were collected for glucose analysis, with EDTA for CCK, GLP-1 and EDTA plus aprotinin for ghrelin analysis. All blood samples were centrifuged at 1,500g for 10 min at 4 °C. Supernatant (serum or plasma) samples were aliquoted and stored locally at −80 °C until shipment to the central lab at Bioiatriki (Athens). Glucose concentration was measured by the enzymatic Hexokinase/G-6-PD method (Alinity c Glucose Reagent Kit 07P55, Abbott) using an Alinity c analyser (a clinical chemistry analyser from Abbott). A conversion factor for glucose, from mg dl−1 to mmol l−1, was applied by multiplying by 0.0555. Insulin concentration was measured by a chemiluminescent microparticle immunoassay (Alinity i Insulin Reagent kit, Abbott) using an Alinity i analyser (immunoassay analyser from Abbott). A conversion factor for insulin, from µIU ml−1 to pmol l−1, was applied by multiplying by 6. HbA1c was measured by an enzymatic assay (Alinity c Haemoglobin A1c Reagent kit, Abbott) using an Alinity c analyser, clinical chemistry (Abbott). One participant had results below the lower detection limit during two CIDs. These latter results were reported as 5.0%. Total cholesterol was measured by an enzymatic assay (Alinity c Cholesterol Reagent kit (Abbott)), and LDL-C and HDL-C were measured by liquid (Alinity c Direct LDL Reagent kit) and accelerator (Alinity c Ultra HDL Reagent kit) selective detergent methods, respectively (both Abbott). All cholesterol concentrations were analysed using an Alinity c analyser. A conversion factor for cholesterol (total, LDL and HDL), from mg dl−1 to mmol l−1, was applied by multiplying by 0.02586. Triglycerides were measured by a glycerol phosphate oxidase method (Alinity c Triglyceride Reagent kit, Abbott) using an Alinity c analyser. A conversion factor for triglycerides, from mg dl−1 to mmol l−1, was applied by multiplying by 0.01129. CCK was measured by a competitive enzyme immunoassay method (RayBio Human CCK Enzyme Immunoassay kit, RayBiotech) using a BioTek Quant analyser (BioTek Instruments). Two participants had results above the range of quantification (4,000 pg ml−1) during two CIDs, and their CCK results were reported as 4,000 pg ml−1.
IHL content
IHL content was analysed in a subgroup (n = 27) of adults at Maastricht (S&SEs group, n = 15; sugar group, n = 12). Proton-magnetic resonance spectroscopy (1H-MRS) was performed using a 3T MR system (Achieva 3T-X Philips Healthcare) at M0, M2 and M12. A 32-channel sense cardiac/torso coil (Philips Healthcare) was used, and a 30 × 30 × 30 mm voxel was placed in the lower hepatic lobe. The STEAM (repetition time, 4,500 ms; echo time, 20 ms; number of signal averages, 128) sequence was used, as described previously42,43. VAPOR water suppression was applied, and an additional water reference scan was obtained (number of signal averages, 16). A custom-written MATLAB script (MATLAB 2014b, The MathWorks) was used to post-process the spectra. Phasing, frequency alignment and eddy current correction were all performed on spectra before signal averaging. CH2 resonance, as a percentage of the sum of CH2 + H2O resonances (CH2/(CH2 + water)), was used as a parameter of IHL content.
Changes in (serious) AEs and concomitant medication use
Any experienced (serious) AEs or changes in medication use were registered during the CIDs. During the WM period, the participants were asked, regardless of intervention, directly about AEs potentially related to the consumption of S&SEs, including gastrointestinal symptoms and headache.
Other outcomes
Subjective appetite sensations over the last 7 days were reported in a questionnaire delivery platform on-site or at home within ±7 days of each CID. The participants answered five questions: (1) how strong was your desire to eat savoury foods; (2) how strong was your desire to eat sweet foods; (3) how satiated have you felt; (4) how hungry have you felt; and (5) how full have you felt? All questions were developed in English and translated into the local language at each site. A visual analogue scale with extremes anchored at each end (0, not at all; 100, extremely) was used. Depending on the device that was used, lines were not necessarily 10 cm long, but the rating was presented as a percentage; for example, a mark at 4 cm on an 8 cm line would correspond to 50%.
Compliance
To assess compliance, participants completed 4-day weighed records of all foods and drinks (three weekdays and one weekend day) at M0 and M12. Only records of a minimum of 3 days were deemed valid and used for further analysis. Daily average intake of energy and macronutrients was calculated by national dietary software at the four intervention sites. In Copenhagen, all information from food records was manually entered into the software programme DankostPro44. This software is based on the official Danish national food composition database (v.4) developed by the National Food Institute at the Technical University of Denmark45. In Maastricht, food intake data were analysed by the Eetmeter food diary and analysis tool (Voedingscentrum). In Harokopio, food intake data were analysed with Nutritionist V diet analysis software (v.2.1, 1999; First Databank), extensively amended to include traditional Greek foods and recipes, as described in the Food Composition Tables and Composition of Greek Cooked Food and Dishes46. Furthermore, the database was updated with nutritional information of processed foods provided by independent research institutes, food companies and fast-food chains. In Pamplona, food records were manually entered in the online application Nutrium (Nutrium.com), which is based on two food databases (BEDCA & CESNID from Spain and the US Department of Agriculture).
Furthermore, intake (g) of products with sugar and S&SEs and the corresponding units were estimated. Products of interest are described elsewhere18. As an objective measure, 24 h urine samples were collected at M0, M6 and M12. The urine samples were weighed and volumes registered (Maastricht, Harokopio and Pamplona) or calculated from the urinary density (Copenhagen). If urinary volume was not registered, urinary weight was divided by 1.0165 g ml−1 (M0, n = 7; M6, n = 0; M12, n = 2). Urinary biomarkers of S&SEs (acesulfame-K, saccharin, sucralose, cyclamate and steviol glucuronide) as well as glucose, fructose and sucrose were analysed by ultra-pressure liquid chromatography coupled to tandem mass spectrometry47. Preparation and analyses followed procedures described elsewhere47 and were conducted at Wageningen University. After correction for dilution, urinary concentrations (ng ml−1) were multiplied by 24 h urine volume and converted to daily excretions (mg day−1). Urinary urea concentration was analysed locally and converted to daily urinary nitrogen excretion (g day−1) by multiplying urea excretion (g day−1) by 0.4664. At Copenhagen and Maastricht, urea was measured by an enzymatic UV test (colourimetry) (ABX Pentra Urea CP, Horiba ABX) using an ABX Pentra 400. At Harokopio, urea was measured by an enzymatic colourimetric (Urease) Alinity c Urea Nitrogen Reagent kit (Abbott Laboratories). At Pamplona, urea was measured by an enzymatic kinetic test (COBAS 8000, Roche Diagnostics).
For the per-protocol population, participants’ compliance was estimated using points (minimum zero points and maximum four points) in relation to four criteria: intake of sugar units and S&SE units and urinary excretion of S&SEs at M6 and at M12. In three out of the four criteria, 75% of the group with the highest (>Q1) or lowest (
Physical activity was measured in a subgroup at Maastricht for seven consecutive days at M0, M6 and M12 using a triaxial accelerometer (activPAL 3TM micro, PAL Technologies). The activPAL was attached to the anterior thigh of the participants and measured posture allocation, step count and 24 h physical activity, distinguishing between sleeping time, low-to-moderate physical activity (time spent standing or walking <100 steps per min), moderate-to-vigorous physical activity (time spent walking ≥100 steps per min or cycling) and sedentary time (time spent sitting or lying down).
Sample size
Sample size calculation was based on body weight results from a previous trial48. It was estimated that a mean difference of 1.5 kg, with a s.d. of ±3.5 kg, 90% power and a two-sided alpha level of 0.05, would require 231 completers. With an estimated dropout of 30%, a minimum of 330 adult participants should be included (approximately 25% per intervention site). The sample size for the 1-year change in gut microbiota required a minimum of 100 participants (n = 50 per intervention group) and was based on a calculation taking into account ~10% change in 20 of the 50 most abundant operational taxonomic units with an alpha of <0.05%. According to this calculation, a total of 40 participants would be enough to detect compositional changes. Furthermore, considering previous work49 in which n = 75 was used to determine differences in beta-diversity, the estimated number of n = 100 participants in our statistical power calculation would indeed be sufficient. No power calculation was performed for the children; hence, the analyses of BMI z-scores were exploratory only.
Statistical analysis
Statistical analyses were conducted in R (v.4.3.1). Baseline characteristics before (M0) and after weight loss or weight stability (M2) are presented as medians (Q1–Q3), and changes at M2 and M12 are presented as unadjusted mean ± s.d. Differences between groups are presented as adjusted mean ± s.e.m. Given that randomization was completed at M0 (that is, 2 months before the 10-month intervention period was initiated), differences in the participant characteristics between groups after weight loss or weight stability (M2) and in changes during the 2-month period (M2–M0) were analysed by analysis of covariance (ANCOVA) adjusted for sex, age, baseline body weight and site. Sex was included as a covariate, given that we know from previous studies that different outcomes can vary between men and women50. The trial was not powered to analyse data from men and women separately.
Change in body weight was calculated as the difference between M0 and M12 (M12 − M0). For the primary ITT analysis, the population was defined as all randomized participants who obtained the required ≥5% weight loss at M2. For the ITT analysis, missing data (that is, body weight at M12) were imputed as the last observation carried forward. The analysis of differences between the groups was conducted using an ANCOVA linear mixed model, with individual change in body weight as the response, and intervention group, baseline body weight, age, sex and site as fixed effects. Additionally, a complete-case analysis (all dropouts omitted) and a per-protocol analysis (only compliant participants defined by the point scoring system) were analysed with the same ANCOVA model. Finally, adult body weight was analysed with the inclusion of all visits (time points) as a linear mixed-effect model with repeated measurements. This analysis had body weight as the response variable and included the fixed effects time-intervention interaction, age, sex and baseline body weight. Site and participants were included as random effects. If there was a significant interaction, post hoc tests were conducted with the R extension package ‘emmeans’ to calculate the estimated marginal mean and s.e.m. at each time point for comparison of the groups. This model was also used to impute missing body weight values for post hoc ITT analyses. For secondary outcomes on continuous data, the main analysis compared the 6-month and 12-month mean changes between the treatment groups by use of the ANCOVA linear mixed model defined above, without imputation of missing values (that is, complete-case analyses). All models were graphically checked by residual plots and quantile–quantile plots to assess model assumptions, mainly the normality assumption, and when relevant, transformed (for example, by logarithm). Furthermore, Poisson regression analysis was used to analyse the predictive effect of group intervention on reported AEs and concomitant medications.
Microbiota analysis
The complete microbiota data processing and analysis pipeline is available at https://github.com/AlexanderUm/SWEET_microbiome. Running the pipeline with the corresponding data will reproduce all of the presented microbiota analysis results and figures. In short, raw reads were preprocessed with the CASAVA pipeline (v.1.8.3) and the Quantitative Insights Into Microbial Ecology 2 (QIIME2; v.2023.9.1) platform. Demultiplexed reads were de-noised into amplicon sequence variants (ASVs) with the DADA2 plug-in51 and were taxonomically annotated with the Naive Bayes classifier trained on the SILVA (v.138) database52. A phylogenetic tree was constructed using the FastTree algorithm and MAFFT alignment. PICRUSt2 (ref. 53) was used to infer metabolic pathways using default settings, and the unstratified MetaCyc pathways abundance table was used for further pathways analysis. Phylogenetic tree, taxonomic, pathways abundance and ASV tables were imported in the R statistical and programming environment (v.4.3.2)54 using the qiime2R package55. For data transformation and visualization, the R packages tidyverse56, ggvegan57, cowplot58, broom59, ComplexHeatmap60, circlize61 and RColorBrewer62 were used. Before further analysis, ASVs with fewer than 50 reads across all samples and those taxonomically assigned to mitochondria or chloroplasts or to kingdoms other than Bacteria or Archaea, as well as those not assigned to any phylum, were removed from the dataset. An appropriate normalization of ASV counts, as specified below, was applied in correspondence with the performed analysis or visualization.
Microbial alpha diversity metrics (Chao1, observed species, Shannon and Simpson indexes) were calculated, and a linear mixed-effects regression analysis (LMM) was performed as implemented in the MicrobiomeStat63 R package for all indexes except observed species. For observed species, a generalized linear mixed-effect model (GLMM) for a Poisson distribution (log link function) was performed using the lme4 package64. Alpha diversity metrics were calculated using the unfiltered ASV table rarefied at even depth (40k). The LMM and GLMM used intervention group (main effect), time of sampling (M0, M2, M6 and M12) (time variable), intervention centre (adjustment variable) as fixed effects and subject ID as the random effect.
Differences in overall microbial composition between intervention groups were assessed using distance-based redundancy analysis (dbRDA) followed by an ANOVA-like permutation test (PERMANOVA) as implemented in the vegan65 R package. The following model was used for dbRDA: DistanceMatrix ~ Time × IntervetionGroup + Condition(Country). The PERMANOVA was performed with 999 permutations and by model terms. Dissimilarity distances (Jaccard, Bray–Curtis, unweighted and weighted UniFrac) for dbRDA were calculated as implemented in the phyloseq66 R package from the ASV count table with total-sum scaling and log2-transformed count.
Differences in individual taxa and metabolic pathway abundance trends over time between intervention groups were tested using the LinDA method as implemented in the MicrobiomeStat R package. Differential abundance was assessed at the genus and family taxonomic levels; in addition, inferred MetaCyc pathways and default count normalization implemented in LinDA was used. Before differential abundance analysis, features with prevalence less than 50% were removed from the dataset. The P values calculated for the LMM were adjusted for multiple testing with false discovery rate correction, and q values less than or equal to 0.1 were considered significant. The LMM model used for differential abundance analysis was identical to the model used for analysis of alpha diversity trends.
Responders and non-responders were identified based on body weight regain (WM index = (weight at CID4 − weight at CID2)/(weight at CID1 – weight at CID2)) or changes in HbA1c and fasting glucose during the WM period (CID4 – CID2)67 (Extended Data Fig. 1a). Microbial composition at each CID was then used to predict the response with a random forest. For classification, genera with a minimum prevalence of 50% in any treatment group were normalized using total-sum-scaling and then log2-transformed. The random forest model was built as implemented in the randomForest package68 and validated using the caret69 package. The random forest models were built with 1,999 trees, and the number of variables per split (mtry) was set to the default. Classification accuracy was assessed with 25 times repeated fivefold cross-validation as implemented in the caret package, and AUC was estimated. As additional validation, the response variable was permuted, and the accuracy of the corresponding random forest model was estimated. This process was repeated ten times, and the results were compared with the original model. The significance of the feature’s contribution to the classification model was evaluated using the permutation approach (199 permutations) as implemented in the rfPermute70 package. The pROC70 package was used to build ROC curves.
Reporting summary
Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.
Source link
