Effect of broccoli sprout extract and baseline gut microbiota on fasting blood glucose in prediabetes: a randomized, placebo-controlled trial

Trial design and oversight

The trial complies with all relevant ethical regulations, and the protocol was approved by the Regional Ethics Committee of Gothenburg (433-18). It started in December 2018 and was conducted as a randomized parallel-arm placebo-controlled double-blind trial in Gothenburg, Sweden (ClinicalTrials.gov NCT03763240) in accordance with the principles of the Declaration of Helsinki and Good Clinical Practice. The study was conducted at Gothia Forum, Sahlgrenska University Hospital, Gothenburg, Sweden, by academic investigators. Funders had no role in data interpretation. The trial was monitored by an independent monitor before, during and after its completion to ensure that it was carried out according to the protocol. All authors had access to the data, were involved in the writing and editing of the paper, vouch for the completeness and accuracy of the data, and agreed to submit the paper for publication.

Participants

A random selection of members of the general population aged 35–75 years in Gothenburg, Sweden, and surrounding municipalities, who had registered addresses and Swedish personal numbers, received an invitation letter with study information and instructions on how to book a time for a screening visit. Gender was determined based on self-report and the official personal number. Participants received travel reimbursement but no other financial compensation.

Inclusion criteria

Individuals were eligible to be included in the trial if all of the following criteria applied:

• Impaired fasting glucose, defined as fasting blood glucose at 6.1–6.9 mmol l⁻¹

• Written informed consent

• Age 35–75 years; participating women of fertile age must have no current pregnancy, which was assessed by a pregnancy test

• BMI 27–45 kg m⁻²

Exclusion criteria

• Diabetes mellitus based on previous documentation or treatment with anti-hyperglycaemic medication or diagnosed according to the World Health Organization criteria (random plasma glucose >11.1 mmol l⁻¹ or fasting glucose >7.0 mmol l⁻¹ or HbA1C ≥ 6.5%)

• Anti-diabetic medication

• Active liver disease

• At screening or at any subsequent visit, a level of aspartate aminotransferase (AST) or alanine aminotransferase (ALT) of more than three times the upper limit of the normal range

• Gastro-intestinal ailments that may interfere with the ability to adequately absorb sulforaphane

• At screening visit, creatinine >130 µmol l⁻¹

• Coagulation disorder or current anti-coagulant therapy, which may be affected by BSE

• Diagnosed with a cardiovascular disease or known cardiovascular event, transient ischaemic attack, coronary by-pass surgery or other coronary vessel intervention within 6 months before enrolment

• Systemic glucocorticoid treatment

• Herbal treatment, defined as food supplement (except multivitamin treatment) with herbal or vegetable extracts that may affect blood glucose

• Participant unable to understand the study information

• Participation in another clinical trial, which may affect the outcome of the present study

• Any other physical or psychiatric condition or treatment that in the judgement of the investigator makes it difficult to participate in the study

Trial procedures

All participants signed a written informed consent before study procedures were initiated. Participants were instructed not to conduct intense physical activity or drink alcohol 24 h before the study visits. They were also instructed to fast starting midnight and not use nicotine on the same day.

At the screening visit, the height and weight of each participant were measured and venous blood samples were drawn for analysis of glucose, creatinine, AST, ALT, GGT, ALP, bilirubin, prothrombin complex and thrombocytes. Blood samples were drawn at 7.30–10.00 a.m.

Individuals with fasting blood glucose between 6.1 and 6.9 mmol l⁻¹ were invited to a second visit approximately 2 weeks later. At this visit, body weight was measured and fasting venous blood samples were drawn for analysis of primary, secondary and safety variables. At this baseline visit, stool samples were also collected.

Individuals who had fasting blood glucose between 6.1 and 6.9 mmol l⁻¹ also at the second visit were randomized to receive BSE or placebo in a double-blind manner. The data from the second visit were used as baseline measures for analyses of primary and secondary variables. If blood glucose was 7.0 mmol l⁻¹ or above, the individual was excluded and referred to primary healthcare.

The randomized participants were instructed to take BSE or placebo once daily in the morning. Concordance with treatment was noted in a diary and also checked at the final visit by counting the remaining doses. Study personnel contacted the participants by phone 2–4 weeks after the initiation of treatment to discuss concordance with treatment and side effects.

The third visit was scheduled on the same weekday as visit 2 (unless it was not possible because of public holidays) 12 weeks after the first dose of the study medication. At this visit, body weight was measured, stool samples were collected and fasting venous blood samples drawn for analysis of primary, secondary and safety variables.

Physical activity was assessed using the International Physical Activity Questionnaire (IPAQ), and dietary habits were assessed using items that had been validated in Swedish national health questionnaires⁴⁰, which the participants completed during the second and third visits.

Randomization

The randomization (in a 1:1 ratio between BSE and placebo) was generated by independent statisticians using a computer-based block randomization algorithm with balanced blocks. Allocation was concealed (via sealed envelopes) from the participants and study personnel until the end of the study. Thus, the generation of the random sequence, participant enrolment by study personnel and the allocation to randomization groups were clearly separated.

Study compounds

BSE containing high amounts of the sulforaphane precursor glucoraphanin was provided by Lantmännen R&D. BSE is a dried powder of an aqueous extract of broccoli sprouts that provides a consistent and stable source of sulforaphane. The active formulation contained BSE with maltodextrin added as a bulking agent, whereas maltodextrin alone was used as placebo. The placebo looked, smelled and tasted similar to the active compound and had the same constituents except BSE. Study doses were provided as dry mixtures in sealed, non-transparent portion-size bags. Each BSE dose delivered 150 μmol of sulforaphane. Sulforaphane content was determined using reverse-phase high-performance liquid chromatography by Eurofins. No sulforaphane was detected in the placebo. The mixtures of BSE and placebo were suspended with approximately 1 dl water and ingested orally once daily in the morning.

Safety studies of BSE in healthy volunteers have revealed no evidence of systematic, clinically significant adverse effects^11,41,42. This has been confirmed in several clinical trials with healthy volunteers as well as, for example, patients with recurrent prostate cancer, where doses of up to 400 μmol sulforaphane have been used^37,42,43. The most commonly reported side effects are indigestion, belching or loose stools^12,41,42,43.

Outcomes

Venous blood samples were taken between 7.30 and 10.00 in the morning. Fasting blood glucose from venous samples was measured at the study centre using a HemoCue Glucose System (HemoCue AB). All other blood analyses were performed at the central hospital laboratory (Gothenburg, Sweden). Homeostasis model assessment-2 estimates of insulin resistance (HOMA-IR) and beta-cell function (HOMA-B) were determined as previously described⁴⁴.

The primary variable was fasting blood glucose, and the primary objective was to test the hypothesis that BSE improves fasting blood glucose using intraindividual comparisons before (baseline) and after treatment in the BSE group relative to the placebo group. The secondary variables were the change from baseline in HbA1c, BMI, insulin resistance (measured by HOMA-IR), insulin secretion (measured by HOMA-B), fasting blood lipids and a fatty liver index based on BMI, waist circumference, triglycerides and GGT⁴⁵. Liver parameters, including GGT, ALP, AST, ALT and bilirubin, were also measured, and haemoglobin, thrombocytes, thyroid-stimulating hormone, creatinine and estimated glomerular filtration rate (based on creatinine) were analysed as safety variables. Insulin clearance was estimated using the fasting C-peptide-to-insulin ratio.

Patient-reported outcomes

Participants completed the IPAQ, which assesses intense and moderate physical activity as well as walking during the past 7 days (ref. ⁴⁶). The questionnaire was completed at baseline and at the last visit by participants. Responses were converted to metabolic equivalent task minutes per week according to the IPAQ scoring protocol⁴⁶.

Dietary habits were assessed using a food frequency questionnaire previously used in public health surveys. The self-reported frequency of intake of vegetables, lentils and root vegetables; fruit and berries; fish and shellfish; sausages; chocolate and sweets; cakes, buns and cookies; cheese; and sugared beverages was recorded and scored according to a reference indicator from the National Food Administration⁴¹. The items were summed to a total score from 0 to 9 (with 9 indicating a diet most adherent to official food recommendations).

Clustering of study participants

The data-driven clustering based on diabetes-relevant traits was conceived in the All New Diabetics In Scania (ANDIS) cohort¹⁴. ANDIS aims to register all incident cases of diabetes in Scania, which is one of the largest regions in Sweden with 1,200,000 inhabitants. Over 27,000 diabetic patients (>90% of the estimated number of eligible cases in the region) are included. The clustering is based on continuous measures of BMI, age, fasting glucose, C-peptide and HbA1c as well as the presence or absence of glutamic acid decarboxylase antibodies (GADA) as a binary variable. The method, which is described in detail in ref. ¹⁴, is based on k-means clustering and has highlighted five clusters of patients with diabetes, each with different pathophysiological characteristics^14,15.

The alignment of study participants with the clusters was performed using the baseline data of each participant. GADA was not measured in this study, but all participants were assumed to have non-autoimmune diabetes based on disease history (type 1 diabetes was an exclusion criterion in the study). The clustering was based on bootstrapping. In every round, the 8,980 individuals used to analyse the original clusters in ref. ¹⁴ were sub-sampled, such that 60% of the cohort was randomly selected and clustered. The centroid, represented by the relative coordinates of the included variables, was determined for each cluster. The study participants were then assigned to one of the clusters based on the nearest Euclidean distance to the cluster centroids. This procedure was repeated in a bootstrapping algorithm, and the number of counts for the different cluster was summed for each study participant. The fraction of repeats that a study participant was assigned to the same cluster was used to determine a cluster alignment score from 0 to 1. A score of 1 means that the participant was assigned to the same cluster in every repeat.

As age at diagnosis of diabetes is used to cluster diabetes patients¹⁴, while age at diagnosis of prediabetes (that is, age at study inclusion) was used in this trial, disease progression and risk for complications may differ, despite pathophysiological similarities between clusters. Thus, the rationale for using the clusters was not to predict complications but to examine the glycaemic response in individuals with different clinical and pathophysiological characteristics.

Replication cohort

The clustering was also performed using baseline data in a separate cohort of individuals taking part in a longitudinal study to examine the influence of lifestyle on diabetes progression (ClinicalTrials.gov NCT05006508). The study complies with all relevant ethical regulations, and the protocol was approved by the Swedish Ethics Review Authority 2021-06830-01. Study participants were recruited by advertisements in 2021–2024. Individuals above 35 years of age across Sweden were eligible to participate after giving informed consent. Those who had not been diagnosed with type 1, type 2 or secondary diabetes completed a diabetes risk assessment questionnaire (the Finnish Diabetes Risk Score questionnaire, ranging from 0 to 26 with higher scores corresponding to higher diabetes risk) at baseline. Participants with a Finnish Diabetes Risk Score at 15 or above were requested to leave blood samples for analysis of fasting glucose, C-peptide (to determine HOMA-B and HOMA-IR) and HbA1c to study their metabolic profile and better understand which factors contribute to the progression of diabetes over time. Those who were 35–75 years old with a fasting blood glucose between 6.1 and 6.9 mmol l⁻¹ at baseline and BMI 27–45 kg m⁻² (corresponding to the study criteria of the BSE trial) were clustered with the same methodology used for the participants of the BSE trial.

DNA extraction, library preparation and shotgun metagenomic sequencing of faecal samples

All study participants collected their own faecal samples at room temperature before visit 2 and 3. The faecal samples were delivered on the same day of sampling to the study centre, where they were stored at −80 °C. To use the samples to the largest extent possible, they were analysed even for participants who had non-complete clinical follow-up data. Total genomic DNA was isolated from 100–150 mg of faecal material using a modification of the International Human Microbiome Standards DNA extraction protocol Q7 (ref. ⁴⁷). Samples were extracted in Lysing Matrix E tubes (MP Biomedicals) containing ASL buffer (Qiagen), vortexed for 2 min and lysed by two cycles of heating at 90 °C for 10 min followed by two bursts of bead beating at 5.5 m s⁻¹ for 60 s in a FastPrep-24 Instrument (MP Biomedicals). After each bead-beating burst, samples were placed on ice for 5 min. Supernatants were collected after each cycle by centrifugation at 4 °C. Supernatants from the two centrifugation steps were pooled, and a 600 µl aliquot from each sample was purified using the QIAamp DNA Mini kit (Qiagen) in the QIAcube (Qiagen) instrument using the procedure for human DNA analysis. Samples were eluted in 200 µl of AE buffer (10 mM Tris·Cl; 0.5 mM EDTA; pH 9.0). Libraries for shotgun metagenomic sequencing were prepared by a PCR-free method; library preparation and sequencing were performed at Novogene on a NovaSeq instrument (Illumina) with 150 bp paired-end reads and at least 6 G data per sample.

Faecal metagenomic profiling and bioinformatic analysis

The metagenomic reads were quality filtered and trimmed using fastq_quality_trimmer from the fastX toolkit (https://github.com/lianos/fastx-toolkit/). To remove human contamination, reads were mapped against the human genome (hg19) using Bowtie2 v2.4.4 (ref. ⁴⁸). Filtered reads passing the quality criteria were then mapped using Kraken2 with default settings against the RefSeq database (release 107). Abundance estimation was performed using Bracken for all reads with a minimum read length of 100 bp. Gene count estimation was performed on a previously published gene catalogue containing 15,186,403 non-redundant microbial genes¹⁷. Kyoto Encyclopedia of Genes and Genomes ontology annotations were then performed for microbial functional profiling based on MEDUSA⁴⁹. Butyrate kinase gene (buk) representing one of the bacterial butyrate-producing pathways²⁷ was profiled using hidden Markov models to screen the gene catalogue and to identify the butyrate producers among the metagenomic species by HMMER⁵⁰.

The BT2156–BT2160 protein sequences were downloaded from RefSeq (WP_008763945- WP_00876394) and mapped against Bacteroides D2 (accession id NZ_CP102261) and Bacteroides DM10 (accession id CP060488) based on the reference genomes of the species used in the RefSeq database (release 107). The gene count estimation of BT2160, the transcriptional regulator of the operon, was performed on the gene catalogue of the non-redundant microbial genes detailed above, and statistical significance was determined based on the proportion of permutation test statistics greater than or equal to the observed statistic (using 10,000 permutations with a random shuffle function in R 4.1.0).

PCo analysis was performed on Bray–Curtis dissimilarity at the species level, calculated based on species abundances, and significance was determined by PERMANOVA test using the adonis2 function with 10,000 permutations. Significantly differential abundant species tables were obtained using the deseq2 package with adjustment for subjects at different visits. The P value adjustment for significantly altered taxa was performed by the default setting in deseq2 using the false discovery rate according to the Benjamini–Hochberg method. Correlation of gut microbiota species abundances with clinical parameters was performed using distance-based redundancy analysis with the capscale function using anova.cca and 10,000 permutations. The functions used in these analyses are implemented in the vegan package (Community Ecology Package-R package version 1.17-8). All statistical analyses involving faecal whole-genome metagenomics were performed in R 4.1.0.

Continuous baseline variables that predict response to BSE

We used XGBoost, an ensemble machine learning technique based on decision trees, to identify continuous baseline variables that predict the change in fasting glucose after treatment of the study drugs. The method develops a multivariable ensemble of prediction models that were used to identify the strongest predictors of response. The optimal values for hyperparameters for each outcome were detected by performing a grid search on several possible combinations of different variables. The hyperparameters include the number of trees, learning rate, minimal loss to expand on a leaf node, maximum tree depth and subsample proportion. All other parameters were used at their default values. The package XGBoost version 1.6.0.1 was used in R 4.1.0.

We computed the relative importance of each variable predicting the outcome using F scores in XGboost, calculated as the sum of Gini improvement among the corresponding splits within a tree averaged over all the trees. In addition, we implemented Shapley Additive Explanations (SHAP), for easy interpretation of the machine learning model output. The SHAP value in this analysis is the mean absolute individual feature-level impact on the model. The training set in our models consisted of a randomly selected subset of 80% of the study participants, and the testing set was composed of the remaining 20%. The model was based on data from the training set; the testing set was independent of the training process and was used only for performance evaluation after the model was established.

Measurement of sulforaphane in serum

The concentration of sulforaphane in serum samples from participants was measured as previously described⁵¹. The methodology is based on analysing dithiocarbamate levels in patient serum by the cyclocondensation reaction for measurement of sulforaphane and its metabolites. Absence of sulforaphane in samples from the placebo group was verified by parallel measurements of the sulforaphane concentration in serum from placebo-treated participants. The difference in the average abundance of BT2160 in baseline and post-treatment samples between participants with low and high sulforaphane concentration in serum was compared using a weighted least squares analysis, adjusted for body surface area.

Statistical analysis

The primary endpoint was the intraindividual change in fasting glucose from baseline in response to BSE or placebo, which was analysed using a linear model adjusted for BMI and variation in HOMA-IR. The comparison of fasting glucose was also complemented with an ANCOVA model. Secondary endpoints included the intraindividual change in secondary variables from baseline in response to BSE or placebo and were analysed using a linear model as for the primary endpoint. Normality was verified for the major clinical variables using normal probability plots.

The full analysis set includes all participants who have clinical measures after randomization, independent of concordance with treatment.

In view of observations in patients with type 2 diabetes that serum triglyceride concentration is associated with the response to BSE, individuals below or above the median serum triglyceride concentration were also analysed separately.

The data-driven clustering method was published after the design of this study, and the investigation of clusters is a post hoc analysis. The primary and secondary variables were compared between BSE and placebo within each cluster of participants using corresponding linear models as applied to the full cohort. The interaction between treatment and subgroup was analysed by a linear model with one term for treatment (BSE or placebo), one term for the subgroup and an interaction term for the treatment and subgroup. Baseline variables were compared between the three clusters using ANOVA followed by Bonferroni corrections to obtain an overall P value for the variation between all three groups and for pairwise comparisons between groups.

The interaction between the abundance of BT2160 (log values) and pathophysiological subgroup (MARD versus MOD and SIRD) was analysed by a linear model with one term for BT2160 abundance, one term for subgroup and an interaction term for BT2160 and subgroup with the change of fasting glucose in response to BSE as the dependent variable. The analysis was adjusted for variation in body surface area between participants using the standard Du Bois formula.

The study was designed to have 80% power to detect a treatment effect of 0.3 mmol l⁻¹ between BSE and placebo. The standard deviation of change in fasting blood glucose over 12 weeks is 0.63 mmol l⁻¹, based on analyses in our longitudinal cohorts of subjects with impaired fasting blood glucose. At alpha 0.05, at least 74 study participants were needed.

Two-sided P values of 0.05 or less were considered to indicate statistical significance. Summary statistics are generally presented as point estimates with 95% CI unadjusted for multiple comparisons. Statistical analyses were performed using SPSS (version 26, IBM) or R 4.1.0.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.