Experimental design and study participants
The study was designed as a non-blinded, non-randomised, one-armed crossover study. Participants were studied for 18 weeks, divided into a run-in period, an intervention period and a wash-out period, each lasting 6 weeks. During the trial period (July 2015 to February 2016), participants were examined at five visits (Fig. 1a) at Frederiksberg Hospital, Denmark.
Participants were eligible for enrolment if they were men, age 18–35 years, had a BMI between 18.5 kg/m2 and 27.5 kg/m2, HbA1c < 39 mmol/mol (5.7%) and plasma creatinine within the normal range. Inclusion criteria included no prescribed medication, including antibiotics, for 2 months prior to recruitment, no previous gastrointestinal surgery discounting appendectomy or chronic illness requiring medical treatment. Participants were instructed not to change dietary habits and lifestyle during the study period.
The intervention was initiated after the second visit, 6 weeks after inclusion. All study participants were instructed to take metformin according to a fixed schedule: 500 mg once daily for the first week, 500 mg twice daily for the second week, 1000 mg+ 500 mg daily for the third week and 1000 mg+ 1000 mg daily for the remaining 3 weeks of the intervention period.
The study was designed for 25 men starting the intervention, expecting a 20% dropout.
The trial was conducted according to the International Conference on Harmonization’s Good Clinical Practice guidelines including the Declaration of Helsinki II. The study was approved by the Ethical Committees of the Capital Region of Denmark (protocol ID: H-7-2014-012) and was registered at www.clinicaltrialsregister.eu (no. 2015-000199-86) and at ClinicalTrials.gov (NCT02546050). All participants gave written informed consent before taking part in the study. No changes to methods were made after trial commencement. The trial was concluded when all the participants had finalised their last visit. No serious adverse events were reported during the trial.
Participants underwent clinical examination five times during the trial: at baseline (visit 1), at the end of the run-in (week 6; visit 2), 3 weeks into the intervention period (week 9; visit 3), at the end of the intervention period (week 12; visit 4) and at the end of the wash-out period (week 18; visit 5). Gastrointestinal symptoms (overall abdominal discomfort and degree of abdominal pain, bloating, constipation, diarrhoea, flatulence, metal taste, nausea and stool consistency satisfaction) were evaluated at all five visits using a digital visual analogue scale (VAS) and recorded as an integer between 0 and 100.
At visits 1, 2, 4 and 5, respectively, participants were examined with height, weight, hip and waist circumference, BP and bioimpedance. Detailed information on anthropometrics can be found in electronic supplementary material (ESM) Methods. A questionnaire on quality of life, physical activity and health status was completed at baseline.
Blood samples were collected at all visits after participants had fasted for 10 h overnight. Alanine aminotransferase and cobalamin were analysed using an enzymatic slide test, and creatinine, glucose, cholesterol, HDL cholesterol, and triglyceride were measured using a colorimetric slide test. LDL cholesterol was calculated using the Friedewald formula . HbA1c was measured using liquid chromatography and insulin was measured using immunoassay. Leukocytes and differential white blood cell count were measured using flow cytometry. Plasma metformin concentration was measured using high performance liquid chromatography followed by tandem mass spectrometry to ensure compliance. Detailed information on biochemical measurements is presented in the ESM Methods. HOMA of insulin resistance and beta cell function were calculated from fasting plasma concentrations of glucose and insulin according to Matthews et al .
Participants were divided into two groups based on change in reported overall abdominal discomfort. An increase in VAS score at visit 3 exceeding 2 × SEM at baseline was considered significant and used as a cut-off. One participant who dropped out before visit 3 due to severe gastrointestinal discomfort was counted among those who developed gastrointestinal adverse effects in spite of missing VAS data. By this definition, 10 of 23 participants experienced an increase in overall abdominal discomfort.
Faecal samples and data processing
Participants were instructed to deliver nine faecal samples throughout the study period. Samples were collected on the day of examination if possible and immediately frozen either at the study site or at home, in which case samples were transferred to the laboratory on dry ice. Samples were stored at −80°C until DNA extraction.
A total of 206 faecal samples were collected but two samples did not contain sufficient material for DNA extraction. Genomic DNA was isolated, followed by PCR amplification of the V4 hypervariable region of the bacterial 16S rRNA gene and sequencing on an Illumina MiSeq platform generating a total of 9,466,021 (mean 43,622 [SD 18,700]) paired-end (250 bp) reads. See ESM Methods for detailed information. Twelve samples that failed during the first run were resequenced. Processing of raw sequencing data was performed using the dada2 (v1.4.0) R package . Following inspection of quality profiles, denoising, merging of reads and removal of chimeric sequences, a total of 7,214,117 reads (mean 34,702 [SD 12,889]; minimum 10,736) in 1764 unique amplicon sequence variants (ASVs) were available for downstream analyses. Taxonomical assignment of ASVs from kingdom to species was performed against the Silva v128 database, using the dada2 implementation of the naive Bayesian RDP classifier .
The primary outcome was compositional change in abundance of ASVs agglomerated at the taxonomic level of genus. Pre-specified secondary outcomes were as follows: compositional change in abundance of non-agglomerated ASVs; change in intra-individual diversity quantified by the number of observed ASVs (richness), Shannon’s entropy and Pielou’s evenness (Shannon’s entropy/loge[richness]); change in community structure and membership quantified by Canberra and Jaccard distances, respectively, and change in anthropometric and biochemical traits. Post hoc analyses included random forest (RF) classification of individuals developing gastrointestinal discomfort and testing for bacterial genera and ASVs responding differently to metformin in these participants, compared with those who did not develop gastrointestinal side effects. Data from 24 participants examined at least once during the intervention period were included in these subanalyses. Data from participants not following protocol (e.g. reduced metformin intake, n = 2) were omitted from analyses from the time they deviated from protocol. In total, 25 participants completed visit 1 and were included in the statistical analyses, but one participant dropped out before initialisation of the intervention.
ASV abundances were agglomerated at genus level using the phyloseq R package, normalised using total sum scaling, and log transformed following addition of a pseudocount corresponding to the lowest non-zero normalised abundance of each taxon. Intervention effect was modelled by a one-way repeated measures ANOVA using mixed linear regression as implemented in the lme4 R package. Models were fitted using restricted maximum likelihood using samples from all time points. The abundance of each genus at each time point during the metformin intervention was compared by a post hoc t test with the mean abundance averaged over the three samples collected during the pre-intervention period. Genera never present in >80% of participants were excluded from analyses. Intervention effect on non-agglomerated ASV abundances, bacterial richness, Shannon’s entropy and Pielou’s evenness was assessed using the same approach considering a two-tailed p value <0.05 as significant. Correction for false discovery rate was done for all genera/ASVs across all time points by the Benjamini–Hochberg method, applying a false discovery rate of 10% for significance.
Intervention effect on community structure and community membership was assessed by permutational multivariate ANOVA (PERMANOVA; vegan R Package v2.4.6) of Canberra and Jaccard distances, contrasting each intervention and post-intervention time point to the average across the three pre-intervention time points. Principal coordinate ordination was performed using the capscale function of the vegan package, specifying an unconstrained model.
RF classification was used to identify bacterial genera that at baseline were discriminative for the development of gastrointestinal adverse effects during the intervention. RF models were trained on total sum scaled abundances of 124 bacterial genera using the caret (v6.0.79) and randomForest (v4.6.12) R packages with 25 iterations of bootstrap resampling. Down-sampling during resampling was applied to address class imbalance. Performance across resamples was evaluated by the area under the receiver operating characteristic (ROC) curve and discriminant genera were ranked by the mean decrease in accuracy. Using the optimal variable settings from the naive RF classifier (mtry = 2), we applied the Boruta (v5.2.0) R package for feature selection of all-relevant discriminant genera identified across 200 repetitions using different random seeds. Selected genera were subsequently used to build an optimal RF classifier, as described above.
We used mixed linear regression to identify bacterial genera and ASVs responding differently to the metformin intervention in participants who developed gastrointestinal side effects compared with those who did not. A response profile model was specified with a main effect of time (categorical) and a time × group interaction as fixed effects and a random intercept for each participant, thereby testing the difference between groups at all time points during the intervention. Data were corrected for multiplicity as described above.
Change in clinical and biochemical traits was assessed using linear mixed model regression ANOVA as outlined above. Logarithmic transformation of the dependent variable was applied if deemed appropriate upon inspection of residual plots and normal probability plots. For VAS data, a constant of 1 was added prior to transformation. Difference in plasma metformin between visit 3 and 4 were tested with a Wilcoxon signed-rank test.
All statistical analyses were made using R v3.4.2 (https://www.R-project.org/).