Donor metabolic characteristics drive effects of faecal microbiota transplantation on recipient insulin sensitivity, energy expenditure and intestinal transit time

Objective Bariatric surgery improves glucose metabolism. Recent data suggest that faecal microbiota transplantation (FMT) using faeces from postbariatric surgery diet-induced obese mice in germ-free mice improves glucose metabolism and intestinal homeostasis. We here investigated whether allogenic FMT using faeces from post-Roux-en-Y gastric bypass donors (RYGB-D) compared with using faeces from metabolic syndrome donors (METS-D) has short-term effects on glucose metabolism, intestinal transit time and adipose tissue inflammation in treatment-naïve, obese, insulin-resistant male subjects. Design Subjects with metabolic syndrome (n=22) received allogenic FMT either from RYGB-D or METS-D. Hepatic and peripheral insulin sensitivity as well as lipolysis were measured at baseline and 2 weeks after FMT by hyperinsulinaemic euglycaemic stable isotope (2H2-glucose and 2H5-glycerol) clamp. Secondary outcome parameters were changes in resting energy expenditure, intestinal transit time, faecal short-chain fatty acids (SCFA) and bile acids, and inflammatory markers in subcutaneous adipose tissue related to intestinal microbiota composition. Faecal SCFA, bile acids, glycaemic control and inflammatory parameters were also evaluated at 8 weeks. Results We observed a significant decrease in insulin sensitivity 2 weeks after allogenic METS-D FMT (median rate of glucose disappearance: from 40.6 to 34.0 µmol/kg/min; p<0.01). Moreover, a trend (p=0.052) towards faster intestinal transit time following RYGB-D FMT was seen. Finally, we observed changes in faecal bile acids (increased lithocholic, deoxycholic and (iso)lithocholic acid after METS-D FMT), inflammatory markers (decreased adipose tissue chemokine ligand 2 (CCL2) gene expression and plasma CCL2 after RYGB-D FMT) and changes in several intestinal microbiota taxa. Conclusion Allogenic FMT using METS-D decreases insulin sensitivity in metabolic syndrome recipients when compared with using post-RYGB-D. Further research is needed to delineate the role of donor characteristics in FMT efficacy in human insulin-resistant subjects. Trial registration number NTR4327.


Hyperinsulinaemic euglycemic clamp with stable isotope tracers and resting energy expenditure
Subjects were admitted at the metabolic unit after an overnight fast for the metabolic flux measurements [2]. Resting energy expenditure (REE) using indirect calorimetry was determined and a catheter was inserted into an antecubital vein for infusion of stable-isotope tracers, insulin, and glucose. Another catheter was inserted into a contralateral hand vein and kept in a thermoregulated (60°C) clear plastic box for sampling of arterialized venous blood.
The exact methodology of the two-step hyperinsulinaemic euglycaemic clamp and the calculation of isotope enrichments and endogenous glucose production (EGP) can be read in the supplementary methods.
At t = 0 h, blood samples were drawn for determination of background enrichments. Then, a primed continuous infusion of isotopes was started: [6,6-2 H2]glucose (prime: 8.8 μmol/kg; continuous: 0.11 μmol · kg −1 · min −1 ), [1,1,2,3,3-2 H5]glycerol (prime: 1.6 μmol/kg; continuous: 0.11 μmol · kg −1 · min −1 ) and continued until the end of the study. After a 2-h equilibration period (14 h of fasting), three blood samples were drawn for isotope enrichments and one sample for glucoregulatory hormones and free fatty acids (FFA). Thereafter (t = 3.0 h), a 2-step hyperinsulinemic euglycemic clamp was started: step one included an infusion of insulin at a rate of 20 mU · m −2 · min −1 (Actrapid 100 IU/mL; Novo Nordisk Farma BV, Alphen a/d Rijn, Netherlands) to assess hepatic insulin sensitivity. Glucose 20% was started to maintain a plasma glucose concentration of 5 mmol/L. [6,6-2 H2]Glucose was added to the 20% glucose solution to achieve glucose enrichments of 1% to approximate the values for enrichment reached in plasma and thereby minimizing changes in isotopic enrichment due to changes in the infusion rate of exogenous glucose [3]. Plasma glucose concentrations were measured every five minutes at the bedside. Five blood samples were drawn at five minute intervals for the measurement of glucose concentrations and isotopic enrichments. Another blood sample was drawn for measurement of glucoregulatory hormones and FFA. Hereafter, insulin infusion was increased to a rate of 60 mU · m −2 · min −1 (step 2) to assess peripheral insulin sensitivity. After another 2 h (t = 7 h), blood sampling was repeated. Endogenous glucose production (EGP) and the peripheral glucose uptake of glucose (Rd) were calculated by using modified versions of the Steele Equations, as described previously [3]. EGP and Rd were expressed as μmol · kg −1 · min −1 .
Total triglyceride hydrolysis/lipolysis (glycerol turnover) was calculated by using formulas for steady state kinetics adapted for stable isotopes and was expressed as μmol · kg −1 · min −1 [4].

Profiling of fecal microbiota composition by sequencing of the 16S rRNA gene
Fecal genomic DNA was isolated from 100 mg of feces using repeated bead beating and purification in the QIAcube (QIAGEN) similar to what previously described, with bead beating at 5.0 m/s for 60 s in a FastPrep®-24 instrument (MP Biomedicals) [5]. Fecal microbiota composition was profiled by sequencing the V4 region of the 16S rRNA gene on an Illumina MiSeq instrument (llumina RTA v1.17.28; MCS v2.5) with 515F and 806R primers designed for dual indexing [6] and the V2 Illumina kit (2x250 bp paired-end reads). 16S rRNA genes were amplified in duplicate reactions as previously described [5].

Pre-processing of 16S rRNA amplicon sequence data
Raw fastq files were quality controlled using FastQC (version 0.11.5). Illumina paired-end reads were processed using the Mothur pipeline (version 1.39.5). After merging forward and reverse reads, contigs were screened to ensure absence of ambiguous bases and a length between 252 and 254 bases. Contigs were aligned to the Silva reference (version 128), dereplicated and subsequently preclustered (allowing a maximum of two differences).
Singletons (sequences with an abundance of one in the entire dataset) were removed. The remaining sequences were chimera-filtered with chimera.vsearch and classified using the Silva taxonomy reference (version 128) with a 80% confidence cutoff. Sequences classified as mitochondria, chloroplasts, eukaryota, as well as un-classified sequences were removed.
Remaining sequences were clustered using the Opticlust distance-based algorithm to 97% OTUs. A phylogenetic tree was constructed with the "double-precision" build of FastTree 2.1 using the abundance-based representative sequences of the OTUs [7]. The resulting OTU table was rarefied to 60 000 counts per sample using the phyloseq package in R. OTUs with a mean relative abundance of less than 0.001% (i.e. a mean of less than one count per two samples) were filtered out. The final dataset was comprised of 801 OTUs in 79 samples.

Machine learning models
Elastic Net regularized classification models [8] with stability selection [9] were implemented in Python 2.7 (www.python.org) as a feature selection tool. In this paper we used the elastic net model which has been widely used for performing feature selection in biological data [10,11]. Elastic net is an embedded sparsity inducing feature selection approach, which combines the sparse regularization and stability selection procedures. This method is a regularization-based method which combines the advantages of two methods such as lasso and ridge regression [8]. Its L2-norm based shrinkage encourages highly correlated features to have similar weights, whereas L1-norm based shrinkage encourages a sparse solution.
Input data (i.e. OTU predictors) were used to predict class membership. For each analysis, input OTUs were filtered to a minimum abundance of at least 10 counts per sample in respective subset of subjects. To train each model, the two hyperparameters (the alpha -the size of the regularization penalty, and the L1 ratio -the ration of L1-norm / L2-norm in the model penalty) were optimized using a 4-fold Cross-Validation procedure on a subset comprising 80% of the data. The model was then tested on the remaining 20% of the data not used in the training. This procedure was repeated 100 times per analysis, each time using different random splittings of the data into train and test subsets. The stability of each feature was calculated as the number of times (out of 100) the respective feature was kept by the model (i.e. the number of times out of 100 runs that the feature had a non-zero regression coefficient). A minimum stability threshold of 60% was applied to all selected features.