care.diabete de la Cuesta-Zuluaga and Associates sing 14 with ty (Qiagen,Hilden,Germany)following metformin (T2D-met"),14 with type 2 di- plemented in the Vegan package of R(22). the es (ND (T2-met) We ine be ingstep with the lysis buffer (0at agnostically identify microbial bic 15 Hz using a stainless stee with a stical An markers.LEfS e uses the nonparam d clinical variables ified DNA e compared across study groups us e noDrop spectrophotometer (Ny ing ANOVA and t tests after check ing fo differential abundance amon ote ns,Fra nce)and sent it to th par then p of Michigan Medical School (Ann Art ants using the unpaired wil MI).The V4 hypervariable region of the lity tests sum test,and final lly uses LD est transformed (natural log for unco and R806 (S'-GGACTACHVGGGTWTC ined variables or ar quare roo of lEfse com with sta statis -3)p rs and equenced sing for p rati tica approaches is that, y and the Statistical analyses were formed ing strategy(19).n addition t with Rv..3.2.2(21】 een each OTU and the grou g cat we sequ the QIamp DNA S 's EB el median 28,699).To limit the effects o nicrobial biomarkers in ou tio an hhe da etained f they had a A 0.05 and 0g10 D M e 12D-me na ment's run. quences were curated ugh Finally,across groups,we tested for dif foll ving th eg sta a erences in e the likeli 13620 of detecting false positives.especially maior butyrate. microbia seq tedat the NCBIand mic units a,including Butyrivibni can be th he a nd com as assessed by quantifying and ssified in each of thes ion of Type 2 iab and preting (4 for A.muciniphil 11 for that had typedi;6self-reported and the erence using ANOVA nd t te te t n We compared the nine whet (fasting bloc 2126 mg/d ne analysis。 bunda of the 28 only ir formin treatment,14 not (1 was Diversity using Results from lEfSe and from the pooled treat with insutin alo with gli d g mu of ph otype e corrected fo logical treat nent for =0 5)calcr including the 2 participants unaware of ated with the GUniF ge ofR(23 age of R(25).Tests were considered betes status)(supplementary a P value. We matched each participant with di nce per OTU.then obtained a dis with thr matrix from pairwis (to th finally eda rela cf20nerP0iemtheohg ndrNSteristiG difference between case and control joining phylogenetic tree using Mothur pants.There were no statistically signif mong en es (a ght or ohese)This left us with atotal sing the adonis function (ANOVA using cal narameters hetw en T2D-met*andTotal microbial DNA was extracted using the QIAamp DNA Stool Mini Kit (Qiagen, Hilden, Germany) following the manufacturer’s instructions, with a slight modification consisting in a bead￾beating step with the lysis buffer (20 s at 15 Hz using a stainless steel bead with a 5-mm diameter). After extraction, we quantified DNA concentration using a NanoDrop spectrophotometer (Nyxor Biotech, Paris, France) and sent it to the Microbial Systems Laboratory, University of Michigan Medical School (Ann Arbor, MI). The V4 hypervariable region of the 16S rRNA gene was amplified using the F515 (59-CACGGTCGKCGGCGCCATT-39) and R806 (59-GGACTACHVGGGTWTC TAAT-39) primers and sequenced using the Illumina MiSeq sequencing platform with V2 chemistry and the dual-index sequencing strategy (19). In addition to DNA from fecal samples, we sequenced negative controls (ultrapure water and the QIAamp DNA Stool Mini Kit’s EB elu￾tion buffer), a DNA extraction blank, and a mock community (HM-782D, BEI Resources, Manassas, VA) in each instru￾ment’s run. Sequences were curated following the MiSeq standard operating procedure implemented by Mothur v.1.36 (20) (see Supplementary Data). Raw sequences were deposited at the NCBI and can be accessed through the BioProject (accession number PRJNA325931). Definition of Type 2 Diabetes and Selection of Control Subjects We identified 28 participants in our study that had type 2 diabetes; 26 self-reported physician-diagnosed diabetes prior to the beginning of the study and 2 were diagnosed through laboratory testing (fasting blood glucose $126 mg/dL and HbA1c$6.5%). Of the 28 participants with type 2 diabetes, 14 were under met￾formin treatment, 14 were not (1 was treated with insulin alone, 2 with gliben￾clamide, and 11 were under no pharmaco￾logical treatment for type 2 diabetes, including the 2 participants unaware of their diabetes status) (Supplementary Table 1). We matched each participant with di￾abetes with three participants without diabetes in our study based on sex, age (to the closest possible age; maximum difference between case and control subjects 6 years; mean 1.5 years; median 1 year), and BMI category (lean, over￾weight, or obese). This left us with a total analytic sample of 112 study participants comprising 14 with type 2 diabetes using metformin (T2D-met+ ), 14 with type 2 di￾abetes not using metformin (T2D-met2), and 84 without diabetes (ND). Statistical Analyses Anthropometric and clinical variables were compared across study groups us￾ing ANOVA and t tests after checking for homoscedasticity and normal distribu￾tion of residuals (using Fligner-Killeen tests of homogeneity of variances and Shapiro-Wilk normality tests). When necessary, variables were appropriately transformed (natural log for uncon￾strained variables or arcsin square root for proportions). Sex ratio and stool con￾sistency were compared using x2 tests. Statistical analyses were performed with R v.3.2.2 (21). Curated DNA sequences ranged from 69 to 102,660 sequences per sample (median 28,699). To limit the effects of uneven sampling, we rarefied the data set to 4,091 sequences per sample, result￾ing in the exclusion of one T2D-met2 par￾ticipant with 69 reads. Although rarefaction may lead to missing low-abundance data, it is a powerful way to reduce the likelihood of detecting false positives, especially among those operational taxonomic units (OTUs) with very low abundance. The gut microbiota structure and com￾position was assessed by quantifying and interpreting similarities based on intra￾and intergroup diversity analyses (a and b diversity, respectively). For a diversity, we calculated Good’s coverage and the number of OTUs of each sample using Mothur and constructed rarefaction curves.We compared these indices among groups of participants using analysis of similarity (ANOSIM) with 1,000 permuta￾tions using the Vegan package of R (22). b Diversity was assessed using phylogeny-based generalized UniFrac dis￾tances (with the a parameter controlling weight on abundant lineages = 0.5) calcu￾lated with the GUniFrac package of R (23). For this, we first reduced the alignment and the OTU table to one representative sequence per OTU, then obtained a dis￾tance matrix from uncorrected pairwise distances between aligned sequences, and finally constructed a relaxed neighbor￾joining phylogenetic tree using Mothur and Clearcut. Comparisons among groups of participants were performed using the adonis function (ANOVA using distance matrices) of the permutational multivariate ANOVA (PERMANOVA) im￾plemented in the Vegan package of R (22). We next used linear discriminant analysis (LDA) effect size (LEfSe) (24) to agnostically identify microbial bio￾markers. LEfSe uses the nonparametric factorial Kruskal-Wallis sum rank test to detect individual OTUs with significant differential abundance among groups of participants, then performs a set of pairwise tests among groups of partici￾pants using the unpaired Wilcoxon rank sum test, and finally uses LDA to esti￾mate the effect size of each differen￾tially abundant OTU (24). The strength of LEfSe compared with standard statis￾tical approaches is that, in addition to providing P values, it provides an estima￾tion of the magnitude of the association between each OTU and the grouping cat￾egories (e.g., metformin, type 2 diabetes) through the LDA score. For stringency, microbial biomarkers in our study were retained if they had a P , 0.05 and a (log10) LDA score $3, i.e., one order of magnitude greater than LEfSe’s default. Finally, across groups, we tested for dif￾ferences in relative abundance of the mucin-degrading A. muciniphila and major butyrate-producing microbial genera, including Butyrivibrio, Roseburia, Subdoligranulum, and Faecalibacterium. For this analysis, we pooled all OTUs classified in each of these phylotypes (4 for A. muciniphila, 11 for Butyrivibrio, 4 for Roseburia, 10 for Subdoligranulum, and 5 for Faecalibacterium) and tested for differences using ANOVA and t tests on arcsin square root transformed relative abundances. This pooling served to examine whether differences in relative abundance of these groups of bacteria occurred across all OTUs or only in specific OTUs. Results from LEfSe and from the pooled analysis of phylotypes were corrected for multiple testing using the Bayesian ap￾proach implemented in the qvalue pack￾age of R (25). Tests were considered significant if they had a P value # 0.05 and a q value #0.1. RESULTS In Table 1 we present the characteristics of T2D-met+ , T2D-met2, and ND partic￾ipants. There were no statistically signif￾icant differences (all P values . 0.10) in demographic, anthropometric, or clini￾cal parameters between T2D-met+ and care.diabetesjournals.org de la Cuesta-Zuluaga and Associates 3