Gut microbiome and stages of diabetes in middle-aged adults: CARDIA microbiome study

Study design

The CARDIA study is a multicenter, longitudinal cohort study of 5115 White and Black men and women from four US metropolitan areas: Birmingham, AL; Chicago, IL; Minneapolis, MN; and Oakland, CA. The details of its design are described elsewhere [15]. Participants were aged 18 to 30 years at baseline in 1985–1986 (Y0) and attended follow-up exams in years 2, 5, 7, 10, 15, 20, 25, and 30 (Y2–Y30) after baseline, with 71% retention among the surviving cohort at Y30. As part of the ongoing cohort study, 615 participants were recruited into a microbiome sub-study at Y30, described briefly below and in detail elsewhere [16]. The comparison in sample characteristics between those included and not included in the microbiome study is presented in Additional file 1: Table S1. All CARDIA field centers received their respective institutional review board approvals, and participants provided written informed consent to all study components at each exam.

Gut microbiome data collection, assay, and preprocessing

Briefly, we followed standard protocols for collection and processing of stool samples [17, 18], as previously described [16]. Participants completed the stool collection in their home using collection tubes pre-filled with RNAlater, along with a short survey pertaining to covariates relevant for the microbiome study, and shipped their samples with provided ice packs and insulated shipping containers and completed questionnaire overnight to the study lab at the Nutrition Research Institute at the University of North Carolina, Chapel Hill, where samples were stored at − 80 °C until processing.

DNA was extracted from 0.2 g of stool using the MoBio PowerSoil kit (or Qiagen DNeasy PowerSoil after the purchase of MoBio by Qiagen). The V3–V4 hypervariable regions were amplified and sequenced using the Illumina MiSeq platform (2 × 300). Forward sequences were processed (quality trimming, denoising, and chimera-removal) through the divisive amplicon denoising algorithm (DADA2) package in R14. The DADA2-formatted Silva database (silva_nr_v138_train_set.fa.gz) was used to assign taxonomy [19].

Assessment of diabetes-related characteristics

Insulin resistance, diabetes duration, and stages of diabetes were the primary diabetes-related characteristics in the present study (Additional file 1: Fig. S1). We used the Homeostatic Model Assessment of Insulin Resistance (HOMA-IR) as a surrogate measure for insulin resistance. HOMA-IR was calculated as follow: [fasting insulin (uU/mL) × fasting glucose (mmol/L)]/22.5 [20]. The average of Y25 and Y30 HOMA-IR was used in the analyses.

Pre-diabetes and T2D were identified according to American Diabetes Association (ADA) criteria [21]. T2D was determined based on the presence of any of the following: a fasting serum glucose (FSG) ≥ 126 mg/dL (available at Y0, Y7 and afterward), or a 2-h (2 h) post-load glucose (2 h-PG) ≥ 200 mg/dL during a 75-g oral glucose tolerance test (available at Y10, Y20 and Y25), or a hemoglobin A1C (HbA1c) ≥ 6.5% (available at Y20 and Y25), or self-report of diabetes medications (e.g., oral hypoglycemic medications or insulin) use [22]. Similarly, pre-diabetes was defined as having a FSG of 100–125 mg/dL, or a 2 h-PG 140–199 mg/dL, or an HbA1c 5.7–6.4% in both Y25 and Y30 and no report of diabetes and no use of diabetes medications across 9 exams.

The duration of diabetes was indicated by the number of years of diabetes and calculated based upon the presence of diabetes at each exam beginning at Y2 [22]. For example, a participant who developed diabetes at Y10 was assigned a total of 20 years as the cumulative duration of diabetes, while a half of year was assigned to participants who developed diabetes at Y30. We characterized diabetes stages into four groups (1) normal; (2) pre-diabetes, (3) diabetes patients without treatment (T2D−, who did not receive treatment at both Y25 and Y30), and (4) diabetes with treatment (T2D+, who received treatment at either Y25 or Y30).

Covariates and confounders

Possible confounders of the associations between gut microbiome and aforementioned diabetes characteristics were identified from the literature [2, 12]. The majority of covariate measures were collected at the Y30 exam, with missing values replaced by Y25 covariates. Age, sex, race, highest educational attainment (high school or less, college, or graduate school), current smoking status (yes/no), current alcohol use (yes/no), and medication use (yes/no), including proton pump inhibitor (PPIs), antihypertensive and lipid-lowering, were assessed through self-reported questionnaires. Body mass index (BMI), resting systolic blood pressure (SBP), and resting diastolic blood pressure (DBP) were collected by trained staff according to a standard protocol. A total physical activity score was calculated based on the Physical Activity Questionnaire [23]. Diet quality score was derived from the interviewer-administered Diet History at the Y20 Exam, as previously described [24].

Participants in the current analyses

In the microbiome sub-cohort of the 615 participants, 607 had viable DNA samples for sequencing. From these 607 participants, we excluded one participant who had diabetes at baseline and one participant with missing smoking status at both Y25 and Y30 exams, resulting in 605 participants for analyses on gut microbiome and diabetes duration and stages of diabetes. For analyses of insulin resistance, we additionally excluded one participant without data on insulin resistance at Y25 or Y30, yielding an analytic sample of 604 (Additional file 1: Fig. S2).

Statistical analyses

We examined associations between gut microbiome composition, measured by within-person α diversity and between-person β diversity, and insulin resistance, diabetes duration, and stages of diabetes and specific taxa with the set of diabetes-related characteristics. We focused our primary analysis on genera, the lowest level of taxonomy from our data.

The α diversity (Shannon index and richness) and β diversity (Bray–Curtis index) at the genus level were calculated using the R package vegan [25]. The α diversity represents the complexity of composition within members of a group. In general, high α diversity is favorable to our health. We calculated α diversity measures using raw genera counts. The β diversity represents the similarity of microbial composition between groups of interest, with high β diversity indicating low similarity. For β diversity analysis, raw genera counts were transformed as log10[(RC/n)(x/N) + 1], where RC is the total raw count for a participant, n is the total count across all genera for a participant, x is the total across all taxa and participants, and N is the total number of participants, as previously described [16, 26]. To investigate difference in β diversity between groups, HOMA-IR was reclassified into two groups based on the median (i.e., ≤ median [2.19], and > median), while diabetes duration was reclassified into three groups (i.e., normal/pre-diabetes, newly diagnosed diabetes [duration < 5 years], and established diabetes [duration ≥ 5 years]).

The associations of α diversity measures with insulin resistance, diabetes duration, and stages of diabetes were assessed by linear regression, adjusting for four sets of covariates sequentially. In model 1, we adjusted for the sequencing run. In model 2, age, sex, race, education level, and field center were added. In model 3, we additionally adjusted for smoking, alcohol use, BMI, physical activity, and diet quality score. Last, in model 4 (the fully adjusted model), we additionally adjusted for the use of PPIs and lipid-lowering drugs. We analyzed associations of β diversity with newly categorized insulin resistance, diabetes duration, and stages using permutational multivariate analysis of variance (PERMANOVA) with covariate adjustment; a p-value was generated through 1000 permutations. To examine post hoc pairwise comparisons, we conducted additional PERMANOVA tests for each pair within categorized diabetes duration and stages. For visualization, principal coordinates analysis (PCoA) based on the Bray–Curtis dissimilarity matrix was applied. We present the first two dimensions from the PCoA according to two groups of HOMA-IR, three categories of diabetes duration, and four diabetes stages. In both α and β diversity analyses, statistical significance was set at a two-tailed p < 0.05.

To limit the possibility of spurious findings due to rare taxa, we restricted analyses to those individual taxa with non-zero counts in at least 75% of participants [16]. As a result, the taxa-specific analysis was based on 107 out of initially 375 genera. The log-transformed genera counts (described above) were used for the analyses. Multivariable linear regression models with the same sets of covariates (described above), were conducted to examine the association of diabetes-related characteristics with microbial taxa abundance. To adjust the p-value for multiple comparisons, we used the Benjamini–Hochberg method for false discovery rate (FDR). In the taxa-specific analysis, statistical significance was set to FDR-adjusted p-value (q-value) < 0.1 [27]. Data analysis was conducted in RStudio version 1.3.959 with R version 4.1.0 (http://www.r-project.org) and SAS version 9.4 (SAS Institute Inc, Cary, NC).

Sensitivity analyses

We conducted two sets of sensitivity analyses based on model 4. First, we added SBP, DBP, and antihypertensive medication use (binary) (model 5). For insulin resistance and diabetes duration analyses, we also investigated whether diabetes medication use (binary) attenuated the main associations of interest by additionally adjusting for use of diabetes medicines (model 6); while for diabetes stages analysis, we further adjusted for diabetes duration (model 6).

留言 (0)

沒有登入
gif