figure b

Introduction

Mounting evidence indicates an intricate relationship between the gut microbiome (GM), a complex consortium of microbes inhabiting the lower gastrointestinal tract, and type 1 diabetes. Both human cohort studies [1,2,3,4,5,6,7] and controlled animal experiments [8,9,10,11,12] have demonstrated that the GM harbours both protective and harmful features that may influence the disease process leading to type 1 diabetes. For example, microbially produced short-chain fatty acids (SCFAs) provided protection from type 1 diabetes [1, 7, 12, 13]. Further complicating such analyses, HLA allele combinations providing increased risk for type 1 diabetes also result in changes in the GM composition, potentially through host regulation and selection [14]. Recent and ongoing trials explore the possibility to modify the microbiome-linked disease risk through various interventions targeting the GM [13, 15].

The GM is also implicated in the natural course of type 1 diabetes after the diagnosis [13, 16,17,18]. For example, existing evidence suggests that the GM is involved in regulating host glycaemic control [13, 19]. An exploratory faecal microbiome transplant (FMT) trial in individuals with recent-onset type 1 diabetes found that autologous FMT halted the decline in endogenous insulin secretion [20]. Together, these emerging data suggest that GM modifications could provide possibilities to intervene in the disease process and even slow down its progression. To this end, a combination of exploratory analyses and controlled experiments is needed to identify microbial strains, metabolites and other GM features that are implicated in host physiology and type 1 diabetes-related biomarkers.

Here, we analysed the GM profiles, for associations with clinical and laboratory data, from people with newly diagnosed type 1 diabetes and unaffected autoantibody (AAB)-positive family members participating in the European, multicentre Innovative approaches to understanding and arresting type 1 diabetes (INNODIA) Natural History Study [21] during a follow-up period of 2–4 years. INNODIA has collected rich information and clinical data on the participants, including fasting C-peptide and HbA1c measurements, to assess endogenous insulin production and glycaemic control, respectively. We report associations between host glycaemic control, diabetes progression and GM features.

Methods

Study population

This study recruited two cohorts from the large INNODIA Natural History Study [21]. The first cohort comprised individuals with newly diagnosed type 1 diabetes (ND cohort) [22] and the other cohort AAB-positive unaffected family members of individuals with type 1 diabetes (UFM cohort) (Table 1). Participants were identified through adult and paediatric diabetes clinics at participating sites and recruited between November 2016 and November 2021. The sex distribution among ND participants was quite even, while there was a female preponderance among the UFM participants, which may reflect a higher willingness to participate in the study among female family members, particularly mothers of ND participants. The overwhelming majority of the participants were white. The ND participants are representative of European individuals with newly diagnosed type 1 diabetes. The UFM participants are representative of European family members of ND individuals, except for the uneven sex distribution. We do not have information on the socioeconomic characteristics of the participants.

Table 1 Clinical and demographic data of the study participants from the ND and UFM cohorts
Full size table

INNODIA selected the first 100 ND cohort individuals for in-depth molecular assays, including GM profiling. Stool samples were available for 98 ND cohort participants. The UFM cohort was included to assess possible GM shifts during the asymptomatic stage of type 1 disease. UFM participants in this study were selected based on stool sample availability. A minority of the UFM participants (10/194; 5.2%) in this study were related to the ND participants. Both ND and UFM participants tested positive for at least one diabetes-associated AAB out of the three analysed (GADA, IA-2A and ZnT8A). Insulin antibodies (IAs) were measured after the ND participants had already received exogenous insulin which is known to induce IA production, and these antibodies are not distinguishable from IAA by the assay used. BMI was calculated as the weight in kilograms divided by the square of the height in metres. Age- and sex-appropriate standard deviation scores (SDSs) were calculated using World Health Organization 2006 and 2007 data [23]. A harmonised protocol for sample collection and storage was used in the study centres [21]. The study followed the guidelines of the Declaration of Helsinki for research on human participants, and the study protocols were approved by the ethical committees of the participating hospitals. Either the parent or participants themselves gave their written informed consent. Sex of the participants was self-reported. Sex and gender were not inclusion or exclusion criteria in the study.

Stool sample collection

Stool samples were collected from the INNODIA ND cohort at baseline (within 6 weeks after the diagnosis of type 1 diabetes) and at 3, 6, 12 and 24 months later. In the INNODIA UFM cohort, stool samples were collected at 6, 12, 18, 24 and 36 months after the screening visit (baseline). The samples were collected in OMNIgene-Gut OMR-200 collection tubes (DNA Genotek, Ottawa, ON, Canada), which stabilise DNA at ambient temperature for up to 60 days. Participants were asked to collect the stool samples at home during the week preceding the next study centre visit and to bring the sample with them to the study centre. The samples were frozen at −80°C in the local study centre and shipped frozen to the INNODIA Biobank for storage at −80°C until analysis.

DNA extraction

DNA was extracted from 250 μl aliquots of the faecal samples in OMNIgene-Gut OMR-200 collection tubes using the NucleoSpin 96 Soil (Macherey-Nagel, Hoerdt, France) kit. Bead beating was done horizontally on a Vortex-Genie 2 (Scientific Industries, NY, USA) at 2700 rev/min for 5 min. One negative control and one positive control (ZymoBIOMICS Microbial Community Standard, Zymo Research, CA, USA) were included per batch of samples from the DNA extraction and throughout the laboratory process.

DNA sequencing

The quality of extracted DNA was evaluated using agarose gel electrophoresis. The quantity of DNA was measured by a Qubit 2.0 fluorometer (ThermoFisher Scientific, MA, USA). DNA was randomly sheared into fragments of 350 bp, on average. Sequencing libraries were constructed using NEBNext Ultra Library Prop Kit (New England Biolabs, Herts, UK). The quality of the DNA libraries was measured using a Qubit 2.0 fluorometer and Agilent 2100 Bioanalyzer (Agilent Technologies, CA, USA), to assess the fragment size distribution. The DNA concentration of the final library was determined by quantitative real-time PCR (qPCR) prior to sequencing. Metagenomic DNA was sequenced using a 2 × 150 bp paired-end protocol on an Illumina platform (Illumina, CA, USA).

Sequence data quality control

Quality control of sequencing reads was conducted using KneadData (v.0.6.1; https://huttenhower.sph.harvard.edu/kneaddata/) to remove low-quality reads and trim low-quality bases. Trimmed reads shorter than 100 bases and reads mapping to human genome GRCh38 were discarded.

Reference gene catalogue

The Clinical Microbiomics Human Gut HG04 gene catalogue, consisting of 14,355,839 genes, was used as a reference gene catalogue. The catalogue is based on 12,170 non-public human gut samples, 9428 publicly available metagenomes from 43 countries [24] and 3567 publicly available genome assemblies from isolated microbial strains. Taxonomic abundance profiles were obtained using the Clinical Microbiomics HGMGS version HG4.D.1 set of 2095 metagenomic species (MGS), each represented by a set of genes with highly coherent abundance profiles and base compositions in the 12,170 metagenomes. The MGS concept is described in [25]. Quality controlled reads were mapped to the gene catalogue using the BWA mem (v.0.7.16a; https://bio-bwa.sourceforge.net/) algorithm [26] with the following criteria: mapping quality (MAPQ) ≥20, sequence identity ≥95% over ≥100 bp.

MGS annotations

An MGS was taxonomically annotated by a BLAST search (https://blast.ncbi.nlm.nih.gov/Blast.cgi) of its genes against the NCBI RefSeq genome database (27 January 2020) combined with rank-specific annotation criteria. An MGS was assigned to a taxon if at least M% of its genes were mapped to the taxon and no more than D% of its genes were mapped to a different taxon. We considered only blast hits with an alignment length ≥100 bp, ≥50% query coverage and percentage identity ≥PID. Parameters M, D and PID were defined for subspecies, species, genus, family, order, class, phylum and superkingdom as follows: PID = (95, 95, 85, 75, 65, 55, 50, 45); M = (75, 75, 60, 50, 40, 30, 25, 20); and D = (10, 10, 10, 20, 20, 20, 20, 15), respectively. Finally, each MGS was processed with CheckM (v.1.1.2; https://ecogenomics.github.io/CheckM/) [27], and the annotation was updated with the CheckM result if this resulted in a lower taxonomic rank.

MGS abundance calculation

MGS abundances were calculated as described in [25]. Briefly, an abundance for each MGS was estimated using gene abundances of 100 signature genes optimised for accurate abundance profiling. MGS abundances were normalised to account for length of the signature genes and the relative abundance of the data.

Functional annotation and profiling

We measured microbiome functional modules using the gut metabolic modules (GMMs) which consist of 103 conserved metabolic pathways, each defined as a series of enzymatic steps represented by KEGG Orthology identifiers (KOs) [28]. Abundance of a functional module was calculated as the proportion of all mapped reads that mapped to a KO belonging to the given module.

Diversity estimation

The α and β diversities were estimated using rarefied MGS abundances. We used the number of entities detected (richness) and Shannon’s index to measure α diversity. The β diversity was calculated using the Bray–Curtis dissimilarity.

Statistical analysis

Associations between microbial α diversities and clinical covariates were tested using linear models. Associations between microbial β diversities and clinical covariates were tested with permutational multivariate analysis of variance (PERMANOVA) as implemented in vegan R package. Associations between individual microbial features (species, functional modules) were tested by linear mixed effect models in MaAsLin2 R package [29], assuming normally distributed data and normally distributed random intercepts per study participant. The statistical models included covariates to correct for effects related to the clinical site, sex and age at diagnosis (age at recruitment for the UFM cohort). Statistical models involving longitudinal analyses additionally included covariates to correct for the time from baseline and participant-specific random intercepts. All data analyses were conducted in R v.4.0.0 [30] in the INNODIA Cloud environment. False discovery rate (FDR) correction of all tests including multiple features was performed using the Benjamini–Hochberg procedure and the resulting q values (i.e. FDR-corrected p values) are reported when appropriate.

Results

We analysed 368 faecal samples collected from 98 individuals newly diagnosed with type 1 diabetes (ND cohort) and 492 faecal samples collected from 194 UFM participants (Fig. 1, Table 1). Participants were recruited across 25 clinical centres in 13 European countries (Table 2). The age of the ND participants varied from 1 to 38 years (mean 12.3, SD 8.64) and the cohort comprised 48 female and 50 male study participants. The average age at the diagnosis was 12.3 years (SD 8.6) and the mean diabetes duration was 3.7 weeks (SD 1.6, minimum 0.7, maximum 6.3 weeks) at the first study visit. At baseline, the average total daily insulin dose was 0.52 IU/kg (SD 0.26), HbA1c 75.3 mmol/mol (SD 23.2; 9.0% [4.0%]) and the mean fasting C-peptide level 272 pmol/l (IQR 236 [Q1, Q3: 106, 342]). The mean plasma glucose reading at baseline was 7.73 mmol/l (IQR 2.77 [Q1, Q3: 5.30, 8.08]). The mean BMI SDS was 0.41 (SD 1.11). Following the results of a previous INNODIA study, we divided the ND cohort participants into three groups based on the change of C-peptide levels over time, depicting the rate of disease progression: rapid decline, slow decline and increasing [31]. The age of the 194 UFM participants varied from 1 to 44 years (mean 21.2, SD 14.1). Clinical and demographic data of both cohorts are presented in Table 1. Stool samples were metagenomically sequenced with an average depth of 9 Gb per sample, corresponding to 30.0 million read pairs per sample (Illumina 2 × 150 paired-end). On average, 87.6% of the high-quality microbiome reads from a sample were mapped to the gut gene catalogue.

Fig. 1
figure 1

Stool sample collection and metagenomes in the ND (N=98 individuals) and UFM (N=194 individuals) cohorts within the INNODIA study. Stool samples were collected at home during the week preceding the next study centre visit. n shows the number of metagenomes generated per time point. T1D, type 1 diabetes

Full size image
Table 2 Number of participants (N) per clinical centre and study cohort
Full size table

We first analysed the metagenomes from the ND cohort to investigate links between disease biomarkers and the GM following the diagnosis of type 1 diabetes. The most prevalent and abundant GM species included several Bifidobacterium and Bacteroides species as well as Faecalibacterium prausnitzii, all common human GM members (Fig. 2). We observed a stark dichotomy in the relative abundance of Prevotella copri, which was highly abundant in a subset of the metagenomes (n=88, 23.8%) and missing in others. P. copri presence was associated with change in abundance of 42 microbiome functional modules (Wilcoxon test, FDR-corrected p<0.05, electronic supplementary material [ESM] Table 1).

Fig. 2
figure 2

Heatmap displaying relative abundances of the 30 most abundant bacterial species in N=370 GM profiles from the participants in the INNODIA ND cohort. Row and column orders were determined by hierarchical clustering using Ward’s clustering criterion

Full size image

We tested for associations between microbial α and β diversities and demographic/clinical/biochemical data at baseline in the ND cohort. We detected shifts in microbial profiles (β diversities) between the clinical centres (PERMANOVA test, R2=0.183, p=0.007). There were no detectable shifts or associations between microbial α or β diversities and participants’ sex, age, BMI SDS, insulin dose, HbA1c, fasting C-peptide measurement or C-peptide/glucose ratio at baseline (α diversities, linear mixed model, p>0.1; β diversities, PERMANOVA test, p>0.05). We observed associations between the abundance of 60 bacterial species and the clinical centres (q<0.20, ESM Table 2).

We next pooled data from the follow-up period to assess longitudinal microbiome changes over the visit schedule in the ND cohort. Twenty-one microbial species and eight functional modules showed longitudinal trends (linear mixed effects model, q<0.20, Fig. 3a,b, ESM Table 2). All these species had an increasing trend over time, and they included both common and rare species with high and low average relative abundances.

Fig. 3
figure 3

Longitudinal changes in the ND cohort during the follow-up. (a) Relative abundance and (b) prevalence of n=21 bacterial species had statistically significant longitudinal trends (linear mixed effects model, q<0.10)

Full size image

We tested for association between HbA1c values at the time of the diagnosis of type 1 diabetes and the microbiome at the baseline visit (within 6 weeks after the diagnosis) in the ND cohort. The relative abundance of F. prausnitzii was inversely correlated with HbA1c (linear mixed effects model, β-coefficient=−0.15 [95% CI −0.24, −0.058], q=0.19, nominal p=0.0019, Fig. 4a, ESM Table 3) and microbial hydrogen metabolism (functional module MF0098) was positively associated with HbA1c (β-coefficient=2.43 [95% CI 0.92, 3.94], q=0.17, nominal p=0.0020, ESM Table 3). Since participants with ketoacidosis at the diagnosis of type 1 diabetes had significantly higher HbA1c (p=5.2 × 10−5; the average HbA1c in participants with ketoacidosis was 117.0 mmol/mol [12.9%], and in participants with no ketoacidosis 93.3 mmol/mol [10.7%]), we conducted sensitivity analysis controlling for ketoacidosis. Both associations above became less statistically significant (F. prausnitzii, q=0.29, nominal p=0.0031; hydrogen metabolism, q=0.35, nominal p=0.0077), indicating that ketoacidosis explained away a fraction of the associations. Ketoacidosis status alone was not associated with any GM features at baseline in the ND cohort.

Fig. 4
figure 4

Associations between gut bacteria at baseline and clinical covariates in the ND cohort. (a) Relative abundance of F. prausnitzii was inversely associated with HbA1c values at the diagnosis of type 1 diabetes (linear model, q=0.19). (b) Relative abundance of C. eutactus correlated with the number of AABs at the baseline visit (linear model, q=0.16)

Full size image

We analysed possible associations of the clinical and laboratory data at the baseline visit (within 6 weeks from the diagnosis) with the baseline microbiome composition in the ND cohort. Tested variables included HbA1c value, insulin dose per kg, fasting C-peptide concentration, fasting C-peptide/glucose ratio and BMI SDS. We did not observe any associations between baseline measurements and microbiome features. We also tested for associations between the number of detected AABs at baseline and the baseline microbiome composition. IAs were excluded from the AAB count since the participants had already received exogenous insulin which is known to induce IA production, and these antibodies are not distinguishable from IAAs by the assay applied. We observed a positive correlation between Coprococcus eutactus and the number of detectable AABs at the baseline visit (β-coefficient=0.41 [95% CI 0.16, 0.66], q=0.16, nominal p=0.0014, Fig. 4b, ESM Table 4) and an inverse correlation between microbial tyrosine degradation and the number of AABs (β-coefficient=−0.48 [95% CI −0.77, −0.19], q=0.17, nominal p=0.0014, ESM Table 4). In addition, we analysed the data for associations between the change in clinical parameters (HbA1c, insulin dose, fasting C-peptide) during the 2 year follow-up and the baseline microbiome composition. Relative abundance of Blautia obeum was inversely associated with change in fasting C-peptide between baseline and visit 5 (β-coefficient=−0.25 [95% CI −0.38, −0.12], q=0.07, Fig. 5, ESM Table 5).

Fig. 5
figure 5

Relative abundance of B. obeum was inversely associated with the change in fasting C-peptide concentration during the first 2 years of clinical type 1 diabetes (linear model, q=0.07) in the ND cohort

Full size image

We compared the microbiome composition between the three groups of participants in the ND cohort based on the change of C-peptide levels over time and observed differences in Streptococcus salivarius (q=0.12), Campylobacter concisus (q=0.15) and Veillonella atypica (q=0.19) abundances between the groups (ESM Table 6). Bacterial α diversity, measured by richness (number of observed species), differed between the groups (repeated measures analysis of variance, corrected for clinical centres, age groups and an interaction between the age groups and progression groups, p=0.021) such that the individuals with rapid C-peptide decline had the lowest bacterial richness, on average (Fig. 6a). As there was a wide age range (1–38 years) within the ND cohort potentially affecting the GM composition, we further divided the participants into three age categories, age <7, age 7–12 and age 13 or older, based on distinct immunohistological profiles [32] and clinical characteristics [33]. Baseline clinical characteristics per age category are shown in ESM Table 7. We found that the association with C-peptide decline was most apparent in the group who were diagnosed with type 1 diabetes before the age of 7 (Fig. 6b–d), although the interaction term between the age groups and the progression groups was not significant (p=0.77).

Fig. 6
figure 6

Difference in microbial richness (denoting number of observed species) according to different C-peptide profiles during the follow-up: (a) in all ND cohort participants and in three age groups, (b) under 7, (c) 7–12 and (d) over 12 years old, within the ND cohort. Individuals were divided into three groups based on the decline of C-peptide levels over time, depicting the rate of disease progression: rapid and slow progression and increasing C-peptide levels [30] Species richness denotes the number of observed species

Full size image

Nineteen individuals in the UFM cohort progressed to clinical type 1 diabetes during or after the follow-up. We compared the GMs of these individuals with the rest of the UFM cohort. Individuals diagnosed with diabetes had increased abundance of Sutterella sp. KLE1602 (longitudinal analysis, β-coefficient=1.20 [95% CI 0.59, 1.80], q=0.033, nominal p=1.2 × 10−4, ESM Table 8), which was also more prevalent in these individuals: 10 out of 19 participants (53%) diagnosed with type 1 diabetes had Sutterella sp. KLE1602 in at least one stool sample compared with 27% (48 of 175) in non-diabetic UFM cohort participants. The progressors also had increased abundance of the functional module pyruvate:ferredoxin oxidoreductase (MF0073, β-coefficient=0.018 [95% CI 0.0092, 0.028], q=0.059, nominal p=9.8 × 10−5, ESM Table 8) in their GMs. We did not observe any longitudinal GM changes or associations between HbA1c levels and GM features, or associations between the stages of type 1 diabetes and microbial α diversities in the UFM cohort.

Discussion

Baseline relative abundance of F. prausnitzii was inversely correlated with HbA1c at the diagnosis of type 1 diabetes. F. prausnitzii is a beneficial commensal bacterium with important health-promoting functions [34], by, for example, production of butyrate which serves as an energy source for colonocytes. Microbial butyrate production is among the strongest mechanisms through which the GM mediates protection against islet autoimmunity and type 1 diabetes [7, 12, 35]. In a pilot trial supplementing resistant starch modified with SCFAs to humans with type 1 diabetes, individuals with the highest SCFA concentrations exhibited the best glycaemic control [13]. Future human cohort studies with faecal and/or serum butyrate quantification could further investigate the role of butyrate in glycaemic control in type 1 diabetes.

We observed differences in microbial richness, the number of observed taxa, according to the disease progression in the ND cohort. Individuals with the fastest decline in C-peptide levels, reflecting decline in endogenous insulin production‚ had the lowest GM richness. We also observed an inverse association between B. obeum and the change in fasting C-peptide levels between baseline and 2 years after the diagnosis. The difference in microbial richness is compatible with the observations from the DIABIMMUNE cohort, which was a longitudinal study in infants and young children with HLA-conferred susceptibility to type 1 diabetes [6]. In that study the decline in microbial diversity appeared prior to the diagnosis of type 1 diabetes. However, such a pattern had been absent in most other human cohort studies, suggesting that microbial diversity alone has a poor predictive value for type 1 diabetes. Together, these observations paint a picture where a multitude of microbial features may be weakly involved in the host–GM interplay in established disease.

UFM cohort members who had been diagnosed with type 1 diabetes during or after the follow-up had higher relative abundance and prevalence of Sutterella sp. KLE1602 compared with other UFM cohort members. While Sutterella spp. are not as extensively studied as many other prevalent gut bacteria, a study in mice found that Sutterella spp. degraded IgA in the gut [36]. The authors showed that mice with Sutterella spp. and resulting low gut IgA levels had increased dextran sulfate sodium (DSS)-induced colon ulceration compared with mice with normal IgA levels and no Sutterella. A randomised double-blind placebo-controlled clinical trial in humans testing the efficacy of FMT in ulcerative colitis found that Sutterella spp. were enriched in individuals who failed to stay in remission following the transplant [37]. Speculatively, Sutterella could disrupt the intestinal antibacterial immune response and indirectly induce local and systemic inflammation by increase in intracellular pathogens and pathobionts [38]. Potential roles of Sutterella spp. and colonic IgA in type 1 diabetes could be further investigated by, for example, including faecal IgA assays and sorting strategies that identify gut bacterial IgA coating in current and upcoming type 1 diabetes cohort studies.

We observed longitudinal GM changes over the 2 year observation period in the ND cohort. These changes included increases in known oral species Bifidobacterium dentium and S. salivarius, as well as several butyrate-producing Ruminococcaceae species and Coprococcus catus. Such longitudinal shifts were absent in the UFM cohort, suggesting the changes in the ND cohort could be related to the manifestation of type 1 diabetes and subsequent lifestyle changes. For example, individuals with newly diagnosed type 1 diabetes are recommended to adapt their dietary patterns and content of their diet which might be reflected in the GM structure.

We found P. copri in 88 (23.8%) metagenomes. Presence of P. copri tended to have significant effects in many functional modules encoded by the entire microbial community. Prevotella-rich microbiomes are more common in non-Western populations, while the GMs in Western countries are more commonly dominated by Bacteroides spp. [39]. Generally, high abundance of Prevotella is associated with a plant-rich, high-fibre diet, and overweight adults consuming a whole-grain diet for 6 weeks lost more weight if their GM harboured Prevotella [40]. Here, the observed dichotomy in Prevotella abundance could reflect long-term dietary habits and other external factors affecting the GM composition. While we did not observe any associations with P. copri (or other Prevotella spp.) and clinical data in our analyses, Prevotella-rich microbiomes could still provide means for stratifying the metagenomes to different community types in further exploratory analyses.

These data included prominent differences in the GM composition between recruiting clinical sites, reflected in β diversities and individual microbial species and functional modules. Such shifts are likely related to geographic, lifestyle, ethnic and other differences between the clinical sites distributed across 13 European countries. These factors are known confounders in microbiome studies since they often prominently correlate with GM structure and therefore pose a challenge to designing large GM studies in rare or otherwise difficult-to-study human conditions. Nonetheless, these data further underscore the necessity of controlling for these and other known confounding factors in GM studies. Specifically, these factors should be considered in trial design for best reproducibility and reliability. Sex was included as a covariate in the statistical analysis, and accordingly sex is not confounding the study results.

Strengths

This study investigated the GM of a relatively large cohort of 98 individuals with newly diagnosed type 1 diabetes with a 2 year longitudinal follow-up involving five visits to the clinic. On all clinic visits, harmonised clinical data on the participants, including fasting C-peptide and HbA1c measurements, were collected. Additionally, stool samples were collected using a harmonised process, enabling robust comparisons between the study centres.

Limitations

The GM data included prominent differences between the clinical sites distributed across 13 European countries. Such shifts are likely related to lifestyle, dietary, ethnic and other differences between the populations but since this information has not been collected in INNODIA we were unable to further dissect the source of the observed variation. Among these, the diet is a major GM modulator and dietary changes following type 1 diabetes could plausibly induce shifts in the GM composition. This hypothesis cannot be investigated here due to the lack of information on participant diets. Stool samples and microbiome data were not available at all time points from all individuals. Even though we are not aware of any systematic stool sampling bias, in theory, any such systematic bias could skew the results. Stimulated peak C-peptide level and area under the curve from mixed meal tolerance test (MMTT) are more accurate measures of beta cell function compared with fasting C-peptide used here, but stimulated C-peptide data were not available at baseline for all participants.

Conclusions

This study reports associations between the GM composition and type 1 diabetes in an INNODIA multicentre human cohort study. Albeit these associations were statistically weak, potential mechanisms behind these associations could be further investigated using animal models and immunological assays. Rigorous, controlled trials are needed to assess the roles of identified bacteria and their potential causality in type 1 diabetes.