Distinct gut metagenomics and metaproteomics signatures in prediabetics and treatment-naïve type 2 diabetics

Abstract

Background

The gut microbiota plays important roles in modulating host metabolism. Previous studies have demonstrated differences in the gut microbiome of T2D and prediabetic individuals compared to healthy individuals, with distinct disease-related microbial profiles being reported in groups of different age and ethnicity. However, confounding factors such as anti-diabetic medication hamper identification of the gut microbial changes in disease development.

Method

We used a combination of in-depth metagenomics and metaproteomics analyses of faecal samples from treatment-naïve type 2 diabetic (TN-T2D, n = 77), pre-diabetic (Pre-DM, n = 80), and normal glucose tolerant (NGT, n = 97) individuals to investigate compositional and functional changes of the gut microbiota and the faecal content of microbial and host proteins in Pre-DM and treatment-naïve T2D individuals to elucidate possible host-microbial interplays characterizing different disease stages.

Findings

We observed distinct differences characterizing the gut microbiota of these three groups and validated several key features in an independent TN-T2D cohort. We also demonstrated that the content of several human antimicrobial peptides and pancreatic enzymes differed in faecal samples between three groups.

Interpretation

Our findings suggest a complex, disease stage-dependent interplay between the gut microbiota and the host and point to the value of metaproteomics to gain further insight into interplays between the gut microbiota and the host.

Fund

The study was supported by the National Natural Science Foundation of China (No. 31601073), the National Key Research and Development Program of China (No. 2017YFC0909703) and the Shenzhen Municipal Government of China (No. JCYJ20170817145809215). The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.

Research in context.

Evidence before this study

An altered human gut microbiota has been strongly linked to metabolic disorders, such as obesity and T2D. However, reported T2D-related microbial features are inconsistent between studies, potentially confounded by age, ethnicity, geography and anti-diabetic medication.

Added value of this study

We report distinct differences characterizing the gut microbiota of NGT, Pre-DM, and TN-T2D. Importantly, we reveal a substantial number of Pre-DM associated microbial and human protein features at both the metagenomics and (or) metaproteomics level and validated several key microbial features in an independent TN-T2D cohort.

Implications of all the available evidence

Our findings suggest that highly complex changes of the intestinal ecosystem may characterize different T2D developmental stages. Longitudinal studies with multi-omics data including information on the gut microbiota, clinical parameters, and dietary information of prediabetic individuals are needed to further our knowledge on disease mechanisms and provide feasible approaches for precision treatment.

Alt-text: Unlabelled Box

1. Introduction

Type 2 diabetes mellitus (T2D) is a chronic heterogeneous disorder associated with hyperglycaemia and low grade inflammation [1,2]. The prevalence has increased dramatically in Westernized countries, and also in China, where 11.6% and 36% of Chinese adults suffer from diabetes and prediabetes (Pre-DM), respectively [3]. Due to complications and comorbidities related to the development of T2D, comprehensive characterization of phenotypic, metabolic and molecular changes of the host and the gut microbiota in Pre-DM and T2D compared to NGT is needed to enable early identification of prediabetic individuals at high risk of T2D development.

Cross-sectional metagenomic studies have linked alterations in the gut microbiome to T2D and prediabetes [[4], [5], [6], [7]]. However, a few recent intervention studies have reported profound impact of antidiabetic drugs on the human gut microbiome, such as metformin, acarbose and glucagon-like peptide-1 (GLP-1) based therapies [[8], [9], [10], [11], [12], [13]], emphasizing the importance of controlling for medication in studies on association between the microbiota and T2D. Moreover, distinct disease-related microbial profiles have been reported in different age and ethnic groups [[4], [5], [6], [7]], making it difficult to identify the microbes possibly involved in disease development. Thus, detailed information on the gut microbial species associated with T2D onset and progression is still limited. Whereas information from metagenomics is limited to identification of the presence of genes, taxa, and their inferred functional capacity, introduction of additional omics approaches including metabolomics, metatranscriptomics, and metaproteomics has increased our knowledge of microbial activity in health and disease [[14], [15], [16], [17]]. For instance, recent metatranscriptomics studies on inflammatory bowel disease and cirrhosis cohorts have revealed considerable discrepancies between data obtained from metagenomics vs metatranscriptomics analyses [17,18]. As metaproteomics enables identification of microbial and human proteins simultaneously in faecal samples [14,19,20], such an approach offers a potential for deciphering both active microbial functions and host-microbiota interactions.

In the present study, we examined 254 stool samples collected from a Chinese cohort combining shotgun metagenomics and metaproteomics analyses. We characterized substantial differences between NGT, Pre-DM and TN-T2D individuals. Of note, consistent aberrations in Pre-DM and TN-T2D individuals included lower abundances of Clostridiales species and higher abundances of Megasphaera elsdenii compared to NGT individuals. Several robust microbial compositional changes were detected at both the DNA and protein levels, such as an enrichment of E. coli in Pre-DM individuals and an increased abundance of Bacteroides spp. in TN-T2D patients. Several Pre-DM-specific features were furthermore uncovered, including a reduced functional potential for processes involved in energy metabolism and bacterial growth. Thus, our findings revealed distinct characteristics of the intestinal ecosystem in the Pre-DM stage. Of note, proteomics analyses revealed that the levels of several antimicrobial peptides (AMPs) and pancreatic enzymes appeared to be lower in faecal samples of the TN-T2D individuals than in the other two groups. These findings suggest that potential specific differences in the host response among groups might influence the composition of the gut microbiota, or vice versa. In conclusion, our study provides a basis for further analyses integrating faecal metagenomics and metaproteomics which may lead to a better understanding of mechanisms underlying the development of Pre-DM and T2D.

2. Materials and methods

2.1. Suzhou T2D study population

The study was approved by the Institutional Review Board of BGI-Shenzhen and the ethical review committee of Suzhou Centre for Disease Control and Prevention (CDC). The study population recruited from community residents from Suzhou, comprised 97 Chinese adults with normal glucose tolerance (NGT), 80 prediabetes patients (Pre-DM) and 77 newly diagnosed, treatment naïve type 2 diabetes patients (TN-T2D). All TN-T2D patients and Pre-DM individuals were screened and newly diagnosed according to the 2011 WHO criteria by a well-trained staff at the Suzhou CDC, as described in detail in a recent published lipidomic study based on this cohort [21]. All enrolled 254 individuals reported no anti-diabetic treatments; thus, none have had taken insulin, or any oral or injectable anti-diabetic medication. However, considering the high prevalence of metabolic adverse conditions in the Suzhou cohort, patients who had been diagnosed with hypertension, dyslipidaemia, or prescribed with drugs for treating these diseases were included. Stool samples for metagenomics were self-collected in 2 ml faecal containers and immediately stored at −80 °C and transported to the laboratory on dry ice. DNA was extracted as previously described [4]. A summary of sample information is presented in Table S1. In addition, shotgun metagenomic datasets of stools from 94 TN-T2D patients from Shanghai, a city near to Suzhou, receiving no anti-diabetic medication [9] were included for validation purposes.

2.2. Method for metagenomics

2.2.1. Generation of a BGISEQ-500-based faecal metagenome data set

In this study, we performed DNA library construction and the combinatorial probe-anchor synthesis (cPAS)-based BGISEQ-500 sequencing for metagenomics (single-end; read length of 100 bp) and applied the same quality control workflow to filter the low-quality reads in accordance with the recently published metagenomic study using this new platform [22]. The remaining high-quality reads were then aligned to hg19 to remove human reads [23]. Metagenomic data statistics is provided in Table S2.

2.2.2. Profiling of metagenomic samples and biodiversity analysis

High-quality non-human reads were aligned to the 9.9 M integrated gene catalogue (IGC) by SOAP2 using the criterion of identity ≥90% [23]. Sequence-based gene abundance profiling was performed as previously described [23]. The relative abundances of phyla, genera, species and KOs were calculated by the sum of the relative abundance of their annotated genes. The alpha diversity (within-sample diversity) was quantified by the Shannon index using the relative abundance profiles at gene, genus and KO levels as described [23]. The beta diversity (between-sample diversity) was calculated using Bray-Curtis dissimilarity (R version 3.3.2, vegan package 2.4–4).

2.2.3. Metagenome-wide association analysis (MWAS)

MWAS was performed on the Suzhou T2D cohort as previously described [4]. Using non-parametric Kruskal-Wallis test (R version 3.3.2 stats package), we identified 266,015 genes showing significantly different abundances between the NGT, Pre-DM and TN-T2D groups (P < .05). After clustering, a total of 126 MLGs (≥100 genes) were generated from these genes. The relative abundance of each MLG was summed using the relative abundance values of all genes from this MLG. The taxonomic annotation of each MLG was determined if >50% of genes in this MLG could be assigned to a certain taxon according to their IGC annotation. The genes of 85 unclassified MLGs were further annotated using a reference sequence database including 1520 high-quality genomes obtained from bacteria isolated from healthy Chinese individuals [24], which resulted in the taxonomic annotations of 11 additional MLGs (See detailed information in Table S5).

2.3. Method for metaproteomics

2.3.1. Sample preparation and LC-MS/MS analysis

Faecal samples from 84 individuals from NGT, Pre-DM, and TN-T2D individuals were used for metaproteome analysis using isobaric tags for relative and absolute quantitation (iTRAQ)–coupled-liquid chromatography tandem mass spectrometry (LC-MS/MS) (Fig. S1). Each group consisted of 28 randomly selected individual samples with matched age, sex and BMI by propensity score matching (R version 3.3.2, MatchIt package 2.4–21) [25] (Table S3). Faecal samples were processed using the filter-aided sample preparation (FASP) protocol [26]. Briefly, 100 mg frozen faeces from each individual were suspended in 500 μl lysis buffer (4% SDS, 100 mM dithiothreitol, 100 mM Tris-HCL (pH = 7.8) with freshly added protease inhibitors (cOmplete™, EDTA-free Protease Inhibitor Cocktail, Roche Applied Science). The samples were incubated for 5 min at 100 °C, followed by sonication to decrease the viscosity. The protein supernatants were collected after centrifugation at 30,000g at 4 °C for 30 min and then quantified using a 2D-quant kit (Sigma). For each diagnostic group, protein extracts in equal amounts from four individuals were pooled, and the selected 28 samples were thus aliquoted into 7 mixtures. A reference sample was created by pooling equal amounts of protein from each of 84 individual sample and 28 samples from self-reported T2D patients. Each mixture containing 100 μg proteins was loaded onto a 10 kDa cut-off spin column (Vivacon 500, Sartorius AG, Goettingen, Germany). The lysate was adjusted to 8 M urea by centrifugation to remove SDS and low-molecular-weight material. After reduction by dithiothreitol (DTT) and alkylation by iodoacetamide (IAM), 8 M urea was added and centrifuged to remove any remaining reagent such as IAM. The urea buffer was then replaced with 0.5 M triethylammonium bicarbonate (TEAB), and the sample was washed with 0.5 M TEAB 5 times. Trypsin (Promega, Madison, WI, USA) was added to digest the protein at a protein: trypsin ratio of 50:1, and the mixtures were incubated for 18 h at 37 °C. The resulting peptides were eluted twice with 100 μl 0.5 M TEAB by centrifugation at 12,000 g for 30 min and vacuum dried. The peptide mixture samples were then dissolved in 0.5 M TEAB and labelled with 8-plex iTRAQ reagents according to the manufacturer's protocol (AB Sciex, USA). For each diagnostic group, 7 mixtures were labelled with tags from I113 to I119. To perform the iTRAQ quantitation throughout the whole experiment, we labelled the reference sample by tag 121 in each iTRAQ run. Thus, three independent 8-plex iTRAQ runs were conducted. Subsequently, labelled peptides were separated on a LC-20AB HPLC system (Shimadzu, Kyoto, Japan) with an Ultremex SCX column (Phenomenon, Torrance, CA) and collected into 20 fractions. Each fraction was analysed via a NanoLC system coupled with a Q Exactive mass spectrometry (Thermo Fisher Scientific, San Jose, CA) as described previously [27].

2.3.2. Database searching and protein identification

For protein database searching, we used Mascot (Version 2.3) [28] as the search engine with the following parameters: trypsin was used as default enzyme and up to two missed cleavages were allowed. Carbamidomethyl (C), iTRAQ8plex (N-term) and iTRAQ8plex (K) were chosen as fixed modifications, and Oxidation (M) was chosen as variable modification. The peptide mass tolerance was set to 10 ppm and the fragment mass tolerance to 0.03 Da.

A two-step search method was applied. The MS/MS spectra were first searched against a collection of three protein sequence databases, including Homo sapiens sequences retrieved from SwissProt (release 2014_11), and human gut microbial protein sequences of IGC genes mapped by sequencing reads from our 254 metagenomic samples. The detailed search parameters are presented in Table S4. The Mascot search yielded a set of scored peptide-spectrum matches (PSMs) and the proteins were inferred from the PSMs. Subsequently, a target-decoy protein database was created containing the above-mentioned proteins and the reversed sequences from these proteins. A second round search based on the target-decoy database was performed to control for false positives as described elsewhere [29]. The PSMs were re-scored by Mascot Percolator [30] integrated into IQuant [31], and filtered at false discovery rate (FDR) ≤ 0.01. To improve the confidence in identification, peptides supported by ≥2 spectra were retained and protein identifications were thus inferred.

2.3.3. Meta-protein generation

Due to the shared similarity of metagenomic protein reference sequences, a microbial peptide hit is typically returned from several proteins within and between species. To avoid inflating numbers and alleviate taxonomic ambiguities of identified microbial proteins, several processes were performed to reduce data redundancy. We first grouped the microbial proteins with at least one shared peptide to generate protein clusters (Fig. S2). Each cluster was then processed according to the maximum parsimony principle. The minimum protein sets containing all peptides of each cluster were selected and defined as the meta-protein representing the cluster (Fig. S2). Individual proteins which only contained unique peptides were also assigned as a meta-protein. All redundant non-meta-protein sequences were thus omitted in subsequent analyses.

2.3.4. Protein quantification

Protein quantification was performed by IQuant [31] in the following three steps.

We first normalized the intensities of iTRAQ reporter ions for all spectra across the eight iTRAQ-labelled samples (I113…I119, I121) using the formula (1) as follows:

$$\overline{s_{i-k}}=\frac{S_{i-k}}{\mathit{median}\left(S_{1-k}:S_{n-k}\right)},\text{where}\mathrm{k}=I113\dots I119,I121$$ (1)

where $\overline{s_{i-k}}$ is the normalized relative intensity of spectrum i in the label k.

The reporter ion ratios were then determined using the formula (2):

$$\overline{r_{i-k}}=\frac{\overline{s_{i-k}}}{\overline{s_{i-121}}},\text{where}\mathrm{k}=I113\dots I119$$ (2)

where $\overline{r_{i-k}}$ is the ratio of relative intensity of spectrum i in the label k, with S_i−121, the relative intensity of the global QC labelled with 121 tags as denominators.

For protein quantification, only unique peptides were taken into consideration. The relative protein ratio was calculated using the mean relative intensity ratio of all unique peptide spectra in each protein using the formula (3):

$$\overline{p_k}=\mathit{mean}\left(\overline{r_{1-k}}:\overline{r_{p-k}}\right),\text{where}k=I113\dots I119$$ (3)

where $\overline{p_k}$ is the protein ratio in label K and acts as an indication of the relative proportions of that protein between the differently labelled samples.

2.3.5. Protein annotation

For microbial meta-proteins, taxonomic and functional annotations of identified proteins were derived from the putative protein-coding IGC genes. As a result, we linked 64.15% (8777 of 11,980) of the meta-proteins with annotation at the phylum or lower taxonomical levels and 80.27% (10,983 of 11,980) with KEGG Ontology (KO) annotation. For human proteins, functional annotations were obtained from UniProtKB/Swiss-Prot (release 2014_11).

2.4. Statistical analyses of metagenomes and metaproteomes

2.4.1. MLG-based random forest classification

Relative abundance data of all MLGs were subjected to random forest (RF) analysis to perform five-fold cross validation (R 3.3.2, caret package 6.0-77) [32]. The combinations of optimal MLGs markers maximising the discrimination accuracy between each two groups were thus determined by RF using an embedded feature selection strategy as previously reported [33]. The importance values of model-selected MLGs were calculated using “mean decrease in accuracy” strategy.

2.4.2. Spearman's rank coefficient correlation

Spearman's rank coefficient correlation (SCC) analysis was used for correlations between number of meta-proteins and metagenomic abundances at the genus level, and between the levels of proteins. The significance cut-off for SCC was set at an FDR adjusted P < .05.

2.4.3. Enrichment analysis of KEGG modules

Differentially enriched KEGG modules were identified according to reporter Z-scores [34]. Z-score for each KO was first calculated from Benjamín-Hochberg (BH)-adjusted P values from Wilcoxon rank-sum tests of comparisons between each two groups. The aggregated Z-score for each module was calculated using Z-scores of all individual KOs belonging to the corresponding module. A module was considered significant at a |reporter Z-score | ≥ 1.96.

2.4.4. Other statistical analyses

Kruskal–Wallis (KW) test was conducted to detect the differences in continuous phenotypic factors, microbial diversity, richness and MLG relative abundances between multi-groups. Dunn's post hoc tests followed KW test to explore the phenotypic differences among three pairwise comparisons (R version 3.3.2, PMCMR package 4.1). For MLG relative abundances, the Dunn's post hoc p-values were adjusted using the Benjamini-Hochberg (BH) method across all 126 identified MLGs for comparisons between each two groups from the Suzhou cohort. The significance cut-off was set as a BH-adjusted Dunn's P value <.05.

Wilcoxon rank-sum test was then performed on relative abundance of 126 MLGs between published TN-T2D patients from Shanghai [9] and NGT or Pre-DM from the Suzhou cohort for validation purposes. The significance cut-off was set as a BH-adjusted P value <.05. Detailed information on enrichment of MLGs between groups is provided in Table S5.

Wilcoxon rank-sum test was performed to detect differences in protein levels between each two groups. Due to a large number of identified proteins in this study, no statistically significant differences in protein intensities between groups could be reached with a significance cut-off of a BH-adjusted P value <.05 after P value adjustment across all proteins. Thus, a P value <.05, and a fold change of protein levels >1.2 or < 0.8 in protein intensities between two groups were used to report proteins with higher or lower trends.

Chi-square test was conducted to detect the distribution of differences in discrete phenotypic factors, such as sex and treatment distribution between groups, and to identify differences in taxonomic and functional assignments between metagenomic and metaproteomic datasets. The significant cut-off was set as a P value <.05.

2.5. Data availability

Metagenomic sequencing data for 254 faecal samples can be accessed from China Nucleotide Sequence Archive (CNSA) with the dataset identifier CNP0000175. The mass spectrometry metaproteomics data have been deposited to the ProteomeXchange Consortium via the PRIDE partner repository with the dataset identifier PXD013452 and https://doi.org/10.6019/PXD013452.

3. Results

3.1. Experimental design

The cohort consisted of 77 TN-T2D patients, 80 Pre-DM individuals and 97 NGT individuals from Suzhou, China (Methods, Table S1). The three groups were matched regarding body mass index (BMI) and sex (P > .05), but individuals with TN-T2D (mean age 66 +/− 8 years) were on average 5 years older than individuals in the two other groups (Table S1). Shotgun metagenomics was performed on faecal samples from all participants, whereas metaproteomics profiling was performed on a subgroup of 84 participants, including 28 age-, sex-, and BMI-matched individuals from each group (Fig. 1).

Fig. 1.

Experimental overview.

254 participants were recruited from the Suzhou cohort and diagnosed as treatment naive T2D patients (TN-T2D, n = 77, red), prediabetic individuals (Pre-DM, n = 80, blue) or individuals with normal glucose tolerance (NGT, n = 97, green). Each participant provided two stool samples. One set of stool samples was used for metagenomic shotgun sequencing, followed by IGC-based taxonomic and functional analyses. The other set of stool samples, comprising a total of 84 samples with 28 age-, BMI- and sex-matched participants from each group, was selected for metaproteomic analyses using isobaric tags for relative and absolute quantitation (iTRAQ)–coupled-liquid chromatography tandem mass spectrometry (iTRAQ-LC-MS/MS) to provide information on the microbial and host proteins present in stool samples.

A total of 11, 980 meta-proteins and 425 human proteins were identified in this study. Microbial gene and protein profiling were used to determine alterations in the abundance of microbial taxa and functions, and human protein profiling was used to identify alterations in the abundance of human proteins in faecal samples from NGT, Pre-DM and TN-T2D individuals. Colour and shapes indicate gut microbial taxa, functional pathway/modules and identified fecal microbial and human proteins enriched in NGT (green), Pre-DM (blue) and TN-T2D (red). Yellow lines indicate metagenomic analyses; grey lines indicate metaproteomic analyses. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

3.2. Distinct metagenomics profiles in Chinese prediabetic and type 2 diabetic individuals

Shotgun metagenomic sequencing of the 254 stool DNA samples was performed using the BGISEQ-500 platform and raw reads were filtered and aligned to the integrated gene catalogue (IGC) of the human gut microbiome to generate gene, taxonomic and functional profiles as previously described (Methods, Table S2). In line with previous studies [[4], [5], [6]], no significant differences in microbial gene-based richness, alpha-diversity, and beta-diversity were found between the NGT, Pre-DM, and TN-T2D individuals (Fig. S3, Kruskal-Wallis (KW) test, P > .05). Using a metagenome-wide association approach [4], we identified 266,015 T2D-associated genes (KW test, P < .05) and clustered these genes into 126 metagenomic linkage groups (MLGs, ≥ 100 genes, Table S5).

We further applied the KW test to detect statistically significant differences in the relative abundances of MLGs between individuals with NGT, Pre-DM, and TN-T2D. Compared to NGT individuals, the abundances of MLGs from the Clostridia class, such as Dialister invisus (MLG-3376) and Roseburia hominis (MLG-14865 and MLG-14920) were significantly lower in individuals with Pre-DM or TN-T2D (Fig. 2A, Table S5, Dunn's test, adjusted P < .05), which is in agreement with previous findings in a Danish T2D cohort [6]. In addition, we found that the abundance of the butyrate-producing Faecalibacterium prausnitzii (MLG-4560) was lower in Pre-DM compared to both NGT and TN-T2D individuals (Fig. 2A, adjusted P < .05). By contrast, MLGs annotated to Escherichia coli (MLG-7919 and MLG-7840), Streptococcus salivarius (MLG-6991), and Eggerthella sp. (MLG-351) were highly enriched in Pre-DM compared to NGT individuals (Fig. 2A, adjusted P < .05). An increased abundance of Streptococcus operational taxonomic units (OTUs) was also recently reported in a Danish prediabetic cohort [7]. Additionally, Pre-DM individuals also exhibited a significant enrichment in E. coli abundance compared to TN-T2D individuals (Fig. 2A, adjusted P < .05). Moreover, we detected significantly lower abundances of Akkermansia muciniphila (MLG-2159) and Clostridium bartlettii (MLG-7540) and higher abundances of Bacteroides caccae (MLG-10234 and MLG-10325), Bacteroides finegoldii (MLG-10154 and MLG-10159), and Collinsella intestinalis (MLG-10084) in TN-T2D patients compared with NGT and Pre-DM individuals (Fig. 2A, adjusted P < .05). Finally, the abundance of Megasphaera elsdenii (MLG-1568) was significantly higher in both TN-T2D and Pre-DM individuals than in NGT individuals (Fig. 2A, adjusted P < .05), in line with the positive correlation between the relative abundance of the genus Megasphaera and T2D recently reported in a large cohort with about 7000 individuals from South China [35]. Several key findings were further validated in faecal samples of 94 treatment naïve T2D patients in Shanghai [9], such as a lower abundance of A. muciniphila and C. bartlettii compared to NGT and Pre-DM individuals, and a lower abundance of E.coli compared to Pre-DM individuals in this study (Fig. 2A, Table S5, Wilcoxon rank-sum test, adjusted P < .05). A summary of gut microbial taxa reported in previously published cross-sectional T2D or prediabetes studies is presented in Table S6.

Fig. 2.

Determination of alterations in the abundance of MLGs and functional modules.

(A) Heatmap of statistically significant annotated MLGs discriminating between TN-T2D, Pre-DM and NGT based on Z-scores. Red, MLGs enriched in high glucose groups, blue, MLGs enriched in low glucose groups. *, indicates MLGs significantly differed between any two groups in the Suzhou cohort; Dunn's test, adjusted P < .05. #, indicates significant MLGs replicated in the treatment naïve T2D patients from Shanghai compared with Pre-DM and NGT in the Suzhou cohort; Wilcoxon rank-sum test, adjusted P < .05 (See Table S5 for full list).

(B) Performance of cross-validated random forest (RF) classification models using relative abundance profiles of gut microbial MLGs, assessed by the area under the ROC curve (AUC), 95% confidence intervals (CI). Orange, AUC for the RF model classifying NGT (n = 97) and Pre-DM (n = 80). Grey, AUC for the RF model classifying NGT (n = 97) and TN-T2D (n = 77). Blue, AUC for the RF model classifying Pre-DM (n = 80) and TN-T2D (n = 77). The best cut-off points are marked on the ROC curves.

(C) Bar plot showing the 10 most discriminating MLGs in the RF models for distinguishing between NGT, Pre-DM and TN-T2D. The bar lengths indicate the importance of the selected MLGs, and colours represent enrichment in NGT (green), Pre-DM (blue) and TN-T2D (red).

(D) Differential enrichment of KEGG modules comparing TN-T2D, Pre-DM and NGT. Dashed lines indicate a reporter score of 1.96, corresponding to 95% confidence in a normal distribution. Colour bars show reporter score values of modules enriched in NGT (green), Pre-DM (blue) and TN-T2D (red). (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

To assess the discriminative power of MLGs in T2D and identify key MLGs differentiating individuals with respect to different disease stages, we applied a feature selection approach and constructed Random Forest (RF) classification models comparing the groups (Methods). Remarkably, the RF models provided high performances regarding classification of samples from the two different disease stages, with area under the ROC curve (AUC) values from 0.90 to 0.94 (Fig. 2B). Apart from taxonomically unclassified MLGs, the most discriminatory MLG for separating TN-T2D and NGT was A. muciniphila. Moreover, MLGs annotated to F. prausnitzii and E. coli both showed to be important in separating Pre-DM samples from TN-T2D and NGT samples (Fig. 2C), indicating the unique microbial DNA signatures of lower abundance of F. prausnitzii and higher abundance of E. coli in Pre-DM individuals. We also validated the predictive power of the RF models between TN-T2D and the other two groups, which showed an accuracy of 76. 6% (72 of 94 patients) for disease prediction in a previously described TN-T2D cohort from Shanghai (Table S7) [9].

We next performed KEGG enrichment analyses to examine possible differential patterns of microbial functional potentials in NGT, Pre-DM and TN-T2D individuals (Table S8). Interestingly, we observed a significant enrichment in modules comprising several sugar phosphotransferase systems (PTS), ATP-binding cassette transporters (ABC transporters) of amino acids, and bacterial secretion systems in the gut microbiota of Pre-DM compared to NGT individuals (reporter score ≥ 1.96, Fig. 2D). Likewise, in line with previous findings in several Chinese cohorts with metabolic diseases, such as atherosclerotic cardiovascular disease (ACVD), obesity and T2D [36], a similar enrichment was found in TN-T2D patients compared with NGT individuals (Fig. 2D). The abundances of the transport system for microcin C, a peptide-nucleotide antibiotic produced by Enterobacteria [37], and the transport system for autoinducer-2 (AI-2), a quorum sensing signalling molecule reported in Proteobacteria [38], were also significant higher in Pre-DM than in NGT individuals (Fig. 2D). Except for enrichment of type II-IV secretion and AI-2 transport systems in Pre-DM vs TN-T2D, we found no other KEGG modules for PTS and ABC transporters to differ significantly in abundance between Pre-DM and TN-T2D individuals (Fig. 2D). However, Pre-DM individuals displayed a significant reduction with respect to several energy and nucleotide metabolism modules compared to both NGT and TN-T2D individuals, including modules of V-type ATPase, pyruvate: ferredoxin oxidoreductase, and bacterial ribosomal proteins (Fig. 2D). We next performed a second-round KEGG enrichment analysis by excluding all E. coli genes. Interestingly, abundances of several of the abovementioned Pre-DM enriched modules of ABC transporters, microcin C and AI-2 transporters in this analysis showed no significant differences between Pre-DM and NGT individuals, suggesting that E. coli may be an important contributor to the observed differences in the functional potential of Pre-DM compared to NGT individuals (Fig. S4). Taken together, these results indicate the possible involvement of substantial compositional and functional disease-related gut microbial changes in the pre-diabetic stage.

3.3. Gut metaproteomics simultaneously identifies faecal levels of microbial and human proteins

To gain further insights into functional changes in the gut microbiota associated with T2D, we conducted metaproteomic analyses using iTRAQ (isobaric peptide tags for relative and absolute quantification) and LC-MS/MS-based protocols on 84 samples, with 28 samples derived from each of the three diagnostic groups (Methods, Fig. S1). Using the strict parameters of 2 peptide-spectrum matches (PSMs) per protein, < 10 ppm mass error and 1% PSM-level FDR (Methods), we identified a total of 145,014 high quality PSMs corresponding to 15,670 proteins, including 15,245 (97.29%) microbial proteins and 425 (2.71%) human proteins (Table S9). As reported [14,19,20], one microbial peptide often exhibits matches to multiple proteins with high sequence similarity, resulting in difficulties in identifying the microbial origin of individual peptides. To alleviate ambiguities, we applied a maximum parsimony principle reported in recent studies [14] [39] and generated 11,980 non-redundant meta-proteins (78.58% of microbial proteins) containing at least one unique microbial peptide. The relative intensities of these unique peptides were further used for meta-protein quantification (Methods, Table S9). The number of identified meta-proteins ranged between 5067 in the Pre-DM samples to 8134 in the TN-T2D samples (Table S9). Venn diagrams showed that only 2782 meta-proteins (34.2%–54.9% of the total number of meta-proteins per group) were shared among the three groups (Fig. S5A), indicating differential microbial expression patterns at the protein level among the groups. Taxonomic annotations indicated a higher percentage of unique Proteobacteria meta-proteins in Pre-DM individuals, compared to the other groups (Chi-square test, P < .05, Fig. S5B), whereas no difference in the distributions of the uniquely detected meta-proteins associated with a wide range of functions was found between the three groups (Fig. S5C).

3.4. Concordance and discordance of microbiota features between metagenomes and metaproteomes

Based on annotated microbial features, we next investigated the consistency as well as the divergence of microbial composition and function at the DNA and protein level. At the phylum level, >90% genes and meta-proteins were consistently assigned to three major phyla, namely Firmicutes, Bacteroidetes and Proteobacteria (Fig. 3A). Despite the overall consistency, we found a significantly higher percentage of the annotated proteins to be assigned to Bacteroidetes (41%) compared to the percentage of genes annotated to Bacteroidetes (25%) (Chi-square test, P < .05, Fig. 3A), suggesting that Bacteroidetes might display an overall higher protein production than the other phyla across the 84 samples. At the genus level, the composition of the metaproteomes was biased towards a limited number of genera. Among 212 common metagenomically-identified genera detected in at least 10% of the 84 samples, only 81 genera (38.21%) could be detected based on metaproteomics (Table S10). Spearman's rank correlation analysis was subsequently performed to determine the relationship between the number of meta-proteins and the abundances at the genus level based on metagenomics. The more abundant a given genus was based on metagenomics analysis, the more of the identified meta-proteins were assigned to this genus (Spearman's correlation coefficient (SCC) = 0.726, P = 5.21E-08, Fig. 3B, Table S9), with Bacteroides (n = 1664), Prevotella (n = 818) and Faecalibacterium (n = 719) harbouring most assigned meta-proteins. For a few genera, such as Anaerotruncus (n = 9), Paraprevotella (n = 9) and Enterococcus (n = 7), we were only able to identify <10 meta-proteins although their median metagenomic abundances were greater than 1E-04 (Table S10).

Fig. 3.

Concordance and discordance of gut microbiome features in metagenomes and metaproteomes.

(A) Taxonomic distribution at the phylum level. Inner circle, metagenomes; Outer circle, metaproteomes.

(B) Spearman's rank correlation between the median relative abundances of genera in metagenomes of 84 samples selected for metaproteomics and the number of identified meta-proteins assigned to the same genus.

(C) Functional distribution at KEGG level 2. Inner circle, metagenomes; Outer circle, metaproteomes.

(D–E) Enrichment analysis of meta-proteins at taxonomic (D) and functional levels (E) comparing NGT, Pre-DM and TN-T2D individuals. Colour bars represent the number of meta-proteins that exhibited trends towards higher levels in NGT (green), Pre-DM (blue) and TN-T2D (red), defined as P < .05 (Wilcoxon rank-sum test) with a fold change of mean intensities >1.2 in pairwise comparisons. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

Comparing KEGG functional categories based on metagenomics and metaproteomics data, we observed large differences in the relative contribution of individual categories between the two datasets (Chi-square test, P < .05, Fig. 3C), in accordance with several previous studies [14,19,20]. For instance, as determined by metaproteomics, 24% and 18% of the proteins were assigned to carbohydrate metabolism and translation categories, whereas the corresponding metagenomic percentages of the two categories were only 11% and 4%, respectively (Fig. 3C). We found that 1508 meta-proteins, accounting for 12.59% of all identified meta-proteins, could be assigned to 10 KEGG orthologues (KO). The top KOs harboured 360 proteins annotated as Ca-activated chloride channel homologues (K07114), whereas the remaining KOs comprised proteins representing abundant house-keeping proteins such as elongation factors, large subunit ribosomal proteins (K02355, K02358 and K02395), chaperones (K04077 and K04043), and glyceraldehyde 3-phosphate dehydrogenase (K00134) as well as flagellin proteins (K02406) (Table S11, Fig. S6).

Aiming to link the microbial protein patterns to metagenomic microbial abundances, we next conducted a fold-change analysis of meta-proteins. In agreement with our metagenomic findings (Fig. 2A), Proteobacteria meta-proteins (mainly from Escherichia, Citrobacter and Enterobacter) exhibited a trend towards an increase in the Pre-DM group, whereas Bacteroides meta-proteins tended to be enriched in TN-T2D individuals (Fig. 3D, Table S12, P < .05 and fold change (FC) of protein intensities >1.2). The levels of several Prevotella meta-proteins tended to be higher in Pre-DM individuals (Fig. 3D, P < .05 and FC > 1.2), although no Prevotella annotated metagenomic MLGs exhibited significantly higher abundance. At the functional level, we observed that the level of meta-proteins involved in carbohydrate metabolism tended to be lower in NGT compared to Pre-DM and TN-T2D individuals (Fig. 3E, Table S11).

3.5. Functional characteristics of excreted human proteins in faeces of T2D individuals

Among the 425 detected human proteins, we identified 218 human proteins that were shared among the NGT, Pre-DM, and TN-T2D groups, accounting for 59.6% to 85.2% of the identified human proteins in each group (Fig. S7A). We next annotated the human proteins with Gene Ontology (GO) terms to obtain insight into the functional characteristics of the human proteins excreted in faeces (Table S13). Among the identified proteins, 181 (42.59%) had previously been identified in faecal samples by metaproteomics, indicative of their general presence (Table S14) [14,19,20]. These included several intestinal mucin proteins, such as MUC-1, MUC-2, MUC-4, MUC5B, MUC12, and MUC-13 as well as members of annexins (ANXA1- ANXA7, a family of calcium-binding proteins) (Table S14). We identified 233 of the faecal human proteins to have tissue-specific annotation, among which 151 proteins (64.81%) were reported to exhibit high expression in the digestive system, and the remaining proteins were annotated to be highly expressed in blood or other tissues such as epidermis (Table S13). Of interest, 18 of the human proteins were annotated as AMPs [40] (Table S13). Several human proteins involved in glucose metabolism, including the sodium/glucose cotransporter 1, were detected in faecal samples of TN-T2D patients only (Fig. S7B). Inhibitors of this protein have been proposed for antidiabetic treatment. Additionally, the TMAO-producing enzyme, dimethylaniline monooxygenase [N-oxide-forming] 3 (FMO3) was also identified exclusively in the TN-T2D group (Table S13). On the other hand, we found that ras GTPase-activating-like protein (IQGAP1) and unconventional myosin-Ic (MYO1C) were uniquely identified in the NGT group (Fig. S7B). Loss of IQGAP1 and MYO1C has been related to impairment of insulin signalling [[41], [42], [43]], but whether their presence in faeces has functional implications remains to be established.

Forty-nine of the human proteins present in faeces were found to show trends towards higher or lower intensities between at least two of the groups (Fig. 4A, Table S15). For instance, we observed a trend towards higher levels of four AMPs, including defensin-5, neutrophil defensin-1, lysozyme c, as well as secreted phospholipase A2, all with important roles in the defence against bacteria [[44], [45], [46]], in faecal samples from NGT individuals than in samples from TN-T2D individuals (Fig. 4A, P < .05 and FC > 1.2). Interestingly, the level of the antimicrobial cathepsin G, reported to inhibit the growth of several organisms from the Proteobacteria phylum [47], tended to be higher in samples from Pre-DM than NGT and TN-T2D (P < .05 and FC > 1.2), and this was coupled to lower levels of alpha-1-antichymotrypsin and alpha-1-antitrypsin, both known inhibitors of cathepsin G [48] (Fig. 4A, P < .05 and FC < 0.8), suggesting that Pre-DM individuals may have initiated strategies to activate a defence system against the enhanced relative abundances of E. coli. Individuals with Pre-DM also exhibited a trend towards lower galectin-3 levels, a lectin with beta-galactoside-binding ability. Galectin-3 has been reported to bind lipopolysaccharides (LPS) from E. coli and play a role as a negative regulator of LPS-mediated inflammation [49]. In addition, galectin-3 was also reported to improve epithelial intercellular contact via desmoglein-2 stabilization [50]. We also found that several proteolytic enzymes tended to occur in lower levels in faeces from TN-T2D individuals. Thus, we found a trend towards lower levels of proteases (trypsin and chymotrypsin and their precursors) and lipases. By contract, we observed a trend towards higher amylase (AMY1) levels in TN-T2D (Fig. 4A). It is interesting to note that the level of dipeptidyl peptidase 4 (DDP4), known to inhibit insulin secretion via its action on GLP-1, tended to be lower in individuals with Pre-DM than in TN-T2D individuals. A network analysis revealed associations between 20 human proteins showing different trends in relation to enrichment in two-pairwise comparisons between NGT, Pre-DM and TN-T2D individuals (Fig. 4B). For instance, we identified a negative correlation between the defensin-5 and TN-T2D-enriched peptidyl-prolyl cis-trans isomerase B (PPIB) (Fig. 4B, Spearman's correlation, adjusted P < .05), the latter previously reported to be associated with islet dysfunction [51].

Fig. 4.

Characterization of human proteins in faecal samples from Chinese NGT, Pre-DM, and TN-T2D individuals.

(A) Heatmap showing levels of 49 human proteins as fold change between each two groups. Only proteins exhibiting trends towards different levels in at least one pairwise comparison are shown. *, P < .05 and fold change of protein levels >1.2 or < 0.8.

(B) Protein-protein interaction network based on 20 human proteins exhibiting trends towards different levels in at least two pair-wise comparisons. The group signatures indicate human proteins with trends towards higher or lower levels in this group compared to others. Orange indicates higher protein levels and blue indicates lower protein levels.

(C) Protein-protein interactions based on selected meta-proteins and human proteins with trends towards different levels between the three groups. Only meta-proteins annotated to the corresponding taxon of the MLGs were selected for the analysis.

The circles indicate human proteins and diamonds indicate meta-proteins. Detailed information on the numbered meta-proteins is presented in Table S12. Colours represent protein of a trend towards a higher level in NGT (green), Pre-DM (blue) and TN-T2D (red). Pink line indicates positive correlation and grey line indicates negative correlation (Spearman's rank correlations, adjusted P < .05). (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

Aiming to investigate possible host-microbial protein interactions in the human gut, we next investigate the possible correlation between the discriminatory bacterial and human proteins. Interestingly, we found significantly negative correlations between several Pre-DM-enriched E. coli proteins and human proteins involved in innate immune responses (HV304, HV305) and adhesion (CEAM6, CEAM7), whereas positive correlations were found between E. coli proteins and cathepsin G, cytochrome c (CYC) and trypsin−1 (TRY1) (Fig. 4C, adjusted P < .05). Conversely, NGT-enriched proteins from F. prausnitzii showed positive correlations with several NGT-enriched digestive enzymes from the exocrine pancreas, such as chymotrypsin-like elastase family member 3A (CEL3A), chymotrypsinogen B2 (CTRB2) and carboxypeptidases (CBPA1 and CBPB1).

4. Discussion

Our comparative study using metagenomics and metaproteomics in normal glucose tolerant, pre-diabetics and treatment naïve T2D individuals provided novel information with regard to disease-stage specifications at the gut bacterial and host level. Previous studies have reported varied signatures in host metabolic stress, immune responses and gut microbiome in prediabetic individuals [52,53]. A substantial number of Pre-DM associated features were revealed at both the metagenomics and metaproteomics level in this study. We observed several consistent signatures in Pre-DM and NGT as compared to TN-T2D, such as a higher abundance of Akkermansia muciniphila and a lower abundance of Bacteroides spp. The former is a well-known mucin degrading bacterium which may alleviate symptoms of the metabolic syndrome in both mice and humans [54,55]. On the other hand, the relative abundance of several butyrate producing Firmicutes species was lower in Pre-DM and TN-T2D compared to NGT. These findings are in line with a gradual disease development through prediabetes to overt T2D.

Of specific note, we also found a higher abundance of Enterobacteriaceae species (dominated by E. coli) and a lower levels of host proteins which potentially are involved in Proteobacteria-specific responses in Pre-DM, such as galectin-3 and proteins within the immunoglobulin superfamily. An increased abundance of gut Enterobacteriaceae has been widely reported in patients with metabolic diseases such as obesity [56] and atherosclerotic cardiovascular disease [57], and in patients with inflammatory bowel diseases [58]. Hexa-acylated LPS and other Enterobacteriaceae-derived molecules, such as extracellular adhesions and flagellins, can trigger proinflammatory pathways via both innate and adaptive immune responses [59]. These unique Pre-DM associated traits might link potential gut microbial cues to an increase in low-grade systemic inflammation.

Only a modest number of relatively highly abundant faecal proteins were identified in the current study. This reflects the current methodological challenges in microbial protein extraction, identification, and annotation as reported previously [60,61], as well as the detection limitations of MS-based proteomics [62]. For instance, we identified <50 proteins from each of several taxa with median abundances in the 0.1% ranges based on metagenomics data (such as NGT-enriched Dialister, Butyrivibrio and Haemophilus). Nevertheless, metaproteomics provides a valuable addition to not only estimating expression of microbial proteins, but also to delineate host-microbial protein interactions in different disease stages. In this regard, we identified trends towards higher levels of several host-derived AMPs in NGT individuals compared to TN-T2D and Pre-DM individuals, suggesting a possible stronger host defence against invading (disease-related) microbes in NGT individuals. By contrast, significant negative associations were found between Pre-DM-enriched E. coli proteins and several human proteins, including AMPs, adhesion molecules and galectin-3, all involved in intestinal barrier function. It is also worth to note the changes in the levels and types of digestive enzymes identified in the faecal samples, where TN-T2D showed enhanced alpha-amylase (AMY1) levels, as compared to pancreatic-derived lipases and proteases. However, the level of pancreatic alpha-amylase (AMYP) was lower in Pre-DM compared to the two other groups. A metaproteomics study has reported lower faecal AMYP levels in type 1 diabetes (T1D) patients compared to their healthy relatives [14], whereas no difference in levels of AMY1 was reported between T1D and controls, suggesting different amylase responses might be present in Pre-DM, TN-T2D and T1D patients. Differences in the levels of secreted digestive enzymes from the exocrine pancreas in NGT, Pre-DM and T2D have to our notice not been addressed previously, although it may be of major importance in relation to the metabolic state in T2D.

Although patients receiving antidiabetic therapy were excluded in this study, there were still 71 individuals who reported the use of other drugs for treating hypertension or dyslipidaemia, and 49 of them reported the use of calcium channel blockers (CCBs), the most often prescribed antihypertension drug. In addition, the Pre-DM and TN-T2D group had a higher prevalence of comorbidity of metabolic disorders, and thus, the proportion of treated individuals was higher in the Pre-DM and TN-T2D group than in the NGT group. In a pioneer study >1000 drugs were screened against 40 representative gut bacterial strains using an in vitro culture system [63]. The drugs included CCBs, angiotensin II receptor blockers (ARBs), beta blocks and statins. Among them, felodipine (CCB) was shown to inhibit the growth of 17 strains [63], including B. caccae which was enriched in the TN-T2D patients in our study. On the other hand, other antihypertensive and antilipidemic drugs have been reported to have much less impact on the 40 gut bacteria (0–5 strains). In addition, our group has recently evaluated the effects of 10 common drugs on the human gut microbiota in a large-scale Chinese cohort (n = 2, 338) and revealed that metformin and acarbose, but not antihypertensive (CCBs, ARBs and diuretics) or antilipidemic medication (statins) exhibited significant standalone impacts on the entire gut microbial community [64]. Although, we have validated several disease related gut microbial alterations in an independent TN-T2D cohort, further studies are still needed to distinguish reproducible disease omics patterns and the observational microbial differences which might be caused by possible confounding variables such as drugs, age, diet and heathy conditions.

Still, together our findings suggest that unique and nonlinear changes of the intestinal ecosystem might exist in Pre-DM individuals before transition to T2D. Further large-scale, longitudinal follow-up studies are needed to delineate how microbial functions changes from prediabetes to diabetes and to address the nature of interactions between the gut microbiota and the host in the transitional phases leading to overt T2D.

Funding sources

This work was supported by grants from the National Natural Science Foundation of China (No. 31601073), the National Key Research and Development Program of China (No. 2017YFC0909703) and the Shenzhen Municipal Government of China (No. JCYJ20170817145809215). The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.

Author contributions

J.L. and H.Zhong designed and coordinated the study. F.L. and J.Zhang oversaw the blood and faecal sample collection. Y.L, B.C., J.C., X. B., Y.Hu and Y.G. participated in sample collection and provided phenotypic information. G.H., B.Z, J.Zi. and S.L. carried out the metaproteomic experiments. H.Zhong., H.R., F.Y., Z.S, and H.Zou performed the bioinformatic analyses of metagenomic data. H.Zhong, H.R., C.F., B.Z, G.H., Y.Z. and J.Wu performed the bioinformatic analyses of metaproteomic data. H.Z. and H.R performed integrative analyses of metagenomic and metaproteomic data. Y.Z. performed revision of the figures. H.Zhong interpreted together with H.R., J.L., X.X., Y.Hou., H.Y., J. Wang, H.J., S.B. and K.K. the data. H.Zhong wrote the manuscript. J.L., K.K., S.B., and L.M. performed revision of the manuscript. All authors read and approved the final manuscript.

Declaration of Competing Interest

The authors declare no competing interests.

Appendix A. Supplementary data

Supplementary material 1

Supplementary material 2