A Fecal Metabolite Signature of Impaired Fasting Glucose: Results From Two Independent Population-Based Cohorts

Type 2 diabetes (T2D) is a leading cause of mortality and morbidity (1), affecting >536.6 million people (10.5% of the total population) worldwide (2), thus representing a huge public health burden (1). The causation of T2D is multifactorial, influenced by host genetics and environmental factors, including diet, obesity, inactivity, and smoking, and the interaction between these factors (3). Furthermore, its onset is gradual, with people progressing through a state of prediabetes (4), and is defined as impaired levels of fasting glucose (IFG), and/or glucose intolerance, and/or elevated hemoglobin A_1c (HbA_1c) (5).

Over the past decade, T2D and prediabetes have been linked by us and others (6–8) to changes in the gut microbiota, and we have recently demonstrated that T2D development is preceded by an alteration in gut microbiota composition (7). A critical challenge in human microbiome research, however, is to characterize and quantify metabolic activity across the full microbial ecosystem (9). The gut microbiome is highly variable, and different bacterial types may have similar metabolic effects on the host. Microbial metabolites are now widely seen as key mediators of the effects of gut microbiome composition on human physiology (10). Fecal metabolites provide a functional readout of the gut microbiome (11,12) and are a novel tool to explore links between gut microbiome composition and activity, host phenotypes, and heritable complex traits, thus improving our understanding of the impact that the gut microbiome can have on its host (11). As the gut microbiome is modifiable with nutritional and lifestyle interventions (13), it is of utmost importance to identify alterations in the fecal metabolites abundances, which reflect metabolic activity perturbations of the human gut microbial ecosystem that might lead to T2D onset.

In the first fecal metabolomics study of prediabetes to date, we aim to identify a fecal metabolite signature of this condition in two independent cohorts to shed light on mechanisms of action underlying T2D onset and development. Addressing this challenge also has long-term implications for future studies into therapies and lifestyle interventions that alter microbial metabolic activity to improve human health.

A flowchart of the study design with the main results is presented in Fig. 1.

We analyzed data from 1,247 nonrelated individuals from UK Adult Twin Registry (TwinsUK) (14), for whom concurrent nontargeted fecal metabolomic profiling (526 metabolites at fasting) and glucose/diabetes information were available (cross-sectional design). Concurrent metagenome sequencing (as a measure of the gut microbiome composition) was also available for a subset of 342 individuals. Subjects were classified into three groups following the American Diabetes Association criteria based on isolated fasting glucose levels (15) at the time of the initial sampling and at subsequent visits (on average, 3.5 [SD 2.0] visits, 4.6 [SD 2.7] years apart): individuals with T2D (fasting glucose ≥7 mmol/L or physician’s letter confirming diabetes diagnosis), individuals with IFG (fasting glucose >5.5 to <7 mmol/L, not on diabetes medication), and subjects without IFG and T2D (fasting glucose >3.9 to ≤5.5 mmol/L) (see Table 1). We refer to “healthy individuals” to indicate individuals without IFG and/or T2D.

Only one twin per twin pair was included in the analyses to eliminate potential bias through correlated error, which might inflate effect estimates.

In a small subanalysis, we included individuals with incident T2D (average follow-up time 2.1 [SD 1.3] years) and an independent subset of healthy individuals who remained healthy during follow-up.

All twins provided informed written consent and the study was approved by St Thomas’ Hospital Research Ethics Committee (REC Ref: EC04/015).

The Cooperative Health Research in the Region of Augsburg (KORA) study is a population-based cohort study. The KORA FF4 study (2013–2014) is the second follow-up of KORA S4 (1999–2001). The 1,007 samples included in the study were collected in the morning between 8:00 a.m. and 10:30 a.m. after at least 8 h of fasting. Metabolon untargeted liquid chromatography/mass spectrometry (MS)-based techniques were applied to measure the metabolites in the KORA cohort (a different version of the platform used in TwinsUK). Healthy individuals and IFG individuals were assigned based on the same criteria as in TwinsUK (described in the above section and in Table 1).

Metabolomics profiling was conducted using ultrahigh-performance liquid chromatography-tandem MS (MS/MS) by the metabolomics provider Metabolon Inc. (Morrisville, NC) on fecal samples from participants in the TwinsUK and KORA cohorts (Supplementary Material). The metabolomic data set measured by Metabolon includes 526 known metabolites for TwinsUK belonging to the following broad categories—amino acids, peptides, carbohydrates, energy intermediates, lipids, nucleotides, cofactors and vitamins, and xenobiotics—of which 357 were also measured in KORA. These include metabolites of established microbial origin (16). A complete list of the included metabolites with their superpathways, subpathways, Kyoto Encyclopedia of Genes and Genomes and Human Metabolome Database identifiers are reported in Supplementary Table 1. We imputed to the day minimum metabolites with <20% missing.

Gut microbiota composition was generated from fecal shotgun metagenomes for a subset of the discovery cohort. DNA extraction, library preparation, and sequencing were conducted as detailed in Visconti et al. (11). For details see the Supplementary Material. Of note, gut microbiota composition is described by species-level genome bins (SGBs), which is the best proxy to define microbial species (17).

Statistical analyses were conducted using R 4.2.2 software. To identify a fecal metabolite signature of prediabetes, we ran logistic regressions adjusting for age, BMI, sex, and multiple testing using the Benjamini and Hochberg method (18) (false discovery rate [FDR] <0.05). We then checked whether the metabolites significantly associated with IFG in the discovery set were also replicated in KORA (P < 0.1). We used a less stringent threshold for KORA because of the winner’s curse (the effect sizes of the most strongly associated variables within a cohort-specific analysis are inflated) (19). Results were meta-analyzed using inverse-variance random-effect meta-analysis. We then created the IFG-metabolite score by linearly combining the replicated metabolites along with covariates. To assess the performance of the score in predicting prevalent IFG and incident T2D, we calculated the area under the curve (AUC) values obtained using fivefold cross-validation (caret package implemented in R [20]). Finally, logistic and Cox regressions were used to investigate the association between the IFG score (Z-scaled) and prevalent IFG risk and incident T2D risk, respectively.

Given the strong association between fecal metabolites and gut microbiome composition (12), we investigated to what extent the gut microbiota composition was associated with each of the replicated metabolites using random forest regressors and classifiers with compositional data and fivefold cross-validation. The performance was calculated using the average of the obtained Spearman correlations between the observed metabolite levels and the levels predicted by the model (denoted as ρ) over the fivefolds used as a test set for the regressors and the average of the obtained AUC values over the testing folds for the classifiers. For details see the Supplementary Material.

We further investigated the associations between their top 100 bacterial features and IFG by running logistic regression models adjusting for covariates and multiple testing species (FDR <0.05). Specifically, we included all of the fecal metabolites that could be predicted by the gut microbiome with an AUC >70%, and we then focused on those that had an outstanding prediction performance (AUC >90%).

Finally, we used formal mediation analysis as implemented in the R package “mediation” with 1,000 nonparametric bootstrap samples (21) to test the mediation effects of the metabolites on the total effect of the gut bacteria on IFG. The mediation model was used to quantify both the direct effect of these gut bacterial species on IFG and the indirect (mediated) effects mentioned above while controlling for age, BMI, and sex. The variance accounted for (VAF) score, which represents the ratio of indirect-to-total effect and determines the proportion of the variance explained by the mediation process, was used to determine the significance of the mediation effect.

The data used in this study are held by the Department of Twin Research at King’s College London. The data can be released to bona fide researchers using our normal procedures overseen by the Wellcome Trust and its guidelines as part of our core funding (https://twinsuk.ac.uk/resources-for-researchers/access-our-data/). The gut microbiome data are available on EBI (https://www.ebi.ac.uk/) under accession number PRJEB32731 (TwinsUK). The KORA FF4 datasets are available upon application through the KORA-PASST (project application self-service tool, https://www.helmholtz-munich.de/epi/research/cohorts/kora-cohort/data-use-and-access-via-korapasst/index.html).

We included 1,247 unrelated individuals from the TwinsUK cohort who had fecal metabolite measures along with glucose/diabetes and prediabetes information. Of these, 142 individuals had IFG (mean fasting glucose 5.9 mmol/L [SD 0.4]) and 1,105 were healthy individuals (mean fasting glucose 4.5 mmol/L [SD 0.3]). Descriptive characteristics of the discovery and replication populations are included in Table 1.

Of the 526 known fecal metabolites analyzed in TwinsUK, the fecal abundances of 26 compounds were associated with IFG after adjusting for age, BMI, sex, and multiple testing (FDR <0.05) (Fig. 2). Identified metabolites were mainly amino acids (n = 7) and lipids (n = 7), but also included xenobiotics (n = 4), cofactors and vitamins (n = 3), nucleotides (n = 2), carbohydrates (n = 2), and one energy-related metabolite (Fig. 2). All significant metabolites, but 3-hydroxyoleate, octadecanedioate (C18-DC), azelate (C9-DC), γ-tocotrienol, and enterolactone, were positively associated with IFG (Fig. 2). Of the 26 metabolites, 18 were also measured in KORA (Supplementary Table 1), and 8 metabolites were replicated (P < 0.1) (Fig. 3). These were the lipid cholesterol (sterol metabolism), the carbohydrate glucuronate (aminosugar metabolism), the cofactors/vitamins nicotinate (nicotinate and nicotinamide metabolism) and protoporphyrin IX (hemoglobin and porphyrin metabolism), the xenobiotics caffeine and 1-methylxanthine (both involved in the xanthine metabolism), the amino acid serine (glycine, serine, and threonine metabolism), and the nucleotide uridine (pyrimidine metabolism). The correlation matrices for the eight fecal metabolites in TwinsUK and KORA are depicted in Supplementary Fig. 1. We combined the results from both cohorts using inverse-variance random-effect meta-analysis (Fig. 3).

We then generated the IFG-metabolite score using TwinsUK individuals:

IFG-metabolite score = −8.79 + 0.07 × glucuronate + 0.25 × protoporphyrin IX + 0.09 × 1-methylxanthine + 0.14 × cholesterol + 0.04 × serine + 0.07 × uridine + 0.04 × nicotinate + 0.17 × caffeine + 0.07 × age + 0.1 × BMI − 0.6 × sex (female = 1)

The IFG-metabolite score was associated with an increased risk of IFG in TwinsUK (odds ratio [OR] 3.9 [95% CI 43.02–5.02], P < 0.0001) and in KORA (OR 1.3 [95% CI 1.16–1.52], P < 0.0001). The association remained significant when further adjusting for clinical covariates (i.e., systolic and diastolic blood pressure, circulating levels of HDL, total cholesterol, and triglycerides, alternative health eating index [aHEI – not available in KORA], activity levels and smoking status) (Table 1) in both cohorts (TwinsUK: OR 3.4 [95% CI 2.65–4.49], P < 0.0001; KORA: OR 1.2 [95% CI 1.06–1.41], P = 0.008). Finally, the IFG-metabolite score accurately predicted prevalent IFG in TwinsUK with an AUC of 79.8% (95% CI 76.3–83.3) in fivefold stratified cross-validation and outperformed the model including only covariates (AUC 77.2% [95% CI = 73.6–81]) by 2.6% (Δ95% CI 2.7–2.1). In KORA, the IFG-metabolite score (top vs. lowest decile) could satisfactory predict prevalent IFG (AUC 65.4 [95% CI 57.9–73]).

In a small independent sample from TwinsUK (descriptive characteristics are shown in Table 1) consisting of 17 healthy individuals (different from the healthy subjects of the IFG data set) and 10 individuals with incident T2D (follow-up time between fecal metabolite measurements and incident events: mean 2.1 [SD 1.3] years), the IFG-metabolite score was also predictive of an increased risk of incident T2D (hazard ratio 4 [95% CI 1.97–8], P = 0.0002) in TwinsUK after further adjusting for baseline circulating glucose levels. It also accurately predicted incident T2D (AUC 83.3% [95% CI 74.4–92.2]), while a model using baseline circulating glucose levels as predictor presented a lower prediction power (AUC 72.4% [95% CI 51.8–92.9]).

We further evaluated the extent to which the gut microbiota was associated with the fecal abundances of the eight replicated metabolites using the AUC obtained by the random forest classifiers and the Spearman correlations (denoted as ρ) between the real abundances and predicted values by the random forest regressors. We included a subset of 342 individuals from TwinsUK with concurrent gut microbiota composition assessed by shotgun metagenomics and fecal metabolites measurements. Descriptive characteristics of this subset are shown in Supplementary Table 2.

The gut microbiome composition was strongly associated with the replicated metabolites, with performance metric values ranging from an AUC of 70.7% (95% CI, 69.1–72.4) and ρ = 0.24 (95% CI, 0.23–0.25) for caffeine to an AUC of 91.4% (95% CI, 90.8–91.9) and ρ = 0.62 (95% CI, 0.62–0.62) for 1-methylxanthine (Fig. 4A and Supplementary Table 3). Protoporphyrin IX was the only metabolite presenting a moderate association (AUC 64.8% [95% CI 63.9–65.6]; ρ = 0.25 [95% CI 0.24–0.26]) (Fig. 4A).

We then investigated whether the abundances from their top 100 bacterial features based on the random forest models were also significantly associated with IFG (Supplementary Table 4). We focused on the fecal metabolites that presented the strongest associations with the gut microbiome composition (AUC >90%—outstanding prediction performance; 1-methylxathine and nicotinate). We identified four characterized gut bacterial species for 1-methylxanthine and nicotinate, of which three overlapping (overlapping: Dorea formicigenerans, Ruminococcus torques, and Faecalibacillus intestinalis; 1-methylxanthine only: Dorea sp. AF24-7LB; nicotinate only: Dorea sp. AF36-15AT), that were positively associated with IFG after adjusting for age, BMI and sex (FDR <0.05) (Supplementary Table 4). We, therefore, performed a formal mediation analysis adjusting for age, BMI, and sex to determine whether 1-methylxanthine and/or nicotinate mediated the associations between these species and IFG. The analysis revealed that 1-methylxanthine acted as a potential mediator in the positive associations of Dorea sp. AF24-7LB (VAF = 10.3%, P = 0.03) and R. torques (VAF = 9.7%, P = 0.04) with IFG, while nicotinate acted as a potential mediator in the positive associations of F. intestinalis (VAF = 22.3%, P = 0.002), D. formicigenerans (VAF = 15.8%, P = 0.002), and R. torques (VAF = 14.1%, P = 0.03) with IFG (Fig. 4B). We further ran mediation analyses for the metabolites that could be predicted by the gut microbiome with an AUC >70%. As reported in Supplementary Fig. 2, uridine, serine, cholesterol, and caffeine were also mediators in the associations between different species (e.g., Dorea spp. and Anaerobutyricum hallii) and IFG. Models were not further adjusted for other comorbidities (e.g., systolic and diastolic blood pressure, circulating levels of HDL, total cholesterol and triglycerides, aHEI, activity levels, and smoking status) as these were not significantly associated with the identified bacterial species or with the metabolites making up the score (Supplementary Table 5).

Here we identify for the first time a fecal metabolite signature of IFG that is associated with prevalent IFG in two independent cohorts and is also predictive of incident T2D in a small subanalysis. The fecal metabolites making up the score are not microbial-derived metabolites but are “host metabolites” (e.g., xenobiotics, cofactors, and vitamins). However, the gut microbiome can accurately predict their fecal abundances (AUC >70%). It is well known that the gut microbiome composition can affect diseases via several mechanisms (22). Circulating microbial metabolites have been reported by us and others to be reflective of gut microbiome diversity and composition (6–8) and predictive of prevalent and incident T2D (7). Taken together, this suggests that the gut microbiome can influence T2D, not only by producing metabolites that enter the bloodstream (7) but also by regulating the absorption or excretion of host-produced compounds, thereby influencing IFG and T2D risk. This hypothesis is further supported by the results of our mediation analysis showing that metabolites making up the score act as partial mediators on the significant associations between several gut microbial species, (e.g., F. intestinalis, D. formicigenerans, R. torques, and Dorea sp. AF24-7LB) and IFG.

Studies have shown that gut microbiome composition differs between individuals with prediabetes/diabetes and healthy subjects (6,7), with compositional shifts correlated with synthesis profile changes of gut bacteria-derived metabolites, including short-chain fatty acids, indolepropionic acid, and trimethylamine (7,22). These “microbial” metabolites enter into the bloodstream and reach different tissues, where they can influence glucose homeostasis and insulin resistance by activating or inhibiting signaling pathways (22). Nevertheless, the identified signature of prediabetes in this study consists of eight metabolites of nonbacterial origin. Serine is a nonessential amino acid mainly obtained by intrinsic synthesis (23). Glucuronate is a sugar acid derived from glucose and involved in the detoxification of xenobiotic compounds (24). Protoporphyrin IX is a cofactor ubiquitously present in the human body as a heme precursor (25). Nicotinate, also known as vitamin B₃ and niacin, is a water-soluble vitamin that can be produced by the human body from tryptophan (26). Cholesterol, which is mainly produced by the liver, is an essential lipid of eukaryotic cell membranes and is also a precursor of bile acids and steroid hormones (27). Uridine is a necessary pyrimidine nucleotide for RNA synthesis produced by several reversible reactions (e.g., dephosphorylation of uridine monophosphate, deamination of a cytidine or combination of uracil and ribose 1-phosphate) (28). Caffeine and 1-methylxanthine are xenobiotics involved in the caffeine metabolism pathway (29).

Strikingly, we find that the gut microbiome is strongly associated with fecal levels of these metabolites, suggesting that the gut microbiome influences the absorption or excretion of compounds involved in various metabolic pathways (e.g., cholesterol, uridine, and glucuronate) and xenobiotics (e.g., caffeine and its derivatives), among others, and such levels of absorption or excretion are directly related to IFG. Our findings lead us to speculate that individuals with prediabetes present gut microbiome composition perturbations, which likewise influence the absorption or excretion of the identified compounds. This is further supported by the mediation analyses, which suggest that the associations between specific gut microbial species, including F. intestinalis, D. formicigenerans, R. torques, and Dorea sp. AF24-7LB, and IFG are mainly reflecting the effect of the gut microbiome in the absorption or excretion of the found compounds.

Under normal conditions, the small intestine can break down, emulsify, and absorb most nutrients, including fats, simple carbohydrates, and proteins (30). For instance, <5 g/day of fat are not absorbed and reach the colon (30). Nonetheless, the absorption capability of the gut can be limited depending on the gut microbiome composition (31). A study conducted by Basolo et al. (31) demonstrated that changes in participants’ gut microbiome composition, due to diet or antibiotic use, impaired nutrient absorption. Several mechanisms might explain how gut microbiome composition might influence absorption, and thus, the disease onset (32–34). For instance, the gut microbiome can affect the gut barrier, which consists of a collection of physical and chemical structures that protect the host from pathogenic invasions and harmful stimuli (32). This can be provoked by the presence of pathogen-associated molecular patterns, such as lipopolysaccharides, in the cell walls of some gram-negative bacteria, which play an important role in intestinal absorption, blood glucose, and inflammation (33). Moreover, changes in the permeability of the gut barrier can be caused by an unbalanced increase in bacteria able to degrade mucin (the main component of mucus, which covers the epithelial surfaces of the gastrointestinal tract) (32). Indeed, in this study, we identify that individuals with prediabetes present larger abundances compared with healthy individuals of the mucin-degraders D. formicigenerans (35) and R. torques (36), which have been previously associated with lower nutrient absorption (36). Finally, some gut microbes can also reduce absorption in the jejunum by altering the expression of intestinal transporters of different types of compounds (34).

Another possible explanation for our findings could be a reduction of specific beneficial bacteria able to use these compounds, thus resulting in increased excretion (27,37). In the case of cholesterol, bacterial members of the genera Bifidobacterium, Lactobacillus, and Peptostreptococcus are needed to convert cholesterol into coprostanol (27). Likewise, an inefficient cholesterol-coprostanol conversion is linked to cardiometabolic diseases (27). For glucuronate, most of it is not absorbed by the small intestine; however, under normal conditions, the amounts that make it to the colon are then efficiently used by Bifidobacterium (37).

This work has several strengths. Our study benefits from a large, accurately phenotyped discovery cohort with metabolomic profiling and gut microbiome composition. We were also able to replicate our findings in a large independent cohort, thus strengthening our findings. Finally, a machine learning algorithm was applied to investigate the prediction of the gut microbiota to the levels of the found eight metabolites, allowing us to simultaneously integrate all the species in the models.

We also note some study limitations. First, the cross-sectional nature of the data used for our primary analysis does not allow us to determine the temporal link between IFG and the identified fecal metabolites.

Second, HbA_1c, postprandial glucose to derive impaired glucose tolerance, which more closely resembles the T2D state (38), and a clinician’s diagnosis were not available in the discovery cohort. Thus, the division of categories in this study is derived from IFG.

Third, the sample size for the subanalysis looking at incident T2D was limited, and we were unable to seek independent replication as, to the best of our knowledge, there are no other cohorts in the world that have measured this fecal metabolome panel and incident T2D. Future studies with larger sample size are therefore needed to test the robustness of the IFG metabolite score to predict incident T2D.

Fourth, there was not a full overlap between metabolites measured in the discovery and validation data sets, which might cause the loss of metabolites of interest to study.

Fifth, the included study groups were unbalanced in age and sex. Hence, although we adjusted all analyses for them and other important clinical variants, the confidence of the results is lowered. In addition, gut microbiota composition data were only available for a subset from the discovery set, and therefore, we could not replicate the mediation analysis in KORA. Furthermore, the Spearman correlations between the predicted (from gut microbiome composition) and actual levels of the metabolites were modest. Indeed, random forest models were trained based on microbial features extracted from metagenomic data, which does not retrieve all species present in a microbiome sample for procedural and technical reasons.

Finally, this study does not include measures of permeability markers, which would contribute to a better understanding of the role of intestinal permeability in the absorption or excretion of the identified compounds.

In conclusion, we are proposing a novel mechanism of how gut microbiome composition affects prediabetes and, consequently, the onset of T2D. The gut microbiome is linked to prediabetes not only by microbial-derived metabolites but also by affecting intestinal absorption or excretion of metabolites of nonmicrobial origin, which are correlated with the risk of IFG and incident T2D. Henceforth, to better understand the onset of T2D, the effect of the gut microbiome in the excretion and/or absorption of host-produced compounds and xenobiotics also needs to be also considered.

Acknowledgments. The authors thank all the participants of TwinsUK for contributing and supporting their research and all participants of KORA for their long-term commitment to the KORA study, the staff for data collection and research data management, and the members of the KORA Study Group (https://www.helmholtz-munich.de/en/epi/cohort/kora) who are responsible for the design and conduct of the study. Finally, the authors thank Caroline Le Roy (Department of Twin Research & Genetic Epidemiology, King’s College, London) for the helpful discussion. Data collection in the KORA study is done in cooperation with the University Hospital of Augsburg.

Funding. This research was funded by the Chronic Disease Research Foundation, and in part by the Wellcome Trust (grant no. 212904/Z/18/Z) and by DiabetesUK (19/0006053). TwinsUK receives funding from the Wellcome Trust, the European Commission H2020 grants SYSCID (contract #733100), the National Institute for Health Research (NIHR) Clinical Research Facility, and the Biomedical Research Centre based at Guy's and St Thomas' NHS Foundation Trust in partnership with King's College London, the Chronic Disease Research Foundation, the UKRI Medical Research Council (MRC)/British Heart Foundation Ancestry and Biological Informative Markers for Stratification of Hypertension (AIM-HY, MR/M016560/1). This work was also supported by UKRI grant MR/W026813/1 to C.M. and A.M.V. The KORA study was initiated and financed by the Helmholtz Zentrum München–German Research Center for Environmental Health, which is funded by the German Federal Ministry of Education and Research (BMBF) and by the State of Bavaria. Stool sample collection and metabolomics analysis in KORA FF4 was supported by iMED, a research alliance within the Helmholtz Association, Germany. C.M., L.P., and A.N. are funded by the Chronic Disease Research Foundation. A.M.V. is supported by the National Institute for Health Research Nottingham Biomedical Research Centre.

Duality of Interest. This study received support from Zoe Ltd. T.D.S. is a cofounder and shareholder of Zoe Ltd. A.M.V., P.W.F., F.A., and N.S. are consultants to Zoe Ltd. K.W. and G.A.M. are employees of Metabolon Inc. No other potential conflicts of interest relevant to this article were reported.

Author Contributions. A.N., F.T., Q.D., A.V., C.C., H.G., F.A., and C.G. curated the data. A.N., F.T., Q.D., and C.M. performed the formal analyses. A.N., A.M.V., and C.M. wrote the manuscript. Q.D., P.L., A.V., C.C., T.B., J.L., H.G., N.W., F.A., K.W., A.-F.B., G.A.M., N.S., M.F., A.P., P.W.F., V.B., T.D.S., J.T.B., and C.G. contributed reagents/materials/analysis tools. T.D.S. and C.M. conceived and designed the experiments. All of the authors revised the manuscript. C.M. is the guarantor of this work and, as such, had full access to all the data in the study and takes responsibility for the integrity of the data and the accuracy of the data analysis.