gut microbiome predictors of treatment response and...

Gut microbiome predictors of treatment response andrecurrence in primary Clostridium difficile infectionS. Khanna*,1, E. Montassier†,‡,1, B. Schmidt*, R. Patel§,¶, D. Knights‡, D. S. Pardi* & P. C. Kashyap*

*Division of Gastroenterology andHepatology, Mayo Clinic, Rochester,MN, USA.†EA 3826 Th�erapeutiques Cliniques etExp�erimentales des Infections, Facult�ede M�edecine, Universit�e de Nantes,Nantes, France.‡Department of Computer Scienceand Engineering, University ofMinnesota, Minneapolis, MN, USA.§Division of Infectious Diseases, MayoClinic, Rochester, MN, USA.¶Division of Clinical Microbiology,Department of Medicine, Mayo Clinic,Rochester, MN, USA.

Correspondence to:Dr P. C. Kashyap and Dr D. S. Pardi,Division of Gastroenterology andHepatology, Mayo Clinic, 200 FirstStreet SW, Rochester, MN 55905,USA.E-mails: [email protected] [email protected]

1Denotes equal contribution

Publication dataSubmitted 13 April 2016First decision 29 April 2016Resubmitted 23 June 2016Resubmitted 10 July 2016Accepted 10 July 2016

The Handling Editor for this article wasProfessor Jonathan Rhodes, and it wasaccepted for publication after full peer-review.

SUMMARY

BackgroundClostridium difficile infection (CDI) may not respond to initial therapy andfrequently recurs, but predictors of response and recurrence are inconsis-tent. The impact of specific alterations in the gut microbiota determiningtreatment response and recurrence in patients with CDI is unknown.

AimTo assess microbial signatures as predictors of treatment response andrecurrence in CDI.

MethodsPre-treatment stool samples and clinical metadata including outcomes werecollected prospectively from patients with their first CDI episode. Next genera-tion 16s rRNA sequencing using MiSeq Illumina platform was performed andchanges in microbial community structure were correlated with CDI outcomes.

ResultsEighty-eight patients (median age 52.7 years, 60.2% female) were included.Treatment failure occurred in 12.5% and recurrence after response in 28.5%.Patients who responded to treatment had an increase in Ruminococcaceae,Rikenellaceae, Clostridiaceae, Bacteroides, Faecalibacterium and Rothia comparedto nonresponders. A risk-index built from this panel of microbes differentiatedresponders (mean 0.07 � 0.24) from nonresponders (0.52 � 0.42; P = 0.0002).Receiver operating characteristic (ROC) curve demonstrated that risk-index wasa strong predictor of treatment response with an area under the curve (AUC) of0.85. Among clinical parameters tested, only proton pump inhibitor use pre-dicted recurrent CDI (OR 3.75, 95% CI 1.27–11.1, P = 0.01). Patients with recur-rent CDI had statistically significant increases in Veillonella, Enterobacteriaceae,Streptococci, Parabacteroides and Lachnospiraceae compared to patients withoutrecurrence and a risk index was able to predict recurrence (AUC = 0.78).

ConclusionGut microbiota signatures predict treatment response and recurrence poten-tially, allowing identification of patients with Clostridium difficile infectionthat may benefit from early institution of alternate therapies.

Aliment Pharmacol Ther

ª 2016 John Wiley & Sons Ltd 1

doi:10.1111/apt.13750

Alimentary Pharmacology and Therapeutics

INTRODUCTIONClostridium difficile infection (CDI) is a common infec-tion in the USA with 450 000 infections and 29 000deaths per year,1–3 with a tremendous economic impact(attributable costs ranging from $8911 to $30 049 perhospitalised patient).4 The pathophysiology of CDI iscomplex with alterations in gut microbiota playing animportant role in susceptibility to CDI.5 A recent studyreported an increase in Firmicutes, Proteobacteria andActinobacteria and decreases in Bacteroidetes in CDIpatients, and an altered ratio of Bacteroidetes to Firmi-cutes was found to be significant after controlling forconfounding factors.6 Another study demonstrated thatseveral bacterial species within the Ruminococcaceae,Lachnospiraceae, Bacteroides and Porphyromonadaceaewere largely absent in CDI cases and highly associatedwith nondiarrhoeal controls.7 Similarly, alterations inRuminococcus gnavus, certain Enterobacteriaceae, Verru-comicrobia and Enterococcus have been related to devel-opment of CDI.7–9 Risk factors for CDI includeincreasing age, antibiotic exposure, proton pump inhibi-tor use, hospitalisation, immunosuppression and comor-bidities.10–15 Among these, gut microbiota changesassociated with antibiotic use have been best studied inanimal models and provide strong evidence for alter-ations in gut microbiota composition and function insusceptibility to primary and recurrent C. difficile.5, 16

The management of CDI includes use of oral antibi-otics but the rate of treatment failure (up to 35%) andrecurrence (60% or higher after 3 episodes) is concern-ing.17, 18 Alternate treatment strategies such as faecalmicrobiota transplant are highly effective in patients withrecurrent CDI.19, 20 Studies have assessed clinical fea-tures which may predict CDI response, but there is alack of robust clinical or microbial biomarkers predictiveof response to primary therapy.21–23 Severe CDI andhospital admission are clinical predictors of metronida-zole failure, but these have not been validated. Increasingage, concomitant antibiotic use, decreased anti-toxin IgGlevels, the presence of comorbidities and potentially theuse of gastric acid suppression medications have beenassociated with recurrent CDI. Anti-toxin IgG levels arenot clinically available.24, 25

Despite the identification of these risk factors, clinicalmodels to predict the risk of recurrent CDI are notrobust enough for routine clinical use and there are nomodels to predict response to initial antibiotic ther-apy.24, 26 Hence, there is a need for biomarkers and bet-ter models to predict these treatment outcomes. Whilethe role gut microbiota alterations play in CDI

susceptibility has been established,27 their role in deter-mining outcomes of CDI treatment has not been investi-gated. We used next generation sequencing tocharacterise microbial communities in patients with pri-mary CDI to identify differences in microbial commu-nity structure and key taxa that can predict treatmentresponse and the risk of recurrence after successful treat-ment.

MATERIALS AND METHODS

Study designWe prospectively recruited 88 patients (median age52.7 years, interquartile range 36.9–65.1; 60.2% female)with their first CDI episode (from 3/2012–9/2013) asidentified from the Clinical Microbiology Laboratory atMayo Clinic, Rochester, Minnesota and collected an ali-quot from the stool samples that led to the diagnosis.Details on clinical data acquisition and analysis are out-lined in Data S1.

SEQUENCING AND ANALYTIC METHODS

16S rRNA gene sequencing and data analysesAfter faecal DNA isolation (fecal DNA kit; MoBio, Carls-bad, CA, USA), amplicons spanning the variable region4 of bacterial 16S rRNA were generated and sequencedusing MiSeq Illumina platform at the Mayo Clinic Medi-cal Genome Facility, Rochester, MN, USA. We used515F TATGGTAATTGTGTGCCAGCMGCCGCGGTAAand barcoded 806R primers AGTCAGTCAGCCGGAC-TACHVGGGTWTCTAAT. The amplicons were ~400base pairs. The sequencing was bidirectional but onlyforward reads were used in the study due to low qualityscores of reverse sequences. We applied built-in func-tions of the Quantitative Insights Into Microbial Ecology(QIIME) pipeline for data analysis.28 The 16S rRNAsequencing data from the Illumina runs were trimmed,demultiplexed, chimera filtered and assigned to opera-tional taxonomic units (OTUs) using packages imple-mented in QIIME 1.8.0 software.28 Further details of 16SrRNA data analysis are outlined in Data S1.

In order to study differences in alpha diversitybetween our cohort of patients and healthy subjects, weincluded data from healthy subjects from a previouslypublished study (global gut dataset).29 The differences ingut microbiota composition between healthy controlsand patients with CDI have been previously described.7

We included adult US individuals and compared the fol-lowing alpha diversity metrics with our cohort of CDI

2 Aliment Pharmacol Ther

ª 2016 John Wiley & Sons Ltd

S. Khanna et al.

patients: phylogenetic diversity whole tree, observed spe-cies and Chao 1 index. The methodology applied in boththe studies was similar as samples were stored at �80 °Cin both studies. DNA extraction was done using MOBIO power soil DNA isolation kit in both studies andthe same protocol (Earth Microbiome Protocol) was fol-lowed including bead beating the faecal samples and thesame region of 16S was amplified using the same pri-mers and the same PCR protocol.30

In order to identify significant associations betweenmicrobial community abundances and clinical metadata,we applied a linear multivariate regression model specifi-cally developed for microbiome data (MaAsLin, Multi-variate microbial Association by Linear models).31 Weused default parameters within MaAsLin as described byMorgan et al.31 Clinical data were selected by boosting,as microbial communities are known to be high dimen-sional, to identify those most associated with each micro-bial feature over potential covariates. Selected clinicaldata were then used in a general linear model includingclinical data as predictor and taxonomic relative abun-dances as response.

Comparisons of relative abundance of taxa betweenresponders and nonresponders, or between patients withand without recurrence after initial response, was per-formed using linear discriminant analysis effect size(LEfSe), a nonparametric Mann–Whitney U test appliedto detect features with significant differential abundancewith respect to the groups compared, followed by a lin-ear discriminant analysis (LDA) to estimate the effectsize of each differentially abundant feature.32 As pro-posed, an LDA score (log 10) >2 was considered signifi-cant.

A risk index was built to differentiate responders fromnonresponders to initial treatment or patients with andwithout recurrence after initial response, based on taxon-omy on a noncollapsed OTU table. All the taxa with aLDA score (log10) >2 were included in the calculation ofthe risk index. In order to build the risk index to predictthe patients who would respond or not respond to treat-ment, the relative abundances (arcsine square root trans-formed) of the taxa associated with the responders totreatment (based on the LEfSe output, all taxa with aLDA score (log10) >2 were summed and the relativeabundances of the taxa associated with nonresponders totreatment (based on the LEfSe output, all taxa with aLDA score (log10) >2 were summed. Then, the differencebetween these two sums (relative abundance of the taxaassociated with no response to treatment minus relativeabundance of the taxa associated with response to

treatment) was calculated, thereby obtaining a risk index.This procedure was repeated n (overall sample size)times to obtain a risk index for each patient in thecohort. Therefore, a risk index was calculated for eachpatient. We had 11 risk indexes in the nonrespondersand 77 risk indexes in the responders, 22 risk indexes inpatients who had a recurrence and 55 risk indexes inpatients who did not have a recurrence of CDI. The util-ity of a risk index to predict a clinical outcome wasrecently reported by our group.33

A leave-one-out cross-validation procedure was alsoconducted. This procedure calculates the risk index onthe n-1 patients (taxa that differentiated patients basedon the LEfSe output) and then tested the risk index inthe held-out patient, that is, the risk index values arepredicted for each patient using a panel of microbesretrained from the other patients. A detailed descriptionof the leave-one-out cross-validation and receiver operat-ing characteristic (ROC) curve analysis are outlined inData S1.

Network analyses were carried out with Cytoscapeusing an edge-weighted spring-embedded layout.34

Importantly, we collapsed OTUs to the genus level andeliminated OTUs present in fewer than 25% of samples,to reduce the very large number of multiple hypothesestested in correlation network analysis. We then per-formed Spearman correlation of taxon–taxon relativeabundance and included only those links with absolutevalue of correlation > 0.5 and false discovery rate (FDR)-corrected P < 0.05.

In order to determine if competition for nutritionalniches in the gut may play a role in determiningresponse to treatment by providing C. difficile with acompetitive disadvantage, we used Phylogenetic Investi-gation of Communities by Reconstruction of UnobservedStates (PICRUSt) to predict Carbohydrate-ActiveEnzymes database–glycoside hydrolase (CAZY-GH)assignments. Details on PICRUSt analysis are outlined inData S1.

RESULTS

Clinical characteristics fail to predict response tostandard therapyThe rate of primary nonresponse following recom-mended treatment was 12.5%. Patients were treated witheither metronidazole or vancomycin (Table 1). Amongthose who had an initial successful response, 28.5% hadrecurrent CDI. Prior antibiotic exposure was noted in59.1% of patients, and 78.8% of those with antibiotic

Aliment Pharmacol Ther 3


Microbiome predictors of C. difficile outcomes

exposure (or 36.1% of all 88 patients) were exposed to asingle agent, 9.7% (or 5.7% of all 88 patients) to a com-bination of antibiotics and exact antibiotic name wasunknown in the remaining six patients. Details onantibiotic use are summarised in Table S1. Chi-squareunivariate analyses demonstrated no significant differ-ences in any clinical variables including initial treatmentin primary nonresponders compared to responders(Table 1).

Gut microbiota signatures prior to treatment predictresponse to therapy in CDIOf the 88 faecal samples collected, a total of 1 449 211high-quality 16S rRNA gene-encoding sequences wereidentified, representing 7470 OTUs. The mean numberof sequences obtained per sample was 16 468 � 4674.Importantly, since samples contained between 6987 and35 494 sequences, diversity analyses were rarefied at6987 sequences per sample to avoid bias.

We assessed the relationship between clinical metadataand microbial measurements by identifying associationsusing MaAsLin (see methods above). This model investi-gated the relationship between microbial taxa collapsedat genus level with metadata of interest while accountingfor other covariates. In our cohort of patients, we did

not find significant associations between gut microbiotaand the clinical characteristics (for example, Age, Sex,BMI, Charlson comorbidity Index) (Table S2). A panelof 36 OTUs that were significantly different between pri-mary nonresponders and responders using LEfSe [corre-sponding to an LDA (log10) >2] were identified.Responders had a significant increase in relative abun-dance of OTUs within Ruminococcaceae, Rikenellaceae,Bacteroides and Faecalibacterium, while nonrespondershad a significant increase in Clostridiaceae, Lach-nospiraceae, Blautia, Coprococcus, Streptococcus, Bifi-dobacterium, Ruminococcus and Actinomyces (Figure 1).

Furthermore, the ability of this panel of microbes todiscriminate between responders and nonrespondersusing ROC curve analysis demonstrated several individ-ual OTUs to be strong predictors of response (Fig-ure S1). A risk index of response to treatment was builtfrom this panel of OTUs. The index, calculated in eachpatient (responders and nonresponders), corresponds tothe difference between the sum of the relative abundanceof all the OTUs associated with nonresponse to treat-ment [i.e. with an LDA (log10) >2] and the sum of therelative abundance of all the OTUs associated withresponse to treatment [i.e. with an LDA (log10) >2]. Themean risk index score was significantly different

Table 1 | Clinical characteristics of all patients

Overall(n = 88)

Treatmentresponder(n = 77)

Treatmentfailure(n = 11) P value*

Treatment successwith no recurrence(n = 55)

Treatment successwith recurrence(n = 22) P value†

Age (median) 52.7 53.8 49.9 0.67 55.6 49.0 0.62Sex (% female) 60.2 59.7 63.6 0.8 58.2 63.6 0.65BMI, mean (kg/m2) 27.5 27.4 28.1 0.75 27.7 26.6 0.44Charlson comorbidity index 1.32 1.32 1.27 0.53 1.34 1.27 0.49Prior antibiotic exposure (%) 59.1 55.8 81.8 0.1 56.4 54.5 0.88Community-acquired (%) 59.1 62.3 36.4 0.13 63.6 59.1 0.9Severe CDI (%) 7 5.2 18.2 0.1 5.5 4.5 0.86Concomitant antibioticexposure (%)

28.4 27.2 36.4 0.5 27.3 27.3 1.0

Concomitant PPIexposure (%)

27.3 25.9 36.3 0.48 18.2 45.5 0.01

Treatment withmetronidazole (%)

70.5 70.1 72.7 0.8 72.7 63.6 0.1

Treatment withvancomycin (%)

23.9 24.7 18.2 0.8 20.0 36.4 0.1

Treatment withvancomycin andmetronidazole

5.6 5.2 9.1 0.8 7.3 0 0.1

CDI, Clostridium difficile infection; BMI, body mass index; PPI, proton pump inhibitor.

* Denotes P value for comparison of treatment responders vs. treatment failures

† Denotes P value for comparison of recurrent infections vs. nonrecurrent infection



S. Khanna et al.

[P < 0.001, one-way ANOVA (analysis of variance) withpost-hoc Tukey HSD (honest significant difference)]between the responders (mean score 0.07 � 0.2) andnon-responders (0.5 � 0.4) as well as healthy subjects(0.03 � 0.1, adult US subjects from global gut study,29

with BMI under 25, n = 60) and non-responders(Figure 2). The risk index was not significantly differentbetween healthy subjects and responders (P = 0.57, one-way ANOVA with post hoc Tukey HSD) suggesting thatthe difference in risk index in responders and nonre-sponders is not due to variability in gut microbiomeamong individuals.

Furthermore, ROC curve analysis showed that thisrisk index was a strong predictor of treatment response,with an area under the curve (AUC) of 0.85. A cut-offof 0.21 was associated with a sensitivity of 77% and aspecificity of 73% (Figure 3a). Importantly, we did notfind a correlation between the risk index and previousantibiotic treatment in terms of response to treatment(Pearson’s product-moment correlation = 0.09,P = 0.37). Thus, nonresponse was not associated withprior antibiotic administration.

In order to assess our risk index, a leave-one-outcross-validation was performed, where the risk index wasbuilt 88 times using n-1 samples each time and then

tested on the held-out sample. Thus, each held-outpatient was treated as a new patient, independently fromthe initial cohort, on whom we tested and subsequentlyrefined the optimal index cut-off to separate respondersand nonresponders. We showed that the risk index wasa strong predictor of response vs. nonresponse

OTU 4476780 Rikenellaceae

OTU 193233 Bacteroides

OTU 175535 BacteroidesOTU 4437814 Bacteroides

OTU 3702906 Ruminococcus gnavusOTU 1654474 LachnospiraceaeOTU 4455767 StreptococcusOTU 526804 StreptococcusOTU 176157 BlautiaOTU 195166 BlautiaOTU 191250 BlautiaOTU 194008 CoprococcusOTU 204681 BlautiaOTU 183925 LachnospiraceaeOTU 190502 LachnospiraceaeOTU 4335781 BifidobacteriumOTU 875735 ActinomycesOTU 190076 LachnospiraceaeOTU 4424239 StreptococcusOTU 294497 ClostridiaceaeOTU 2225203 Clostridiaceae

OTU 359750 Clostridiaceae

OTU 4448928 Clostridium

OTU 4459634 ClostridiumOTU 4445673 Clostridium

OTU 4344207 Streptococcus anginosus

OTU 4439603 Streptococcus



–4 –2 0

Linear discriminant analysis score (log10)

2 4

OTU 315982 ClostridiumperfringensOTU 300501 Clostridium perfringens

OTU 4412788 Clostridium perfringens

OTU 4370657 Clostridium perfringensOTU 4479317 Clostridium perfringens

OTU 185763 Faecalibacterium prausnitzii

Non responders

Responders

OTU 4439469 Ruminococcaceae

Figure 1 | Summary of the operational taxonomic units associated with response or nonresponse to treatment usinglinear discriminant analysis effect size analysis (LEfSe). Representative bacteria with relative abundance of at least 1%with significant differences are represented.

CDI treatmentnon response

–0.5

0.0

0.5

Ris

k in

dex

1.0

CDI treatmentresponse

Healthy subjects

Figure 2 | Clostridium difficile infection (CDI) risk indexbased on the differentiated operational taxonomic unitsin nonresponders, responders and healthy controlscalculated using the Mann–Whitney U test:***P < 0.001. The boxplots denote top quartile, medianand bottom quartile and individual dots representindividual patient samples.




(permutation test performed on the difference betweenthe mean of those without response and the mean ofthose with response to treatment with 999 random per-mutations, P < 0.0001; Figure 3b). We also determinedwith this leave-one-out procedure that a CDI risk indexthreshold of 0.11 best predicts response to treatment,yielding a sensitivity of 70% at a specificity of 73%(mean AUC = 0.81). Thus, we found that our risk index,based on a panel of 36 OTUs that were significantly dif-ferent between primary nonresponders and respondersusing LEfSe, accurately identified patients with CDI,likely to not respond to conventional treatment.

OTU networks are disrupted in primarynonrespondersCorrelations between OTU networks at the genus levelwere computed to differentiate responders and nonre-sponders and demonstrated a threefold (56 vs. 16,ratio = 3.35) decrease in the number of strong taxon–taxon correlations (absolute value of Spearman correla-tion >0.5 and false discovery rate-corrected P < 0.05) innonresponders compared to responders, and most of thedecreased nodes between responders and nonresponderswere members of the phylum Firmicutes (50 vs. 10,ratio = 5) and Actinobacteria (7 vs. 1, ratio = 7) (Fig-ure 4).

Gut microbial diversity is not significantly different inprimary nonresponders compared to respondersUnweighted and weighted UniFrac-based principal coor-dinate analysis (PCoA) did not show significant

differences in beta diversity between responders andnonresponders (unweighted UniFrac distance metric:R = �0.09, P = 0.9; weighted UniFrac distance metric:R = �0.004, P = 0.5) (Figure S2). Patients with CDI (re-sponders and nonresponders) had significantly decreasedalpha diversity (P < 0.001 for the 3 measures phyloge-netic diversity whole tree, observed species and Chao 1index; one-way ANOVA with post hoc Tukey HSD) com-pared to healthy subjects (adult US subjects from globalgut study29). Among patients with CDI, although therewas a trend towards decreased alpha diversity in nonre-sponders compared to responders, this difference wasnot significant for the 3 measures (phylogenetic diversitywhole tree, observed species and Chao 1 index; one-wayANOVA with post hoc Tukey HSD) (Figure S3).

Gut microbiota functional repertoire is significantlydifferent in responders and nonrespondersWe imputed functional aspects of the microbiota from16S rRNA data using PICRUSt to predict CAZY-GHassignments. Based on LEfSe, we found that GH70 (dex-transucrase) and GH38 (a-mannosidase) were increased,whereas GH59 (b-galactosidase) and two carbohydrate-binding modules, CBM16 and CBM42, were significantlydecreased in nonresponders compared to responderswith LDA score (log 10) >2 (Figure S4).

Gut microbiota signatures predict CDI recurrenceAmong the patients who initially responded to treat-ment, 28.5% had recurrent CDI. The median time torecurrence was 23 days (range 15–56 days). There were

0.0 0.2 0.4

Area under the curve = 0.85

0.6 0.8 1.0

0.0

0.2

0.4

Tru

e po

sitiv

e ra

te

False positive rate

0.6

0.8

1.0

0.0 0.2 0.4

Mean area under thecurve = 0.81

0.6 0.8 1.0

0.0

0.2

0.4

Tru

e po

sitiv

e ra

te

False positive rate

0.6

0.8

1.0

(a) (b)

Figure 3 | (a) Receiver operating characteristic (ROC) curve analysis of the CDI risk index in pre-treatment faecalsamples collected following 10-fold cross-validation. (b) ROC curve analysis of the CDI risk index in pre-treatmentfaecal samples collected following 10-fold cross-validation of all the held-out samples from the leave-one-out cross-validation analysis. Following the leave-one-out procedure, the ROC curves of each leave-one-out procedure are inblue and the mean ROC curve of all the procedure is in black.



S. Khanna et al.

no significant differences in patients with and withoutrecurrent CDI (Table 1) among the clinical variablesanalysed except PPI use which predicted recurrent CDIon univariate analysis (odds ratio 3.75, 95% confidenceinterval 1.27–11.1, P = 0.01) and multivariable analysisafter controlling for age (P = 0.0007) and comorbidities(P = 0.0009) in separate multivariable models. However,PPI use was not associated with alterations in the gutmicrobiota (Table S2).

Relative abundance of 11 OTUs was significantly dif-ferent between those with and without recurrence.Patients with recurrence had a significant increase inVeillonella, Enterobacteriaceae (Erwinia), Streptococcus,Parabacteroides and Lachnospiraceae using LEfSe (Fig-ure 5). Several individual microbes were strong predic-tors of recurrence (Figure S5).

A risk index of recurrence built from this panel ofmicrobes differentiated between patients with and with-out recurrence. This index included all OTUs that had aLDA score (log10) >2, as previously described. This indexwas significantly different in patients who did not have arecurrence (mean score 0.09 � 0.08) and those who did(0.19 � 0.12) (Mann–Whitney U test, P = 0.0001)

(Figure 6a). The ROC curve analysis showed that thisrisk index was a strong predictor of recurrence, with anAUC of 0.78 (Figure 6b). Moreover, a cut-off of 0.13had a sensitivity of 78% and a specificity of 68%. Asdescribed previously, a leave-one-out cross-validationwas performed, where the risk index was built 77 timesusing n-1 samples each time and then tested on theheld-out sample. This procedure showed that the gutmicrobiota was a strong predictor of recurrence (permu-tation test performed on the difference between themean of those with recurrence and the mean of thosewithout recurrence with 999 random permutations,P < 0.0001; Figure 6c). This procedure also determinedthat a CDI risk index classification threshold of 0.11 bestpredicts recurrence with a sensitivity of 75% at a speci-ficity of 69% (mean AUC = 0.75). Thus, we found thatour risk index can identify patients with CDI who arelikely to have recurrence after initial response. Impor-tantly, a supervised learning method using a RandomForest model was also applied but was unable to accu-rately assign samples to their source population based ontaxonomic profiles at the OTU level and was outper-formed by the above risk index approach.

Enterococcaceae

Bifidobacterium

Coprococcus

Roseburia

[Ruminococcus]Klebsiella

Pantoea

Enterobacter

Xenorhabdus

Clostridiaceae

Pasteurellaceae

Haemophilus

Enterococcaceae

Enterococcus

Clostridiales

Dorea

Lachnospiraceae

Blautia

Enterobacteriaceae

Escherichia

Clostridium

Clostridiaceae

Peptostreptococcaceae

Turicibacter

Ruminococcaceae

Faecalibacterium Enysipelotrichaceae

Eggerthella

SMB53

GranulicatellaActinomyces

Pasteurellaceae

Haemophilus

Streptococcus

Streptococcus

Akkermansia

FusobacteriumShigella Coprobacillus Bacteroides

Escherichia Lactococcus Actinomyces

Rothia Gemellaceae

Enterococcus

Pantoea

Klebsiella

Enterobacter

Citrobacter

Lactobacillales

Lactobacillales

Erwinia

TrabulsiellaXenorhabdus

Aquamonas

Shigella

Proteobacteria

FusobacteriaVerrucomicrobiaActinobacteria

FirmicutesBacteroidetes

(a) (b)

Figure 4 | Spearman correlation demonstrating taxon–taxon relative abundance including links with absolute value ofcorrelation >0.5 and false discovery rate-corrected P < 0.05. Network analyses displayed with Cytoscape using anedge-weighted spring embedded layout in responders (a) and nonresponders (b). Positive correlations are shown asblue links between nodes and only significant correlations are depicted. The size of the node was based on thenumber of correlations associated with the corresponding taxon. The colour of the node is based on the phylum level:Actinobacteria (blue), Bacteroidetes (green), Firmicutes (purple), Proteobacteria (yellow), Fusobacteria (pink) andVerrucomicrobia (grey). There is a threefold decrease in the number of taxon correlations in the nonresponders andmost of the decreased nodes were Firmicutes and Actinobacteria.




Pre-treatment Gut microbial diversity is notsignificantly different in patients with and withoutrecurrent CDIUnweighted and weighted UniFrac-based PCoA did notshow significant differences in beta diversity betweenpatients with and without recurrence (unweighted Uni-Frac distance metric: R = �0.1, P = 0.9; weighted Uni-Frac distance metric: R = 0.02, P = 0.3) (Figure S6).Moreover, there was no significant difference in alphadiversity based on three different metrics between CDIpatients with and without recurrence (Figure S7).

Gut microbiota functional repertoire is significantlydifferent in those with and without recurrenceAs described above, we again imputed functional aspectsof the microbiota from 16S rRNA data using PICRUSt

to predict CAZY-GH assignments. Based on the LEfSetool, GT30 (b-fucosidase) and a carbohydrate-bindingmodule (CBM20) were increased in patients who had arecurrence compared to patients who did not with LDAscore (log10) >2 (Figure S8).

DISCUSSIONIn this study, we report specific gut microbiota signa-tures associated with the initial response to treatmentand recurrence after successful treatment in patients withprimary CDI. We have developed a risk index based oncompositional differences among patients with CDI,which can help predict response to treatment and recur-rence in patients with CDI. This is a step towards identi-fying CDI patients who would be candidates for earlyalternative therapies such as faecal microbiota

–4 –2

RecurrenceNo recurrence OTU 4458959 Veillonella

OTU 4310208 Veillonella dispar

OTU 3801267 Veillonella parvula

OTU 191251 Parabacteroides

OTU 533969 Lachnospiraceae

OTU 577294 Parabacteroides distasonis

OTU 1010113 Erwinia


OTU 2497335 Parabacteroides distasonis

OTU 4457268 Enterobacteriaceae

OTU 197286 Enterobacteriaceae

0

Linear discriminant analysis score (log10)

2 4

Figure 5 | Summary of theoperational taxonomic unitsassociated with recurrence(n = 22) vs. nonrecurrence(n = 55) after successfultreatment response usinglinear discriminant analysiseffect size analysis (LEfSe).

0.0

No recurrence Recurrence

0.1

0.2

Ris

k in

dex

0.3

0.4

0.5

0.0

0.2

0.4

Tru

e po

sitiv

e ra

te

0.6

0.8

1.0

0.0 0.2 0.4

False positive rate

Area under the curve = 0.78Mean area under thecurve = 0.75

0.6 0.8 1.0 0.0 0.2 0.4

False positive rate

0.6 0.8 1.0

0.0

0.2

0.4

Tru

e po

sitiv

e ra

te

0.6

0.8

1.0

(a) (b) (c)

Figure 6 | (a) Clostridium difficile infection (CDI) risk index based on the differentiated operational taxonomic units inpatients with success vs. recurrence. Mann–Whitney U test: ***P < 0.001. Boxplots denote top quartile, median andbottom quartile. (b) Receiver operating characteristic (ROC) curve analysis of the CDI risk index in faecal samplescollected prior to treatment predicting recurrence. (c) ROC curve analysis of the CDI risk index in faecal samplescollected prior to treatment. Following the leave-one-out procedure, the ROC curves of each leave-one-out procedureare in blue and the mean ROC curve of all the procedure is in black.



S. Khanna et al.

transplantation (FMT). Although a significant percentageof patients either fail primary treatment or have recur-rence after successful treatment, there are no robustmodels to predict these outcomes. Gut microbiota playan important role in resistance to colonisation by patho-gens such as C. difficile, and this resistance can bealtered following disruption of normal gut microbialcommunity structure following antibiotic use or otheralterations in host physiology. Several mechanisms thatpromote C. difficile colonisation have been attributed toalteration of gut microbiota including increase in pri-mary bile acids, increased availability of nutrients suchas succinate and sialic acid, and decrease in butyrate pro-ducers in the gut.35, 36 Given the impact of gut micro-biota in the pathogenesis of CDI, we profiled themicrobial community in pre-treatment stool samples andidentified biologically relevant microbial signatures thatpredict treatment response and risk of recurrent CDI.

Data on clinical predictors of response to treatment inCDI are sparse. Clinical predictors of metronidazole fail-ure in CDI include severe CDI, recent cephalosporin use,CDI on hospital admission and intra-hospital transfer. Inrecurrent CDI, studies showing increasing age, concomi-tant antibiotic exposure and proton pump inhibitor useas risk factors are balanced by other studies with a lackof correlation between these variables and recurrentCDI.24, 25, 37 In our study, PPI use was the only clinicalvariable significantly associated with recurrent CDI.However, most patients had community-acquired CDIand were younger than patients with hospital-acquiredCDI, which likely explains the lack of significant differ-ences in some previously described clinical predictors ofrecurrent CDI such as age and comorbidities.

In our cohort, there was a significant decrease in overallmicrobial diversity in patients with CDI as compared tohealthy individuals, but among patients with CDI, therewas no significant decrease in the microbial diversityamong those who failed primary therapy or had recur-rence after initial success compared to CDI patients whodid not have these adverse outcomes. We found anincrease in taxa within Faecalibacterium, Ruminococ-caceae, Bacteroides and Rikenellaceae, among others inprimary responders. Interestingly, these taxa have beenassociated with colonisation resistance against C. difficilein adult patients.27 Faecalibacterium prausnitzii is largelyconsidered a beneficial bacterium as significant reductionsin this taxa have been reported in patients with inflamma-tory bowel disease compared to healthy individuals.38

F. prausnitzii and members of Ruminococcaceae are cap-able of producing butyrate, similar to C. difficile.39, 40

Several studies indicate that CDI is accompanied by adepletion of butyrogenic bacteria,9 suggesting either aninhibitory effect of butyrate or the loss of a nutrient niche,which can be occupied by C. difficile. Thus, the presenceof F. prausnitzii and Ruminococcaceae may improve theresponse to treatment by providing niche competition ordirect inhibition. An increase in Bacteriodetes representedby greater abundance of the families Bacteroidaceae,Rikenellaceae and Porphyromonadaceae has been reportedfollowing successful FMT for treatment of CDI.41 Interest-ingly a higher relative abundance of Bacteroides and Rick-nellaceae was seen in primary responders in our study,suggesting they may synergise with conventional treat-ment for exclusion of C. difficile. While the mechanism ofaction remains unclear, recent studies have shown thatcertain Bacteroides spp. may play a role in preventinginfection with C. difficile.42, 43 B. fragilis influences thedevelopment of the immune response and B. thetaio-taomicron stimulates Paneth cells to produce antibacterialpeptides, which may prevent pathogens from colonis-ing.44, 45 While indirect, these data suggest a synergisticrole for members of the gut microbial community in aug-menting treatment response.

Interestingly, even though overall gut microbial diver-sity is not significantly different in responders comparedto nonresponders, there is a threefold increase in the num-ber of strong taxon–taxon correlations in responders ascompared to nonresponders, and most of the decreasednodes in nonresponders were members of the phylumFirmicutes. This suggests that in addition to differences inindividual taxa, the overall microbial community structureplays an important role in augmenting response toprimary therapy by more effectively excluding C. difficile.

In patients with recurrent CDI we found increases intaxa that were different than in primary nonresponders.Several of these have been previously associated withrecurrent CDI or potentially promote the growth of C. dif-ficile. There was an increase in Enterobacteriaceae, Veil-lonella and Parabacteroides among others in patients withrecurrent CDI compared to patients without recurrence.Several studies both on mice and on humans demonstratean increase of Enterobacteriaceae in CDI.7, 27, 46 A persis-tent expansion of Enterobacteriaceae has also been shownafter treatment with clindamycin suggesting it may play arole in susceptibility to colonisation with C. difficile.Parabacteroides diastonis, a succinate producer, has beenpreviously associated with CDI47, 48 perhaps sinceincreased succinate availability facilitates expansion ofC. difficile. The low concentration of succinate present inthe microbiota of conventional mice is transiently elevated




upon antibiotic treatment or chemically induced intestinalmotility disturbance, and C. difficile exploits this succinatespike to expand in the perturbed intestine.35 Further stud-ies are needed to identify mechanisms by which these bac-teria may contribute to failure of treatment or recurrenceafter successful treatment.

The primary strength of our study is the ability topredict the clinically important outcomes of treatmentfailure and recurrence in patients with primary CDI bycharacterising the gut microbiota of pre-treatment stoolsamples. It is possible that strain level differences inC. difficile contribute to the outcome but these were notassessed in this study. This study helps define a newdiagnostic paradigm, but as with any initial finding, thereare limitations. The risk index developed in our studylikely needs to be prospectively validated in a largercohort of CDI patients given the relatively small samplesize. We did not have information on the patients’ diet-ary histories at the time of diagnosis so we were unableto assess how diet may influence microbial compositionand CDI outcomes. Additionally, detailed analyses com-paring microbiome changes due to prior antibiotic expo-sure would need to be considered in future studies. Wealso used healthy controls from a previously publishedstudy and there could be a study effect given the sampleswere run at different times as has been pointed out inprevious studies. This is usually attributed to differencesin DNA extraction protocol, primers and data analysis.49

However, as these were not different between our studyand the healthy controls, we do not expect a study effect.Prior studies, which have investigated study effects havelooked at healthy controls from different studies, wherestudy-specific effects can be seen. However, as we areinvestigating differences in healthy controls and patientswith CDI, we expect the true biological effect to over-come inter-individual study differences. While this wasnot a primary aim of our study we do note that ourfindings are similar to those previously reported.7 Boththese studies used Illumina platform for sequencing, weused Miseq and the healthy controls were sequenced onHiseq. A cross-validation has been previously publishedshowing that the data from these two platforms do notintroduce a bias in the results.30 The risk index deter-mined in our study and applied to the healthy controlswas similar in responders and healthy controls (from dif-ferent studies) and different from nonresponders, furthersupporting the validity of our finding.

In conclusion, gut microbiota signatures can be used topredict response to and recurrence after initial treatment inpatients with CDI. This finding will need to be validated in

a larger cohort and future work will focus on understand-ing interaction of individual taxa with C. difficile to under-stand mechanisms by which they may determine outcomeof treatment. Nevertheless, these biomarkers potentiallyallow identification of subsets patients that may be initiallytreated with more effective therapies such as newer antibi-otics, faecal microbiota transplantation or defined micro-bial consortia instead of a prolonged therapeutic trial.

SUPPORTING INFORMATIONAdditional Supporting Information may be found in theonline version of this article:Figure S1. Receiver operating characteristic (ROC)

curve analysis of the most distinctive operational taxo-nomic units in fecal samples collected prior to treatmentfollowing 10-fold jack-knifing. The 10 ROC curves are inblue and the mean ROC curve is depicted in black.Figure S2. Difference in beta-diversity among respon-

ders and non-responders. Analyses performed on 16SrRNA V4 region data, with a rarefaction depth of 6,987reads per sample. (A) Principal Coordinate Analysis(PCoA) of unweighted UniFrac distances. Proportion ofvariance explained by each principal coordinate axis isdenoted in the corresponding axis label. (B) PrincipalCoordinate Analysis (PCoA) of weighted UniFrac dis-tances. Proportion of variance explained by each princi-pal coordinate axis is denoted in the corresponding axislabel. There are no significant differences in beta diver-sity among responders and non-responders samples (un-weighted UniFrac distance metric: R = -0.09, p = 0.9;weighted UniFrac distance metric: R = -0.004, p = 0.5).Figure S3. Difference in alpha-diversity among

responders and non-responders using both phylogenetic(Faith’s phylogenetic diversity) and non-phylogenetic(observed species, Shannon index, chao 1 index) richnessmetrics. Analyses were performed on 16S rRNA V4region data, with a rarefaction depth of 6,987 reads persample. Whiskers in the boxplot represent the range ofminimum and maximum alpha diversity values within apopulation, excluding outliers.Figure S4. Differences in glycoside hydrolases: Signifi-

cantly different Carbohydrate-Active Enzymes database -glycoside hydrolases (CAZY GH) in samples collectedprior to treatment between responders (n = 77) andnon-responders (n = 11) using Linear discriminant anal-ysis Effect Size analysis (LEfSe). Boxplots denote topquartile, median and bottom quartile.Figure S5. Receiver operating characteristic analysis of

the most distinctive operational taxonomic units in fecalsamples collected prior to treatment following 10-fold



S. Khanna et al.

jack-knifing. The 10 ROC curves are in blue and themean ROC curve is in black.Figure S6. Difference in beta-diversity among patients

with and without recurrence after successful treatment.Analyses performed on 16S rRNA V4 region data, witha rarefaction depth of 6,987 reads per sample. (A) Prin-cipal Coordinate Analysis (PCoA) of unweighted Uni-Frac distances. Proportion of variance explained by eachprincipal coordinate axis is denoted in the correspondingaxis label. (B) Principal Coordinate Analysis (PCoA) ofweighted UniFrac distances. Proportion of varianceexplained by each principal coordinate axis is denoted inthe corresponding axis label. There are no significant dif-ferences between those with and without recurrence (un-weighted UniFrac distance metric: R = -0.1, p = 0.9;weighted UniFrac distance metric: R = 0.02, p = 0.3).Figure S7. Difference in alpha-diversity among

patients with and without recurrence after successfultreatment. Analyses were performed on 16S rRNA V4region data, with a rarefaction depth of 6,987 reads persample. Whiskers in the boxplot represent the range ofminimum and maximum alpha diversity values within apopulation, excluding outliers. There was no significantdifference in alpha diversity based on four different met-rics between those with and without recurrence.Figure S8. Differences in glycoside hydrolases: Signifi-

cantly different GT30 (b-fucosidase) and a carbohydrate-binding module (CBM20) in patients who had a recur-rence compared to patients who did not with LDA score(log10) >2 using Linear discriminant analysis Effect Sizeanalysis (LEfSe). Boxplots denote top quartile, medianand bottom quartile.Table S1. Prior antibiotic exposure in patients with

primary Clostridium difficile infection.Table S2. Correlations between clinical characteristics

and bacterial taxaData S1. Data acquisition and analyses.Data S2. R Code to generate risk index.

AUTHORSHIPGuarantor of the article: Purna C. Kashyap.Author contributions: S.K. was involved in study concept, dataacquisition, data analysis, data interpretation, drafting and revisionof manuscript; E.M. was involved in data acquisition, data analysis,data interpretation, drafting and revision of manuscript; B.S. wasinvolved in data acquisition, data analysis and data interpretation;R.P. was involved in data analysis, data interpretation, drafting andrevision of manuscript; D.K. was involved in study concept, dataanalysis, data interpretation, drafting and revision of manuscriptD.S.P. was involved in study concept, data analysis, data interpreta-tion, drafting and revision of manuscript; P.C.K. was involved instudy concept, data acquisition, data analysis, data interpretation,drafting and revision of manuscript.

All authors have reviewed and approved the final draft submitted.

ACKNOWLEDGEMENTSDeclaration of personal interests: Sahil Khanna has received researchgrant support from Merck Pharmaceuticals and has consulted withRebiotix, Inc (consultation paid to Mayo Clinic) and Summit phar-maceuticals. Dr. Pardi has received funding from Seres therapeutics,Salix pharmaceuticals, Atlantic pharmaceuticals, Takeda pharmaceu-ticals, Merck pharmaceuticals, and Janssen pharmaceuticals. Dr.Pardi has also served as a Consultant to Merck pharmaceuticals andOtsuka pharmaceuticals. Dr. Patel reports grants from BioFire,Check-Points, Curetis, 3M, Merck, Hutchison Biofilm Medical Solu-tions, Accelerate Diagnostics, Allergan, and The Medicines Com-pany. Dr. Patel is a consultant to Curetis, Roche, Qvella, andDiaxonhit. In addition, Dr. Patel has a patent on Bordetella pertus-sis/parapertussis PCR with royalties paid by TIB, a patent on adevice/method for sonication with royalties paid by Samsung toMayo Clinic, and a patent on an anti-biofilm substance issued. Dr.Patel serves on an Actelion data monitoring board. Dr. Patelreceives travel reimbursement and an editor’s stipend from ASMand IDSA, and honoraria from the USMLE, Up-to-Date and theInfectious Diseases Board Review Course.Declaration of funding interests: This research was made possible bysupport from NIH K08 DK100638 (PCK), Global Probiotic Council(PCK), Minnesota Partnership for Biotechnology and Genomics(PCK) and Center for Individualized Medicine, Mayo Clinic, Roche-ster and CTSA Grant Number UL1 TR000135 from the NationalCenter for Advancing Translational Sciences (NCATS), a componentof the National Institutes of Health (NIH) and the National Instituteof Diabetes and Digestive and Kidney Diseases of the National Insti-tutes of Health under Award Number P30 DK084567 (ClinicalCore). Its contents are solely the responsibility of the authors anddo not necessarily represent the official view of NIH. EM was sup-ported by the Robert Tournut award from the French National Soci-ety of Gastroenterology.

REFERENCES1. Kelly CP, LaMont JT. Clostridium

difficile–more difficult than ever. N EnglJ Med 2008; 359: 1932–40.

2. Magill SS, Edwards JR, Bamberg W,et al.Multistate point-prevalence surveyof health care-associated infections. NEngl J Med 2014; 370: 1198–208.

3. Lessa FC, Mu Y, Bamberg WM, et al.Burden of Clostridium difficile infection

in the United States. N Engl J Med2015; 372: 825–34.

4. Nanwa N, Kendzerska T, Krahn M,et al. The economic impact ofClostridium difficile infection: asystematic review. Am J Gastroenterol2015; 110: 511–9.

5. Buffie CG, Bucci V, Stein RR, et al.Precision microbiome reconstitution

restores bile acid mediated resistance toClostridium difficile. Nature 2015; 517:205–8.

6. Manges AR, Labbe A, Loo VG, et al.Comparative metagenomic study ofalterations to the intestinal microbiotaand risk of nosocomial Clostridumdifficile-associated disease. J Infect Dis2010; 202: 1877–84.




7. Schubert AM, Rogers MA, Ring C,et al. Microbiome data distinguishpatients with Clostridium difficileinfection and non-C. difficile-associateddiarrhea from healthy controls. mBio2014; 5: e01021–14.

8. Rousseau C, Poilane I, De Pontual L,Maherault AC, Le Monnier A, CollignonA. Clostridium difficile carriage inhealthy infants in the community: apotential reservoir for pathogenic strains.Clin Infect Dis 2012; 55: 1209–15.

9. Antharam VC, Li EC, Ishmael A, et al.Intestinal dysbiosis and depletion ofbutyrogenic bacteria in Clostridiumdifficile infection and nosocomialdiarrhea. J Clin Microbiol 2013; 51:2884–92.

10. Khanna S, Pardi DS. Clostridium difficileinfection: new insights into management.Mayo Clin Proc 2012; 87: 1106–17.

11. Leffler DA, Lamont JT. Clostridiumdifficile infection. N Engl J Med 2015;373: 287–8.

12. Press A, Ku B, McCullagh L, Rosen L,Richardson S, McGinn T. Developing aclinical prediction rule for firsthospital-onset Clostridium difficileinfections. A retrospective observationalstudy. Infect Control Hosp Epidemiol2016; 37: 896–900.

13. Settle CD, Wilcox MH, Fawley WN,Corrado OJ, Hawkey PM. Prospectivestudy of the risk of Clostridium difficilediarrhoea in elderly patients followingtreatment with cefotaxime orpiperacillin-tazobactam. AlimentPharmacol Ther 1998; 12: 1217–23.

14. Turco R, Martinelli M, Miele E, et al.Proton pump inhibitors as a risk factorfor paediatric Clostridium difficileinfection. Aliment Pharmacol Ther2010; 31: 754–9.

15. Dalton BR, Lye-Maccannell T,Henderson EA, Maccannell DR, LouieTJ. Proton pump inhibitors increasesignificantly the risk of Clostridiumdifficile infection in a low-endemicity,non-outbreak hospital setting. AlimentPharmacol Ther 2009; 29: 626–34.

16. Allegretti JR, Kearney S, Li N, et al.Recurrent Clostridium difficile infectionassociates with distinct bile acid andmicrobiome profiles. AlimentPharmacol Ther 2016; 43: 1142–53.

17. Leffler DA, Lamont JT. Clostridiumdifficile infection. N Engl J Med 2015;372: 1539–48.

18. Khanna S, Pardi DS. Clostridiumdifficile infection: managementstrategies for a difficult disease. TherAdv Gastroenterol 2014; 7: 72–86.

19. Cammarota G, Masucci L, Ianiro G,et al. Randomised clinical trial: faecalmicrobiota transplantation bycolonoscopy vs. vancomycin for thetreatment of recurrent Clostridium

difficile infection. Aliment PharmacolTher 2015; 41: 835–43.

20. Guo B, Harstall C, Louie T, Veldhuyzenvan Zanten S, Dieleman LA. Systematicreview: faecal transplantation for thetreatment of Clostridium difficile-associated disease. Aliment PharmacolTher 2012; 35: 865–75.

21. Louie TJ, Miller MA, Mullane KM,et al. Fidaxomicin versus vancomycinfor Clostridium difficile infection. NEngl J Med 2011; 364: 422–31.

22. Zar FA, Bakkanagari SR, Moorthi KM,Davis MB. A comparison of vancomycinand metronidazole for the treatment ofClostridium difficile-associated diarrhea,stratified by disease severity. Clin InfectDis 2007; 45: 302–7.

23. Johnson S, Louie TJ, Gerding DN,et al. Vancomycin, metronidazole, ortolevamer for Clostridium difficileinfection: results from twomultinational, randomized, controlledtrials. Clin Infect Dis 2014; 59:345–54.

24. Hu MY, Katchar K, Kyne L, et al.Prospective derivation and validation ofa clinical prediction rule for recurrentClostridium difficile infection.Gastroenterology 2009; 136: 1206–14.

25. Shivashankar R, Khanna S, KammerPP, et al. Clinical predictors ofrecurrent Clostridium difficile infectionin out-patients. Aliment PharmacolTher 2014; 40: 518–22.

26. Eyre DW, Walker AS, Wyllie D, et al.Predictors of first recurrence ofClostridium difficile infection:implications for initial management. ClinInfect Dis 2012; 55(Suppl. 2): S77–87.

27. Seekatz AM, Young VB. Clostridiumdifficile and the microbiota. J ClinInvest 2014; 124: 4182–9.

28. Caporaso JG, Kuczynski J, StombaughJ, et al. QIIME allows analysis ofhigh-throughput communitysequencing data. Nat Methods 2010; 7:335–6.

29. Yatsunenko T, Rey FE, Manary MJ,et al. Human gut microbiome viewedacross age and geography. Nature 2012;486: 222–7.

30. Caporaso JG, Lauber CL, Walters WA,et al. Ultra-high-throughput microbialcommunity analysis on the IlluminaHiSeq and MiSeq platforms. ISME J2012; 6: 1621–4.

31. Morgan XC, Tickle TL, Sokol H, et al.Dysfunction of the intestinalmicrobiome in inflammatory boweldisease and treatment. Genome Biol2012; 13: R79.

32. Segata N, Izard J, Waldron L, et al.Metagenomic biomarker discovery andexplanation. Genome Biol 2011; 12: R60.

33. Montassier E, Al-Ghalith GA, Ward T,et al. Pretreatment gut microbiome

predicts chemotherapy-relatedbloodstream infection. Genome Med2016; 8: 49.

34. Smoot ME, Ono K, Ruscheinski J,Wang PL, Ideker T. Cytoscape 2.8: newfeatures for data integration andnetwork visualization. Bioinformatics2011; 27: 431–2.

35. Ferreyra JA, Wu KJ, Hryckowian AJ,Bouley DM, Weimer BC, SonnenburgJL. Gut microbiota-produced succinatepromotes C. difficile infection afterantibiotic treatment or motilitydisturbance. Cell Host Microbe 2014;16: 770–7.

36. Ng KM, Ferreyra JA, Higginbottom SK,et al. Microbiota-liberated host sugarsfacilitate post-antibiotic expansion ofenteric pathogens. Nature 2013; 502:96–9.

37. McDonald EG, Milligan J, Frenette C,Lee TC. Continuous proton pumpinhibitor therapy and the associatedrisk of recurrent Clostridium difficileinfection. JAMA Intern Med 2015; 175:784–91.

38. Cao Y, Shen J, Ran ZH. Associationbetween Faecalibacterium prausnitziireduction and inflammatory boweldisease: a meta-analysis and systematicreview of the literature. GastroenterolRes Pract 2014; 2014: 872725.

39. Pryde SE, Duncan SH, Hold GL, StewartCS, Flint HJ. The microbiology ofbutyrate formation in the human colon.FEMS Microbiol Lett 2002; 217: 133–9.

40. Rios-Covian D, Gueimonde M, DuncanSH, Flint HJ, de Los Reyes-Gavilan CG.Enhanced butyrate formation by cross-feeding between Faecalibacteriumprausnitzii and Bifidobacteriumadolescentis. FEMS Microbiol Lett 2015;362: doi: 10.1093/femsle/fnv176.

41. Hamilton MJ, Weingarden AR, UnnoT, Khoruts A, Sadowsky MJ. High-throughput DNA sequence analysisreveals stable engraftment of gutmicrobiota following transplantation ofpreviously frozen fecal bacteria. GutMicrobes 2013; 4: 125–35.

42. Hopkins MJ, Macfarlane GT. Changesin predominant bacterial populations inhuman faeces with age and withClostridium difficile infection. J MedMicrobiol 2002; 51: 448–54.

43. Hopkins MJ, Macfarlane GT.Nondigestible oligosaccharides enhancebacterial colonization resistance againstClostridium difficile in vitro. ApplEnviron Microbiol 2003; 69: 1920–7.

44. Hooper LV, Stappenbeck TS, Hong CV,Gordon JI. Angiogenins: a new class ofmicrobicidal proteins involved in innateimmunity. Nat Immunol 2003; 4: 269–73.

45. Mazmanian SK, Kasper DL. Thelove-hate relationship between bacterialpolysaccharides and the host immune



S. Khanna et al.

http://dx.doi.org/10.1093/femsle/fnv176

system. Nat Rev Immunol 2006; 6:849–58.

46. Rea MC, O’Sullivan O, Shanahan F,et al. Clostridium difficile carriagein elderly subjects and associatedchanges in the intestinalmicrobiota. J Clin Microbiol 2012; 50:867–75.

47. Chang JY, Antonopoulos DA, Kalra A,et al. Decreased diversity of the fecalmicrobiome in recurrent Clostridiumdifficile-associated diarrhea. J Infect Dis2008; 197: 435–8.

48. Lawley TD, Clare S, Walker AW, et al.Targeted restoration of the intestinalmicrobiota with a simple, defined

bacteriotherapy resolves relapsingClostridium difficile disease in mice.PLoS Pathog 2012; 8: e1002995.

49. Lozupone CA, Stombaugh J, GonzalezA, et al. Meta-analyses of studies of thehuman microbiota. Genome Res 2013;23: 1704–14.




gut microbiome predictors of treatment response and...

Documents