Multi-omics investigation of Clostridioides difficile-colonized patients reveals protective commensal carbohydrate metabolism

Clostridioides difficile infection (CDI) imposes a substantial burden on the health care system in the United States. Understanding the biological basis for the spectrum of C. difficile-related disease manifestations is imperative to improving treatment and prevention of CDI. Here, we investigate the correlates of asymptomatic C. difficile colonization using a multi-omics approach, comparing the fecal microbiome and metabolome profiles of patients with CDI versus asymptomatically-colonized patients. We find that microbiomes of asymptomatic patients are significantly enriched for species in the class Clostridia relative to those of symptomatic patients. Asymptomatic patient microbiomes were enriched with fucose, rhamnose, and sucrose degradation pathways relative to CDI patient microbiomes. Fecal metabolomics corroborates this result: we identify carbohydrate compounds enriched in asymptomatic patients relative to CDI patients, and correlated with a number of commensal Clostridia. Further, we reveal that across C. difficile isolates, the carbohydrates rhamnose and lactulose do not serve as robust growth substrates in vitro, corroborating their enriched detection in our metagenomic and metabolite profiling of asymptomatic individuals. We conclude that in asymptomatically-colonized individuals, carbohydrate metabolism by other commensal Clostridia may prevent CDI by inhibiting C. difficile proliferation. These insights into C. difficile colonization and putative commensal competition suggest novel avenues to develop probiotic or prebiotic therapeutics against CDI.


62
Clostridioides difficile infection (CDI) remains a significant cause of morbidity and mortality in the 63 health care setting and in the community [1]. Antibiotic treatments, among other risk factors associated 64 with weakened colonization resistance, increase susceptibility to CDI [2,3]. C. difficile residence in the 65 human gastrointestinal (GI) tract may result in a spectrum of disease, from asymptomatic colonization to 66 severe and sometimes fatal manifestations of CDI [4]. Diagnosis of CDI relies on detection of the protein 67 toxin, most commonly by enzyme immunoassay (EIA), or the detection of the toxin-encoding genes, by 68 nucleic acid amplification test (NAAT). These diagnostic tools serve as rough benchmarks for assessing 69 severity of disease. Discrepancies between the results of these assays, as in the case of patients with 70 clinically significant diarrhea (CSD) who are EIA negative (EIA-) but NAAT positive for toxigenic C. difficile 71 (Cx+), highlight the complexity of states in which C. difficile can exist in the GI tract. Clarifying the 72 biological differences between asymptomatic colonization (Cx+/EIA-) and CDI (Cx+/EIA+) will be critical 73 for identifying mechanisms of colonization resistance, and for defining novel probiotic or prebiotic 74 avenues for treatment or prevention of CDI [5,6].

75
C. difficile enters the GI tract as a spore, germinates in the presence of primary bile acids, and 76 replicates through consumption of amino acids and other microbiota/host-derived nutrients [7]. It is not a 77 coincidence that many of these metabolic cues are characteristic of a perturbed microbiome [8,9]. The 78 hallmark of C. difficile pathogenesis is the expression of the toxin locus encoded on the tcd operon; this 79 locus is tightly regulated by nutrient levels [10]. Correspondingly, it is hypothesized that an environment 80 replete of nutrients induces toxinogenesis, allowing C. difficile to restructure the gut environment and 81 acquire nutrients through inflammation [11,12]. Patients who are colonized but have no detectable C. 82 difficile toxin in their stool suggests that these patients' microbiomes may be less permissive towards CDI 83 development. Identification of metabolic traits within the microbiome of asymptomatic, C. difficile-colonized patients could reveal a number of potential therapeutic pathways towards precise amelioration 85 of symptomatic C. difficile disease.

86
A multitude of probiotic and prebiotic approaches have demonstrated efficacy to curb C. difficile 87 proliferation in vivo [6,13,14]. While restoration of the microbiota through fecal microbiota transplantation 88 can provide colonization resistance [15], the molecular mechanisms of how this resistance is conferred 89 remain unclear. Recent studies using a murine model of infection have indicated that the administration of 90 carbohydrates (both complex and simple) in the diet can be used to curb or prevent CDI [16][17][18].

91
Paradoxically, integrated metabolomics and transcriptomics data collected during murine C. difficile 92 colonization indicates that simple carbohydrates are imperative for pathogen replication [11]. It is critical 93 to understand the mechanism by which catabolism of specific carbohydrates could inhibit C. difficile 94 proliferation in the human GI tract.

95
Here, we perform a multi-level investigation of two relevant patient populations, those colonized 96 with C. difficile but EIA negative (asymptomatically colonized) and those who are EIA positive (CDI) to 97 understand the microbial and metabolic features that may underlie protection from CDI. First, we use 98 microbiome analyses to identify a number of non-C. difficile, clostridial species that are negatively 99 correlated with C. difficile in asymptomatically-colonized individuals. Secondly, interrogation of a 100 metabolomics dataset from the same patient population [19] reveals increased abundance of a number of 101 carbohydrate metabolites in asymptomatic patients. Finally, we show that these metabolites enriched in 102 asymptomatically-colonized individuals are largely non-utilizable by C. difficile isolates. Together, these 103 datasets reveal that asymptomatically-colonized patients are defined by an interaction of clostridial 104 species and carbohydrate metabolites that may serve as a last-line of resistance against CDI in colonized 105 patients.

108
The clinical outcome of C. difficile colonization of a host is heavily influenced by taxonomic and 109 metabolic constituents of the microbiome. While colonization in many cases can lead to CDI, evidence of 110 cases of colonization without CDI implies some level of protection from outright disease. We sought to 111 identify biological features that distinguish asymptomatically-colonized patients from those with CDI [20].
In a retrospective human cohort of 102 patients with clinically significant diarrhea (CSD), two groups of 113 patients were identified: those diagnosed with CDI (Cx+/EIA+) or those asymptomatically -colonized 114 (Cx+/EIA-), as defined previously [19]. Because antibiotics are a well-known risk factor for CDI, we 115 analyzed previous antibiotic orders (within one month prior to diagnosis) for patients in the Cx+/EIA-and 116 Cx+/EIA+ cohort, as a proxy for antibiotic exposure (Supplementary Table 1

121
Antibiotics increase susceptibility to CDI through disruption of colonization resistance, mainly 122 conferred to the host via the gut microbiome [23]. We hypothesized that our asymptomatic patients would 123 have increased microbiome-mediated colonization resistance relative to CDI patients. To determine the 124 microbial correlates of disease state, we performed shotgun metagenomic sequencing on patient stool 125 samples from the asymptomatic (n=54) and CDI (n=48) groups. We examined community structure in 126 stool metagenomes and found that there was no significant difference (Wilcoxon rank-sum, P=0.78). in 127 alpha-diversity (Shannon diversity) between patient groups ( Figure 1A). In addition, we interrogated beta-128 diversity (Bray-Curtis dissimilarity) between microbiomes of asymptomatic and CDI patients and found no 129 clustering by EIA status (PERMANOVA, P=0.69) or levels of C. difficile ( Figure 1B). Previous comparative 130 microbiome studies have revealed phylum-level differences in CDI cases versus controls not colonized 131 with C. difficile [24]. In contrast, we found no significant differences in relative abundance of bacterial 132 phyla between asymptomatically-colonized patients and patients with CDI (Supplementary Figure 1B).

133
Instead, we hypothesized that differences between these states may manifest at higher resolution. We 134 used a multivariable regression model, as implemented by MaAslin2[25], to identify microbial taxa 135 predictive of either group. C. difficile was the strongest predictor of CDI state, whereas non-C. difficile 136 Clostridia were predictive of asymptomatic state ( Figure 1C, FDR < 0.25; Supplementary Table 3).

137
Correspondingly, we saw increased C. difficile relative abundance in CDI patients and increased levels of 138 a number of non-C. difficile clostridial species, including Eubacterium spp., Dorea spp., and 139 Lachnospiraceae spp. in asymptomatic patients (Supplementary Figure 1C). Given that C. difficile levels were an overt predictor of CDI, we analyzed patient microbiomes regardless of state to understand 141 microbial features that might correlate with C. difficile. Using CoNet [26], a software package that employs 142 multiple measures of correlation to define microbial networks, we found that C. difficile anti-correlated with 143 a number of previously identified clostridial taxa (Supplementary Figure 1D). Finally, given the differences 144 in antibiotic exposure in these cohorts, we also interrogated taxonomic features that were predictive of 145 antibiotic exposure. Interestingly, we found that taxonomic features that were predictive of CDI state were 146 also associated with antibiotic exposure ( Figure 1D). Our data indicates that patients with C. difficile 147 colonization or CDI do not have grossly different gut microbiome community structures but instead have 148 distinctive alterations in a subset species from class Clostridia in the microbiota.

149
Colonization resistance is often conferred through the presence of commensals that outcompete 150 pathogens in a metabolic niche of the gut [27]. To identify metabolic pathways in other clostridia that 151 might enable them to outcompete C. difficile, we defined metabolic potential in patient microbiomes using 152 HUMAnN2 to quantify microbial pathway abundances. In line with our taxonomic analysis, we found no 153 significant differences in alpha-or beta-diversity between overall metabolic pathway composition in the 154 two patient microbiome groups (Supplementary Figure 2A,B). Therefore, we trained an elastic net model 155 to identify specific pathways associated with disease ( Figure 2A). We found a number of carbohydrate 156 degradation pathways and amino acid biosynthetic pathways associated with the asymptomatically-157 colonized (Cx+/EIA-) patients, including 'sucrose degradation III', 'fucose and rhamnose degradation', and 158 'L-methionine biosynthesis I.' Investigation of the genera that encode such pathways revealed that the 159 sucrose degradation III pathway was increased in asymptomatic patients, largely due to Blautia spp. and 160 Faecalibacterium spp., among a number of other Firmicutes genera ( Figure 2B). Interestingly, the fucose 161 and rhamnose degradation pathways were entirely defined by Escherichia spp., presumably E. coli. This

162
suggests that metabolic functions such as fucose/rhamnose degradation may be confined to a smaller 163 number of taxa than carbohydrate degradation pathways such as sucrose degradation. Our metabolic 164 pathway analyses highlight differentially abundant carbohydrate degradation processes that could 165 contribute to colonization resistance against C. difficile in patient microbiomes.

166
To define metabolic determinants of protection from C. difficile in asymptomatic patients, we 167 leveraged metabolomics profiles of these patients' stools [19]. Ordination of Euclidean distances between Cx+/EIA-and Cx+/EIA+ stool metabolomes revealed no significant differences in metabolome structure 169 (Supplementary Figure 2C). We again used MaAslin2 to determine metabolites associated with each 170 disease state. Consistent with previous analysis, a number of end-product Stickland fermentation 171 metabolites (4-methypentanoic acid and 5-aminovalerate) were associated with CDI patients. While we 172 found that 4-hydroxyproline was the strongest predictor of asymptomatically-colonized patients, many of 173 the significant metabolites that were associated with asymptomatic patients were predicted to be 174 carbohydrates ( Figure 2C, FDR < 0.25; Supplementary

179
Supplementary Figure 3). These data reveal a carbohydrate signature that is depleted in CDI patients.

180
Notably, fructose and rhamnose are either substrates or products of the 'sucrose degradation III' and 181 'fucose and rhamnose degradation' pathways, which we found to be enriched in asymptomatic patients.

182
The co-occurrence of these microbial pathways and their corresponding metabolites suggest that the 183 presence of a commensal carbohydrate metabolism that could antagonize C. difficile pathogenesis.

184
We hypothesized that the differential abundance of identified stool metabolites in these patient 185 cohorts is related to the metabolism of specific microbes in their microbiomes. We performed a sparse 186 partial least-squares-discriminatory analysis (sPLS-DA) with the mixOmics package to define

189
Of the strongest metagenomic variable weights, four out of five species (C. difficile, a Lachnospiraceae 190 spp., Anaerostipes hadrus, and Clostridium clostridioforme) were also significantly associated with an EIA 191 state ( Figure 1). In the metabolomics block of the latent component ( Figure 3A), the eight highest-192 weighted metabolites were also discovered by previous analyses (Figure 2). Using the variables defining 193 the latent component, we performed correlational analyses ( Figure 3B) and found a number of striking 194 correlations. C. difficile and Stickland metabolites (5-amino-valeric acid and 4-methylpentanoic acid) [19] 195 were positively correlated, whereas C. difficile had negative relationships with fructose, rhamnose, and hydroxyproline. Moreover, we found that carbohydrates, such as fructose and rhamnose, were correlated 197 with the presence of commensal Lachnospiraceae and Streptococcus species; the Lachnospiraceae spp.

198
was also previously anti-correlated with C. difficile (Figure 1). This network revealed both expected 199 relationships, highlighting the known pathophysiology of CDI, and novel commensal-carbohydrate 200 relationships that define asymptomatic colonization.

201
Our examination of taxa, metabolic pathways, and metabolites revealed that presence of non-C.

202
difficile clostridial taxa that might provide colonization resistance against C. difficile through their 203 metabolism. We hypothesized that the observed inverse relationship of certain carbohydrates to C.

204
difficile might indicate that these metabolites are not be digestible by C. difficile or that these metabolites 205 are end-products of a more complex commensal metabolism that is exclusionary to C. difficile. Using a 206 set of clinical C. difficile isolates cultured from this patient cohort (6 different ribotypes), we examined 207 growth of C. difficile on carbohydrates associated with asymptomatic patients. Using a defined minimal 208 media (CDMM[28]), we found that C. difficile isolates grew robustly on fructose as expected (median 209 maximum A600 of 0.72), but did not proliferate on rhamnose or lactulose (median maximum A600 of 0.18 210 and 0.20 respectively). Notably, in the case of sorbitol, we found that a subset of strains, including the 211 reference strain VPI1064, grew to a maximum A600 of greater than 0.35 ( Figure 4A,B). Given that we had 212 found sucrose degradation as a metabolic pathway enriched in asymptomatic patients, we hypothesized 213 that C. difficile would be unable to use this carbohydrate. Indeed, when grown on sucrose as the sole 214 carbon source, strains achieved a median maximum A600 of ~4.2-fold less than that of growth on fructose.

215
Though C. difficile cannot grow on rhamnose as the sole carbohydrate, in other organisms 216 rhamnose has substantial transcriptional influence over carbon catabolite gene clusters [29,30]. We 217 wanted to rule out the possibility that rhamnose may impact C. difficile through possibly cryptic 218 transcriptional reprogramming, perhaps contributing to C. difficile repression in vivo. Accordingly, we 219 performed whole transcriptome RNA sequencing on C. difficile cultures exposed to a metabolizable 220 substrate, fructose, or a non-metabolizable substrate, rhamnose. In the presence of fructose, we found 221 555 genes significantly altered (adjusted p-value <0.05 and |fold-change| > 2)(Supplementary Table 5).

222
Some of the most altered genes were indicative of carbon catabolite repression of sugar transport and 223 upregulation of glycolytic processes to metabolize fructose. In contrast, we found only 3 genes significantly increased in the rhamnose condition (Supplementary Figure 4). The lack of striking systems-225 level or targeted (toxin expression, sporulation) regulation by rhamnose, and C. difficile's inability to utilize 226 it, leads us to conclude that its association with asymptomatic patients' microbiomes is not through direct 227 interaction or suppression of C. difficile. Instead, we speculate that rhamnose may be the byproduct of a

246
The results from our cross-sectional multi-omics profiling of these CDI-related human cohorts

256
On the other hand, identification of the metabolites sorbitol and 4-hydroxyproline in the stool of 257 asymptomatic patients colonized with toxigenic C. difficile may indicate low-level inflammation. Recent 258 data indicate that hydroxyproline and sorbitol are host-derived metabolites, reflective of some amount of 259 collagen degradation and toxin-mediated inflammation [12,41]. In a mouse model of CDI, the presence of 260 sorbitol and mannitol before pathogenesis was interpreted as a "pre-colonized state" [11,23]

281
We also found an increase in rhamnose in asymptomatic patients, and confirmed that across a 282 number of clinical C. difficile isolates, growth with rhamnose as the sole carbohydrate source is severely 283 decreased relative to the robust growth observed with fructose. Transcriptional profiling of rhamnose-  Our multi-omics analyses of a colonized asymptomatic patient population support a growing body of 297 literature concerning commensal metabolism as a tool against C. difficile. Evidence from both mouse 298 models of disease and human studies indicate that administration of polysaccharides or 'microbial 299 accessible carbohydrates' may prevent C. difficile proliferation or decrease the risk of CDI [16-18, 43, 49].

300
Recently, a probiotics-based attempt to design a consortium of mucosal sugar utilizers revealed its ability 301 to decrease C. difficile colonization in vivo [14], indicating that increasing mucosal metabolism, or 302 carbohydrate catabolism, may be another route to strengthening commensal resistance to C. difficile.

303
Given the plethora of prebiotics and probiotics explored in the C. difficile field, we emphasize the need for 304 an approach that harnesses both probiotic-and prebiotic-based components to inhibit the proliferation of  adapter sequences and DeconSeq[53] to remove human sequences. Samples that were less than 15% bacterial DNA during initial sequencing were discarded, and all samples were sequenced to a depth of at 335 least 5 million reads.

336
We performed taxonomic profiling of metagenomic sequences using MetaPhlAn2 [54] and 337 functional pathway profiling using HUMAnN2 [55]. Metacyc pathway abundances were normalized to 338 relative abundances using the humann2_renorm.py function. The humann_barplot function was used to 339 assess taxonomic composition of metabolic pathways. Custom python scripts were used to parse 340 metaphlan "_profiled_metagenome.txt" and humann2 "pathwayabundance.txt" files. Data were imported 341 to R to analyze community composition and differential associations. MaAslin2 was used to fit a logistic

351
For both microbiome and metabolomics data, the nearZeroVar function of the caret package was used to 352 remove low-prevalent or invariant taxa/pathways/metabolites. These filtered data sets were analyzed for 353 community composition and differential association. Alpha-diversity and beta-diversity were calculated 354 using the vegan package. Bray-Curtis dissimilarity was used as a beta-diversity metric for microbial taxa 355 and metabolic pathways, while Euclidean distance was used as a beta-diversity metric for metabolomes.

356
The MaAslin2 package was used to fit a logistic regression model for microbial taxa associated with EIA 357 status as the fixed variable. No normalizations were applied, and data was log-transformed, with all other 358 default settings applied. This package was also used to fit a logistic regression model for microbial taxa 359 associated with antibiotic exposure. Antibiotic exposure data was converted to a binary variable for this 360 analysis, and microbiome data was treated as above. Finally, MaAslin2 was used to fit a logistic 361 regression model for metabolites associated with EIA status, with the same settings as above.

380
Because isomeric sugars generate very similar spectra, we utilized both spectral similarity and retention 381 time to identify sugar metabolites (Supplementary Figure 2B).

384
The metagenomics relative data was imputed with min(abundance>0)/2, and the metabolomics data was 385 imputed with a value of 1. For both filtered datasets, a centered log-ratio transformation was used to 386 analyze filtered metagenomics and metabolomics datasets above. The mixOmics package in R was used Five mL of each strain (in biological triplicate) were grown to log-phase (OD600 ~ 0.4)in TY and 406 exposed to TY-rhamnose or TY-fructose (with each carbohydrate at 30 mM). Cells were harvested by 407 adding one volume of 1:1(v/v) acetone/ethanol to the culture to arrest growth and RNA degradation.

408
Sample were spun at 4000 x g for 5 minutes. The cell pellet was washed with 500 µl TE buffer (0.5 M 409 EDTA, 1 M Tris pH 7.4) and spun down to remove the supernatant. The cell pellet was resuspend in one 410 mL Trizol and two rounds of bead-beating at 4500rpm for 45s were performed. 300µl of chloroform was 411 added to the suspension, lysates were vortexed, and centrifuged at 4000 rpm for 10 min at 4C. The 412 aqueous layer was removed and RNA was precipitated using isopropanol, washed with 70% ethanol, and 413 resolubilized in TE buffer. Total RNA was treated with Turbo DNase (for two rounds of digestion