Computational approaches to understanding Clostridioides difficile metabolism and virulence

Abstract

The progress of infection by Clostridioides difficile is strongly influenced by metabolic cues it encounters as it colonizes the gastrointestinal tract. Both colonization and regulation of virulence have a multi-factorial interaction between host, microbiome, and gene expression cascades. While these connections with metabolism have been understood for some time, many mechanisms of control have remained difficult to directly assay due to high metabolic variability among C. difficile isolates and difficult genetic systems. Computational systems offer a means to interrogate structure of complex or noisy datasets and generate useful, tractable hypotheses to be tested in the laboratory. Recently, in silico techniques have provided powerful insights into metabolic elements of C. difficile infection ranging from virulence regulation to interactions with the gut microbiota. In this review, we introduce and provide context to the methods of computational modeling that have been applied to C. difficile metabolism and virulence thus far. The techniques discussed here have laid the foundation for future multi-scale efforts aimed at understanding the complex interplay of metabolic activity between pathogen, host, and surrounding microbial community in the regulation of C. difficile pathogenesis.

The role for multiple layers of metabolism in C. difficile virulence induction

Infection by the bacterium Clostridioides (formerly Clostridium) difficile is the most frequent cause of nosocomial diarrhea, with increasing prevalence of community-acquired infection [1]. The hallmark inflammatory disease is induced by secretion of two major toxins (TcdA and TcdB), while select hypervirulent lineages encode for and produce an additional binary toxin (CDT) [2]. Severe cases of C. difficile infection (CDI) can lead to plaques along the intestinal wall composed of high levels of damaged epithelia and neutrophil infiltrate, known as pseudomembranous colitis, and in the most serious instances cause complete distension of the colon [3]. Susceptibility to colonization of the gastrointestinal tract by C. difficile is dependent on perturbation of the gut microbiota, typically through antibiotic use, which has been linked to concurrent shifts in the metabolic environment of the gut [4]. Upon infection the colonic metabolome is further altered to assist accelerated pathogen outgrowth, including increases in growth substrates and other molecules that negatively affect the host or surrounding microbial community [5,6]. Changes in metabolite concentrations appear to not only be the direct result of C. difficile metabolic activity, but also from indirect impacts that the toxins have on the microbiome through actions on the host epithelium [7,8]. Due to growing rates of antibiotic resistance among isolates, conserved elements of C. difficile metabolism in vivo present enticing new targets [9]. As a consequence, a deeper consideration for the role of metabolism in overall C. difficile pathogenesis must be taken to identify novel treatment strategies [10].

A primary factor driving these observed trends in the metabolome are concordant shifts in population density of microbial groups present in the microbiota, which under normal conditions resist C. difficile colonization [11]. The loss of active metabolic functionality in the perturbed microbiota reduces competition for growth nutrients like carbohydrates and peptides along with lowered conversion of germination signals like secondary bile acids [12]. In response to the variable metabolic environment during infection, C. difficile possesses a comparatively large genome with commensurate diverse metabolic capacity [13]. Aside from several more common pathways for anaerobic carbohydrate fermentation [14], C. difficile also utilizes a relatively uncommon form of catabolism known as Stickland fermentation through which pairs of amino acids are inversely reduced and oxidized to ultimately generate ATP and replenish NADH levels [15]. Furthermore, intermediates involved at different points along each of these pathways have been shown to influence expression of numerous phenotypes including virulence [16].

Activation of toxin production itself is almost entirely metabolically regulated, and similar regulation appears to be in place for other virulence factors including flagella, biofilm formation, and sporulation [17–19]. Multiple pleiotropic metabolite-sensitive “master regulator” transcription factors have been identified in C. difficile (e.g. CodY, SigH, CcpA, PrdR), in which several homologs are known to control components of central metabolism, energy generation, and virulence across multiple gram-positive pathogens [16]. Additionally, C. difficile also has been shown to employ phase variation which is a mechanism to promote metabolic and phenotypic diversity among its population to aid in survival following abrupt environmental changes [20]. Further complicating the issue, significant genetic variation has also been found to exist among strains of C. difficile, which has downstream impacts on regulation of metabolic pathways [21,22]. A standard laboratory technique to interrogate metabolic control mechanisms and their connections to virulence among other bacterial species have been genetic knockout studies, however genetic manipulation in C. difficile has been historically difficult [23]. These features combined, along with previously described sensitivity to changes in host and microbiome physiology, make CDI a highly complex and integrated problem which has proven difficult to directly interrogate.

Computational approaches to understand metabolism as it relates to C. difficile virulence

Over the previous two decades, there has been an exponential increase in the application of computational methods to the study of biological systems and infectious disease. Essentially if we understand at least some of the drivers defining the observed biological outcome, computational modeling of complex systems can allow for more rapid identification of knowledge gaps and suggest possible mechanisms. Specifically within the field of C. difficile research, the number of computational modeling approaches to understanding the infection have grown continuously over time. As many elements of CDI have been historically difficult to manipulate experimentally, computational modeling has presented a valuable toolset to drive laboratory research. As metabolism is inherently connected to numerous aspects of C. difficile physiology and virulence expression through an intricate network of interactions, computational methods of signal deconvolution become even more enticing. Moving forward, this principle grows in importance as omics technologies become less expensive, more accessible, and produce larger volumes of data for characterizing biological systems.

Current approaches each reside largely along a spectrum ranging from empirical to mechanistic models (Figure 1). Empirical methods primarily focus on capturing and describing emergent patterns within data, without significant prior knowledge of specific drivers for a given phenotype. Alternatively, mechanistic models rely on explicit interaction rules for elements within the system and allow for a degree of extrapolation with novel factors that influence predictions. Importantly, these categories of techniques are not entirely mutually exclusive and elements of both can be found across approaches and choice of model depends more on the biological question of interest. Within C. difficile research thus far, there have been two additional subgroups of computational methods within each of the major categories. While distinct, each family of algorithms have sought to make new connections between metabolic pathways and presence or expression of virulence traits in C. difficile, and are subsequently discussed in order of increasing computational granularity. We also attempt to provide an entry point for each concept that will be useful to any reader, regardless of previous expertise.

Figure 1). Forms of in silico analysis applied to C. difficile metabolism and virulence.

Generalized breakdown of computational approaches used thus far to understand aspects of C. difficile metabolism or infection. The techniques in recent studies fall into four general categories: (A) Multiple nucleotide or peptide sequence alignment; (B) Gene product, metabolite, or microbial species discrete interaction networks; (C) Statistical modeling with machine learning; or (D) Genome-scale metabolic network reconstructions and flux balance analysis. These methods can be placed within two even more general classifications; either empirical which are more descriptive and rely on emergent patterns in the input data, or mechanistic which require predefined syntax or interaction partners for predictions and often require significant upfront knowledge of a system. Many approaches utilized in the studies covered in this review overlap in some form between groupings, and are listed A–D in order of generally increasing granularity.

Comparative sequence alignment

The most common computational approach to comparing metabolic capacity or other genotypic differences between bacterial strains/species is nucleotide sequence alignment (Figure 1A). The process generally begins with raw read collection and quality curation, which is followed by reference-based assembly to form usable regions of long contiguous sequence (contigs) [24]. With either newly assembled contigs or downloaded complete genomes, the user can then perform global or local alignment for the specific regions of interest. Through this comparison of related sequences, an odds ratio is calculated for the level of relatedness between the target regions. These calculations ultimately provide context for structural, functional, or evolutionary differences across bacterial lineages, also allowing for the application of additional metadata to nucleic acid signature frequencies to make connections with related characteristics such as clinical outcomes or metabolic pathways. While not directly a method for predictive modeling of drivers for downstream phenotypes, these techniques have already provided important insights into C. difficile outbreak analysis, antimicrobial resistance, and pathogenicity [25,26].

With the increasing access and affordability of high-throughput sequencing, more frequent multi-strain whole genome characterization has become possible in the clinic. Several studies have now indicated that this procedure vastly outclasses previous targeted amplification-based approaches for correctly distinguishing pathogenic strains [27], and draft genomes for a significant number of C. difficile clinical isolates are now publicly available. Recently, a large-scale analysis effort for genomes from >12,000 clinical C. difficile isolates from around the globe was published and indicated high levels of divergence among virulence genes [28]. Genomic characterization of clinical isolates has supported strong connections to metabolism as mutations arise over the course of recurrent infection [29]. In another recent study, a group leveraged whole genomic alignment across 134 fully-sequenced clinical isolates of C. difficile and found significantly higher levels of genetic variability among strains with higher rates of recurrence [21]. Furthermore, the authors also found that these mutations occurred more frequently in metabolic genes involved in the catabolism of amino acids which may have implications in the physiology of specific hosts or their microbiomes. These results not only highlight the value of sequence-based analyses in finding novel connections between C. difficile metabolism and clinical outcomes, but also the importance of applying secondary clinical metadata to a preclinical computational analysis when possible.

Generalized interaction networks

Another frequently used form of analysis is the construction of directional interaction networks based on known biological syntax (Figure 1B). The structure and scope of these networks are primarily user-defined, often represented by a defined list of previously established or inferred interactions. An advantage to these models is that they are extremely computationally lightweight and able to be constructed even from minimal, strictly qualitative biological data. These systems then allow for rapid investigations into behaviors of the system as a whole and make tractable hypotheses about components or processes where less is understood. Specifically in terms of CDI, analysis of transcriptomic abundance data from C. difficile during infection using a relatively simple interaction network structured by general metabolic pathway annotations, demonstrated distinct patterns of catabolic preferences across distinct infection models that correlate with long term colonization patterns [30], underscoring the value of added structure to omics data in infectious disease research.

A further approach within this class of computational methods that has achieved success in CDI research are Boolean dynamic models. Generally speaking, nodes within these networks are composed of discrete interaction partners (e.g. proteins, metabolites, or bacterial species), with directed edges representing binary interactions (e.g. regulatory relationships or transformations). The interaction rules themselves are represented as simple equations containing Boolean operators (OR, AND, NOT, or AND NOT) which drive the behavior of the connected nodes. Recently, one group expanded upon a previously published Boolean architecture of C. difficile and microbiota population interactions which initially found that metabolically similar species tended to co-localize and those with higher degree of metabolic specialization actually had a greater likelihood of excluding the pathogen [31]. The follow-up study expanded on these findings by adding a random search of possible interplay to the model to uncover previously unknown interaction rules in the CDI microbiome, and concluded that under certain conditions populations of the genus Enterococcus directly augments the colonization of C. difficile in the gut [32]. Several lines of investigation have positively correlated abundance of Enterococcus species with CDI incidence [33], and these results demonstrate the potential of Boolean models in uncovering previously unknown ecological relationships. Alternatively, another study utilized 151 published transcriptomes of C. difficile across diverse growth conditions to generate a transcriptional regulatory network of known and inferred C. difficile transcription factors and their targets, with very similar syntactic relationships to a Boolean network. Through this platform, they found that ethanolamine metabolism is likely co-regulated with Stickland fermentation loci and expression may increase concordantly with toxin-mediated inflammation [34]. In addition to the novel patterns of C. difficile physiology, these results show that these models also can accurately recapitulate complex layers of gene regulation in C. difficile as they relate to disease.

An iteration of mathematical modeling that is more focused on exploring the mechanisms driving a given system is the design of ordinary differential equations (ODEs). Principally, an ODE model consists of a single or discrete series of polynomial equations that describe rates of change of variables in the chosen training dataset. As the number of interacting variables included in these equations grows, it creates a form of interaction network. ODEs have been applied extensively in biomedical research, especially within the areas of cell signaling and epidemiology [35]. Although not widely applied to C. difficile metabolism and virulence research, a recent study sought to construct an ODE model for the interactions of C. difficile with specific bacterial competitors, access to germination signals, and the presence of growth antagonists during infection [36]. Their model was able to capture the relationship between bacterial conversion of secondary bile acids and the ability of C. difficile to successfully colonize the gut, emphasizing the importance of members from Lachnospiraceae and Lactobacillaceae in colonization resistance. This work lays the foundation for future efforts to create models that also integrate factors like competition for growth substrates, toxin-induced damage, and immune response effects from the host.

Statistical modeling

Several forms of computational analysis that have been applied to C. difficile metabolism fit in the category of statistical modeling methods like regression analysis or machine learning (Figure 1C). In many ways these models are the most empirical of the techniques covered in this review. These approaches leverage systems of equations to first fit trends in data before assessing emergent properties, instead of using biological syntax to define a model to then measure which elements of the data are explained and which are not. The most broadly applied technique in this group is an approach to supervised machine learning known as Random Forest, which creates ensembles of decision trees with stochastically selected features at each branch point resulting in classification models that are robust to overfitting. Multiple studies have leveraged this method using omics datasets from infection to address questions surrounding metabolism in the gut microbiome of CDI [5,12,37]. A strong example of this combined paired 16S gene sequencing and metabolomics profiling data into a single machine learning analysis which suggested that underrepresented bacterial species in the antibiotic-treated gut may have a disproportionate impact on the metabolome and subsequent patterns of CDI clearance [5]. Collectively, these analyses have contributed to deeper understanding for the landscape of metabolic changes of the gut that precede or occur in response to C. difficile colonization. While the strength of statistical models is that they are readily applied to new datasets and are generally unbiased, they can begin to struggle when attempting to identify underlying biological mechanisms.

Genome-scale metabolic network reconstructions

The final category of techniques described here includes genome-scale metabolic network reconstructions (or GENREs), which are computational formalizations of the metabolic capacity encoded in the genome of a given organism (Figure 1D). Mathematically these models are represented as a large matrix of stoichiometric coefficients for metabolic reactions and their substrates and products, with corresponding reaction activity constrained by known biological and physical parameters. To then predict the flow of metabolites through the cell an objective function, usually generation of biomass, is selected for maximization using a linear programming method known as flux balance analysis (FBA). Although the generation of an initial network reconstruction can be partially automated based on genome and biochemical reaction annotations, creating high fidelity GENREs requires extensive literature-driven curation for target organism-specific metabolic processes [38]. New models are then typically validated against experimental screens for gene essentiality and carbon substrate usage to ensure reliability of downstream metabolic predictions [39]. Overall, this method is a heavily bottom-up approach to constructing a functional model of the target species’ metabolism. However, in the end these models create a platform for growth simulation and rapid prediction for the impacts of genotype on many observable metabolic phenotypes that has enhanced research efforts for a number of human bacterial pathogens [40].

The first curated GENREs for C. difficile focused on the widely utilized laboratory strain 630 due to the large volume of previously published molecular and phenotypic characterization for the isolate [41–43]. These studies performed systematic in silico knockout screens, identifying essential genes for infection-like conditions and cross-referenced the results with established inhibitors for protein homologs to quickly generate a list of novel potentially druggable therapeutic targets. Due to the close connection to gene annotation and metabolic function inherent to GENREs, they also provide a strong platform for the contextualization of a variety of omics data types; including genomics, RNA-Seq, proteomics, and metabolomics. This ability to contextualize high-dimensional omics data was exemplified in a recent study, a group generated a new GENRE of str. 630 and used it as a platform for large-scale analysis of whole genome alignment data [44]. By first integrating high throughput phenotypic screens of carbon source utilization with full genome sequencing of three separate strain 630 isolates, the authors streamlined a semi-automated curation pipeline to rapidly generate a high-quality GENRE for the laboratory strain. They applied this reconstruction to an alignment of 415 C. difficile genomes to identify subnetworks of metabolism with lower conservation across the strains. This analysis of core genomic metabolic capacity revealed that while Stickland fermentation pathways are generally conserved, but found that simple polysaccharide catabolic pathways were much more susceptible to mutations.

Generation of context-specific models of active metabolism has also been an effective approach to understanding the metabolism of C. difficile most associated with changes in virulence during infection. Along these lines, a group created GENREs for both str 630 as well as a hypervirulent isolate str. R20291 and leveraged computational methods for integration of transcriptomic data into network models to more accurately simulate growth under conditions of increased virulence expression [45]. This analysis indicated significantly increased usage of the Pentose Phosphate Pathway in addition to higher levels of both cytidine and N-acetylneuraminate usage when expression of multiple virulence factors is suppressed. Expanding to the higher levels of complexity, a recent study also integrated the abundances of bacterial species within samples from 93 recurrent CDI patients and 40 fecal microbial transplant samples to create models of active community metabolism [46]. Building off of their previous work simulating C. difficile growth in competition with prominent members of the gut microbiota during infection [47], the authors performed whole-community metabolic modeling and found that members of Enterobacteriaceae may create favorable conditions for C. difficile colonization through increased consumption of tyrosine and tryptophan, as well as reduced bioconversion of secondary bile acids. Both of these amino acids are not known to be preferred Stickland fermentation substrates of C. difficile and may indicate decreased competitive exclusion of the pathogen [13].

Synthesis and Future Efforts

As C. difficile virulence expression determinants are at the convergence of host, microbiome, and pathogen metabolic and regulatory networks, computational models are powerful means of signal deconvolution to identify novel driving mechanisms. These methods have provided a window into multiple aspects of C. difficile metabolism as they relate to virulence that may have been otherwise inaccessible experimentally. Many of the studies discussed in this review leveraged a combination of techniques, and the success of this integration of approaches highlights how these analyses can reveal novel insights not possible through a single computational method. Predictions have ultimately provided a means to create testable hypotheses to accelerate wet lab research and in turn inform further refinements to in silico models, completing the cycle of systems biology. Moving forward as datasets become even more available and modeling techniques improve, these systems will only increase in value for understanding new relationships in C. difficile metabolism and virulence.

In the near future, hybrid and multiscale modeling approaches will likely become significantly more integral to CDI studies that seek to capture variation from across organizational levels of biology (Figure 2), making more complex mechanism inferences possible [48]. Additionally, as omic profiling becomes more accessible and computational frameworks for integrating this data become more robust, greater degrees of multi-omic integration should become possible. Combined, these factors promote a more complete understanding for the total ecosystem of CDI and will open new avenues for possible treatments and interventions. Systems approaches can allow for the identification of novel points of regulation that may be targeted to downregulate pathogenicity and avoid further accumulation of antibiotic resistance. Layering of secondary regulatory networks with GENREs of other bacterial pathogens has already proven successful for uncovering novel metabolic regulation of virulence expression [45]. Additionally, since metabolism of the microbiome is so strongly tied to the progression of CDI, systems models should add utility in deciphering active mechanisms in the success of fecal microbial transplant and suggest strategies for generating more targeted consortia against CDI [49–51]. Furthermore, dynamic computational models incorporating both host and microbial elements could help us to understand shifts in the metabolic environment of the gut as they relate to increased inflammation and inform possible interventions to reduce excess epithelial damage in response to the infection. On a final note, this field of research would also greatly benefit from making computational methods more accessible to non-experts. Future efforts to generate broadly applicable in silico tool sets, with more intuitive interfaces and detailed user instructions should be encouraged by both reviewers and funding agencies.

Figure 2). Example schema for multi-scale modeling approach for the role of metabolism in CDI.

Numerous opportunities exist for integration of multiple layers of metabolic and virulence modeling in CDI research, with one possible realization shown here: (A) An ODE model of metabolite concentrations from diet and other microbes entering or leaving the system, also accounting for those being moved across the host epithelium. (B) A Boolean network determining immune cell response to changes in metabolite and microbial antigen concentrations. (C) Multiple interacting GENREs for members of the microbiota as well as C. difficile with simulated growth patterns driving changes to their metabolic environment. (D) Another Boolean network determining virulence expression within C. difficile itself that is sensitive to metabolite availability. (E) Finally, supervised machine learning for integration and processing for the large amount of signal generated by such a multi-layered dynamic system to enable usable hypothesis generation for focused downstream testing.

Highlights.

• C. difficile responds to a variety of metabolic signals for determining virulence expression, but the combinatorial nature of these signals makes their study in isolation difficult

• Computational modeling allows for a reductionist and systematic approach to assessing the role of metabolic pathways in bacterial pathogenesis

• A variety of in silico techniques have been implemented to explore specific elements of C. difficile metabolic capacity as it relates to infection

• Multi-scale computational modeling efforts may deconvolute complex interactions between pathogen, microbiota, and host to greatly augment laboratory research