Multi-Omics Strategies for Investigating the Microbiome in Toxicology Research

Abstract

Microbial communities on and within the host contact environmental pollutants, toxic compounds, and other xenobiotic compounds. These communities of bacteria, fungi, viruses, and archaea possess diverse metabolic potential to catabolize compounds and produce new metabolites. Microbes alter chemical disposition thus making the microbiome a natural subject of interest for toxicology. Sequencing and metabolomics technologies permit the study of microbiomes altered by acute or long-term exposure to xenobiotics. These investigations have already contributed to and are helping to re-interpret traditional understandings of toxicology. The purpose of this review is to provide a survey of the current methods used to characterize microbes within the context of toxicology. This will include discussion of commonly used techniques for conducting omic-based experiments, their respective strengths and deficiencies, and how forward-looking techniques may address present shortcomings. Finally, a perspective will be provided regarding common assumptions that currently impede microbiome studies from producing causal explanations of toxicologic mechanisms.

Many roles have been proposed for the trillions of microorganisms that live on and within humans, and interest in understanding their effects on human health remains strong. Advances in sequencing platforms, bioinformatics, analytical chemistry, and related technologies have helped to reveal new interactions between microbes and their host, generating new interest into this area of research. Within the broader field of microbiome research, the human gut microbiome is the most studied and is likely to have an important role in toxicology; therefore, the focus in this review will center on the gut microbiome. However, the technologies and analytical methods described herein can and are being applied to investigate other microbiomes and their interactions with toxic chemicals including environmental pollutants.

The microbiome was originally defined by Whipps et al. (1988), and updates to this definition were proposed by Berg et al. (2020) to incorporate new discoveries and technological developments. The definitions provided by Berg et al. (2020) are followed in this review. In summary, microbiota is defined as the composition of five kingdoms of microorganisms (Bacteria, Archaea, Fungi, Protists, and informally, algae), while the microbiome considers microbiota and their “theatre of activity.” This contains the complete assembly of proteins, metabolites, mobile genetic elements (including viruses), and physical environmental conditions (Berg et al., 2020). The host influences the microbiome through metabolism, dietary choices, pharmaceuticals (Geller et al., 2017; Lee and Ko, 2014), and by providing the physical environmental conditions (Ding and Schloss, 2014; Donaldson et al., 2016). The intertwining of influences by microbiota on their host and vice versa dictate evolutionary pressures on all organisms involved (Drew et al., 2021; Reese et al., 2021; Singh et al., 2021).

Evolutionary links between host and microbe are an essential component of understanding how changes to microbial community composition affect the microbiome, resulting in an impact on human health. Changes to microbiota due to diets rich in processed foods (Asnicar et al., 2021; Chassaing et al., 2022) and industrialization (Groussin et al., 2021; Mancabelli et al., 2017; Reese et al., 2021) have produced dramatic changes in community profiles (Turnbaugh et al., 2009; Voreades et al., 2014), horizontal gene transfer (Groussin et al., 2021), and in disease (Blacher et al., 2019; Elhenawy et al., 2021; Gao and Liu, 2017; Yoon et al., 2021). Along with changes to the diet, industrialization has expanded the human exposome (Logan et al., 2018) with new environmental toxicants and ingested xenobiotics, shaping the microbiome. Together, these changes have produced microbial communities that can be less diverse, more susceptible to perturbations, and less genetically stable (Chen et al., 2021). These evolutionary dynamics and environmental changes provide a case for the importance of the microbiome in toxicology.

The human body hosts multiple microbiomes relevant to toxicology, including the gut, respiratory (Perrone et al. 2021), skin (Byrd et al. 2018), and oral (Willis and Gabaldón, 2020) microbiomes. These distinct environments differ in physical conditions, microbial composition, and exposure to toxicants, however, overlap between them provides opportunity for interaction (Park et al. 2021; Pathak et al., 2021). While these microbiomes are gaining increased attention, the gut microbiome remains the focus of microbiome-related research due to the gut acting as the primary means of nutrient and metabolite uptake (Donaldson et al., 2016) for distribution throughout the body (Chen et al., 2019; Cryan et al., 2019; Tripathi et al., 2018). Consequently, the microbiome has direct impacts on multiple organ systems and related disease. For example, indoxyl sulfate, a uremic toxin (Ellis et al., 2016) with links to cardiovascular disease (Gao and Liu, 2017), is produced in the liver from indole (Gao and Liu, 2017) and reported to be nearly absent in hemodialysis patients without colons (Aronov et al., 2011) and in germ-free mice (Wikoff et al., 2009), indicating a possible link to the gut microbiota. The reach of microbiome has been proposed to extend to the central nervous system with implications in neurological development (Rothenberg et al., 2021), Parkinson’s disease (Kim et al., 2019), amyotrophic lateral sclerosis (Blacher et al., 2019), and depressive behavior (Lukić et al., 2019; Partrick et al., 2021), in addition to the more direct effects of the gut microbiota on the gastrointestinal tract. Tryptophan-derived metabolites from microbial metabolism act as agonists for aryl hydrocarbon receptor (AHR; Gao et al., 2018; Hubbard et al., 2015), a ligand-activated transcription factor, which is linked to maintenance of intestinal barrier integrity (Metidji et al., 2018), immune response (Gutiérrez-Vázquez and Quintana, 2018), inflammation (DiNatale et al., 2010), neurological (Wei et al., 2021), among other effects (Beischlag et al., 2008; Dong et al., 2020). Additionally, physical interactions between bacteria and intestinal epithelial cells have been investigated, as in the case of type IV secretion system in strains of Escherichia coli and its role in Crohn’s disease (Elhenawy et al., 2021), or in the attachment of bacteria to epithelial cells (Donaldson et al., 2016).

The negative consequences of disrupting the microbiota should remain a consideration for toxicology as the introduction of xenobiotic compounds may lead to acute and/or long-term (Shao and Zhu, 2020; Zhao et al., 2020) changes to the microbiome. Three primary avenues exist for the microbiota to interact with ingested xenobiotic compounds in the gastrointestinal tract (Figure 1). The most direct interactions are toxic effects leading to microbial cell damage or death. Gao et al. (2017) provided an example of this dynamic by linking perturbations in microbiota composition to changes in the metabolome in mice exposed to lead. For example, upregulation of coenzyme A disulfide reductase levels and an increase in abundance of the gene encoding MutT, a nudix hydrolase, suggest bacteria respond to oxidative stress caused by lead-generated reactive oxygen species, while a significant increase in abundance of phosphate ABC transporter gene may indicate that bacteria sequester lead by forming phosphate salts as a second means of protecting itself from lead. The authors suggest that in addition to damaging bacterial cells, they may protect the host from lead exposure, offering new insights related to lead-induced disease in the host (Gao et al., 2017). Microbes may also metabolize compounds related to drug treatments, leading to side effects (LoGuidice et al., 2012), less effective treatments (Enright et al., 2016; Geller et al., 2017), or mitigating toxic effects (Sharma et al., 2020). Geller et al. (2017) demonstrated the importance of understanding this interaction by significantly improving the efficacy of gemcitabine, a chemotherapy drug used to treat patients with pancreatic, lung, breast, or bladder cancers, against resistant colon carcinoma tumors. Mycoplasma hyorhinis and other bacteria possessing the long isoform of cytidine deaminase are capable of metabolizing gemcitabine, a chemotherapeutic drug, into the inactive metabolite 2,2′-difluorodexoxyuridine. Colonization of tumor microenvironments by these specific bacteria confer gemcitabine resistance to the tumor, reducing or eliminating the efficacy of this cancer therapy. Co-treatment with gemcitabine and ciprofloxacin, a broad spectrum antibiotic, resulted in significantly reduced tumor growth in a subcutaneous model of colon carcinoma in immunocompetent mice (Geller et al., 2017). The study illustrates the importance of considering the microbiome when evaluating the effect of xenobiotic compounds on human health, including introductions through medical treatments. Finally, the microbiota can produce toxic compounds through the metabolism of xenobiotics or through general metabolism. Dall’Erta et al. (2013) showed that Fusarium mycotoxins, deoxynivalonel and zeralenone, are hydrolyzed by gut microbiota, releasing toxic aglycones. Using an in vitro process designed to simulate gastrointestinal digestion, the authors reported three masked mycotoxins, deoxynivalenol-3-glucoside, zearalenone-14-glucoside, and zeralenone-14-sulfate, undergo a significant conversion to active mycotoxins, relative to microbe-free controls (Dall’Erta et al., 2013). Masked mycotoxins describe a class of molecules that are likely to remain undetected due to differences in physicochemical properties (Berthiller et al., 2013), leading to underrepresentation of mycotoxins in cereal commodities (Dall’Erta et al., 2013). The authors posit that the microbe-driven metabolism of masked mycotoxins could lead to higher practical amounts of mycotoxin in cereal commodities than has been conventionally accepted (Dall’Erta et al., 2013). Ignoring the microbiome and its microbiota risk oversimplifies complex mechanisms involving how xenobiotic compounds affect human health.

Figure 1.

A, Exposure to xenobiotics may have a toxic effect on members of the microbiota, altering the community composition and affecting the host. B, Microbial metabolism of xenobiotics may mitigate the host’s response to the compound. This may provide protection to the host against toxicants or reduce the effectiveness of pharmaceutical treatments. C, Members of the microbiota may obscure the cause of host toxicity by generating toxins from xenobiotic compounds. D, 16S rRNA gene sequencing, shotgun metagenomics, and metatranscriptomics provide detailed profiles of the taxonomic composition and genetics of the gut microbiota. E, Metabolite profiling with LC-MS/MS is used to identify the metabolic contributions of different organisms to the microbiome and identify relationships between detected compounds. F, Statistics methods simplify datasets with hundreds or thousands of features to identify similarities between samples and predict the causes of feature variability. G, Integrating sequencing and metabolomics techniques can provide a means to generate testable hypotheses that can be assessed using gnotobiotic mice and synthetic communities. Created with BioRender.com.

Study of the microbiome has rapidly changed perspectives in toxicology and offers new insights into old research questions. Recent research trends surrounding the role of the AHR as a xenobiotic sensor provides an example of how study of the microbiome may offer new insights and avenues of research. The AHR is present in the cytoplasm of human cells and possesses a promiscuous ligand binding domain, allowing for interaction with a variety of xenobiotic compounds, most notably 2,3,7,8-tetrachlorodibenzo-p-dioxin (TCDD). The discovery and synthesis of TCDD by Sandermann (1957) and an accidental exposure of a laboratory assistant who developed chloracne, a painful skin condition, and recognized symptom of TCDD exposure (Beischlag et al., 2008), set in motion extensive research into the toxicity of TCDD. The next 20 years saw multiple industrial incidents involving the release of TCDD and the recognition by Dow Industrials of possible danger associated with TCDD in the manufacturing of certain industrial chemicals (Beischlag et al., 2008) Research into TCDD exposure led to essential characterizations (Burbach et al., 1992; Poland et al., 1976) of the AHR which further illustrated the importance of AHR in toxicology research. Early research into the AHR has been centered around its role in inducing metabolism of synthetic xenobiotic compounds, such as TCDD, polycyclic aromatic hydrocarbons, or polychlorinated biphenyls (Beischlag et al., 2008; Hubbard et al., 2017). AHR is highly expressed at barrier sites like gut, lung, and skin which are the sites most likely to come into contact with environmental pollutants, but also with microbes (Metidji et al., 2018). In 1991, an early link between the AHR and the microbiome was established by identifying that bacterial metabolism of tryptophan produce AHR ligands (Perdew and Babbs, 1991). The subsequent decades have seen increased interest in investigating this connection between AHR, microbes, and the microbiome. As an example, in 2014, Moura-Alves et al. (2014) connected microbial metabolites, phenazines, and naphthoquinone phthiocol, produced by bacterial pathogens to the AHR. This study is now one of many demonstrating a connection between microbially produced AHR ligands and host immune response (Gutiérrez-Vázquez and Quintana, 2018; Moura-Alves et al., 2014). Furthermore, investigations began into the role of metabolites generated by commensal bacteria, especially tryptophan metabolites (Brown et al., 2020; Dong et al., 2020; Jin et al., 2014; Roager and Licht, 2018), and their ability to activate AHR (Hubbard et al., 2015; Jin et al., 2014; Korecka et al., 2016). This new line of inquiry has expanded understanding of how microbial metabolites influence inflammatory bowel diseases, cell proliferation (Nyström et al., 2021), tumorigenesis (Wu et al., 2018), and maintenance of gut homeostasis (Murray and Perdew, 2017) through AHR activation. AHR was formerly considered only in the context of environmental and synthetic pollutants, but recent research now mandates that the microbiome must be considered to fully understand the impact of xenobiotic compounds on human health. AHR provides an example of how studying the microbiome offers new insights into understanding long-asked questions within toxicology, and as technologies for studying the microbiome improve, more examples will come to light. This has already been demonstrated with other xenobiotic receptors. The pregnane X receptor is well known for its role in drug metabolism, but its activity and subsequent influence on the microbiome are also reported (Dutta et al., 2021; Dvořák et al., 2020). A similar example can be made of the constitutive androstane receptor and its reported role in modifying the composition of gut microbiota (Dempsey et al., 2019; Wahlang et al., 2021), further highlighting the value of incorporating microbiome research into toxicology.

The main objective of this review is to provide an outline for current methods used to study the microbiota, their metabolism, and how to integrate technologies to investigate their effect on the host (Figure 1). Presented alongside these methods are their strengths and deficiencies, as well as forward-looking techniques that may supplement current techniques, or replace them entirely. Through the described techniques, current and future, it is possible to conduct reproducible, thorough, and expand upon the insightful research into the microbiome and its interactions with xenobiotic compounds.

SEQUENCING-BASED METHODS

 

Study of the microbiome is made possible by rapid advances in sequencing technologies and in computational tools that process the large amounts of generated data (Figure 2, Table 1). The challenges to growing many gut microbes using traditional culturing methods make culture-independent sequencing technologies not just a powerful means of understanding microbiota community profiles, but, at present, the only practical means. Metagenomics, the sampling and sequencing of an entire community of organisms (Hugenholtz and Tyson, 2008), is the main approach to profile communities of the microbiome. Two primary approaches are used for metagenomics: amplification and sequencing of the bacterial 16S rRNA gene (Johnson et al., 2019), and the untargeted sequencing of all present DNA in a sample, referred to as shotgun metagenomics (Quince et al., 2017). In addition, metatranscriptomics is as a method for studying the entirety of mRNA transcripts present in a community to survey gene expression and predict metabolic changes (Lau et al., 2018; Zhang et al., 2021). These techniques are invaluable resources for studying the microbiome, but they possess disadvantages intrinsic to their design that must be considered.

Figure 2.

Sequencing-based analysis of the microbiome begins with the extraction of nucleic acids, library preparation, and high-throughput sequencing. Sequencing data is processed using bioinformatics software or pipelines based on the sequencing technique. Popular and user-friendly software for taxonomic identification using the 16S rRNA gene include QIIME2, DADA2, and mothur. Whole metagenome sequencing, also called shotgun metagenomics, and metatranscript sequencing data is often processed using complete software pipelines like HUMAnN2 or SqueezeMeta. All three sequencing techniques provide taxonomic profiling of the microbiome, but shotgun metagenomics and metatranscriptomics also provide profiles of functional potential or expressed genes. Created with BioRender.com.

Table 1.

Widely Used, Open-Source Software Tools for Microbiome Research

Tool	Data Type	Short Description	Analysis	Visualization	Availability and Resources
QIIME2 (Bolyen et al., 2019)	16S rRNA amplicon sequencing	An accessible and complete workflow for processing raw 16S amplicon sequence reads to taxonomic classification and statistical analysis. Additional plugins expand functionality, including the ability to process metabolomics data.	Taxonomic classification, diversity statistics, differential, and statistical analysis. Third party plugins may expand analysis options.	Generates taxonomic bar plots, diversity statistics, and ordination plots. Community generated scripts also available.	GUI, Unix command line, Python 3 API. Robust tutorials provided.
Mother (Schloss, 2020; Schloss et al., 2009)	16S rRNA amplicon sequencing	Processes 16S rRNA gene sequence data to generate taxonomic classifications based on OTUs or ASVs. Provides limited analysis options.	Taxonomic classification, diversity statistics, and basic statistical analysis.	No visualizations.	Unix command line. Tutorials provided.
DADA2 (Callahan et al., 2016)	16S rRNA amplicon sequencing	Processes 16S rRNA gene sequence data based on amplicon sequence variants.	Taxonomic classification and error evaluation.	No visualizations.	R package. Tutorial provided.
Kraken2 (Wood et al., 2019)	Shotgun metagenomics	Assembles amplicon sequencing reads into K-mers to match to reference genomes for identification.	Taxonomic classification and functional pathway analysis.	No visualizations.	Unix command line.
HUMAnN2 (Franzosa et al., 2018)	Shotgun metagenomics, metatranscriptomics	Software pipeline for identifying taxonomy and functional genes based on species or strain-specific marker genes.	Taxonomic classification and functional pathway analysis.	Limited visualizations.	Unix command line. Tutorial provided.
SqueezeMeta (Tamames and Puente-Sánchez, 2019)	Shotgun metagenomics, Metatranscriptomics	An automated pipeline for processing amplicon or long-read sequencing data with little technical background.	Taxonomic classification and functional pathway analysis.	No visualizations.	Unix command line.
XCMS Online (Tautenhahn et al., 2012)	LC-MS metabolomics	A complete workflow for detecting features, retention time correction, peak alignment, annotation, and statistical analysis.	Basic statistical and pathway analysis.	Visualization of statistical analysis.	Web interface. Tutorial provided.
MetaboAnalyst 5.0 (Pang et al., 2021)	NMR, LC-MS metabolomics	A complete workflow for data processing raw data through functional, pathway, and statistical analysis. Other multi-omics data can be integrated into analysis.	Basic and complex statistical analysis, multi-omic integration, pathway, and functional analysis.	Visualization of statistical and network analysis, and for evaluating data processing.	Web interface and R package. Tutorials provided.
MzMine2 (Pluskal et al., 2010)	LC-MS metabolomics	A complete workflow for processing LC-MS raw data through statistical analysis and visualization.	Basic statistical analysis with data export to statistical software.	Visualization of statistical analysis and evaluating data processing.	Linux, OS X, and Windows. Tutorials provided.
MS-DIAL 4 (Tsugawa et al., 2020)	GC-MS, LC-MS metabolomics	A workflow for processing untargeted MS metabolomics data with support for basic data analysis. Processed data may be exported to other software for complex analysis.	Basic statistical analysis with data export to statistical software.	Visualization of PCA, pathway analysis, and molecular spectrum networking.	Linux, OS X, and Windows. Tutorial provided.

16S rRNA Gene Sequencing

The use of the bacterial 16S rRNA gene for community profiling is the most developed and commonly used community profiling technique (Johnson et al., 2019). For analysis of complex communities, single variable regions, like the V4 or V6, or multiple regions, such as V1–V3 or V3–V5, are most often sequenced by Illumina sequencing platforms for use in microbial identification (Johnson et al., 2019). Sequences are clustered into operational taxonomic units by clustering pipelines, often QIIME2 (Bolyen et al., 2019) or mothur (Schloss, 2020; Schloss et al., 2009), which can be compared to reference databases (Balvočiūtė and Huson, 2017; Siegwald et al., 2019), such as SILVA (Quast et al., 2013), RDP (Cole et al., 2014), or RefSeq (O’Leary et al., 2016), for classification often at the genus taxonomic level and rarely at the species level. 16S rRNA amplicon sequencing holds three significant advantages over alternative metagenomics techniques: low starting biomass requirements, low cost, and ease of data processing. PCR amplification of the 16S RNA gene is a critical step in the sequencing workflow, and despite the introduction of PCR bias (Sze and Schloss, 2019; Witzke et al., 2020), amplification allows for a lower starting biomass requirement than shotgun metagenomics (Witzke et al., 2020). The combination of short target sequences and the declining price of Illumina sequencing technology means high sampling depth is possible for studies with limited scope (Sze and Schloss, 2019).

Data processing, compared to shotgun metagenomics, lowers the computational resources required, lowering costs, and offering a simpler pipeline (Bolyen et al., 2019; Schloss, 2020). Additionally, tutorials, workshops, and other resources are widely available for users of 16S pipelines, especially for users of QIIME2 (forum.qiime2.org) and mothur (forum.mothur.org), making the process of learning the data analysis pipeline easier. Yet, processing metagenomic data is not simple and multiple options exist for each step of analysis. Pipeline and database performance benchmarks using characterized “mock” communities have shown that different analysis pipelines report different results (Prodan et al., 2020; Straub et al., 2020). Also, debate continues over the fundamental idea behind how to classify and report taxa using the 16S rRNA gene. It is often argued that operational taxonomic units are classified into genera or species by arbitrary cutoffs, resulting in imprecise classification and difficulty in reproducing results (Callahan et al., 2017; Tsukuda et al., 2017). An alternative to operational taxonomic units was proposed and implemented with the software DADA2 (Callahan et al., 2016), which attempts to classify organisms based on sequence differences, as few as single nucleotide differences, referred to as amplicon sequence variants (Callahan et al., 2017). Despite the adoption of both operational taxonomic units and amplicon sequence variants methods in common pipelines, both are used in publications to profile the microbiota. Alternatively, many authors choose to provide analysis of bacterial communities using multiple pipelines or both taxonomic classification methods (Joos et al., 2020; Jun et al., 2020; Martinson et al., 2019) to demonstrate agreement. Cutting-edge sequencing technologies and computational tools may serve to alter this debate and diminish the drawbacks of 16S rRNA amplicon sequencing. The development of long-read sequencing platforms—enabled by Pacific Biosciences and Oxford Nanopore Technologies—allow for sequencing the entirety of the 16S rRNA gene in a single read. When combined with circular consensus sequencing, which allows for repeated sequencing of the same DNA molecule, long-read sequencing is capable of accurately sequencing complete 16S genes in a high-throughput method (De Coster et al., 2021; Johnson et al., 2019). Full sequences of the 16S gene are improving classification accuracy and taxonomic resolution to the species or strain level (Fuks et al., 2018; Johnson et al., 2019).

To move beyond simple taxonomic descriptions, advances have been made in using 16S-based taxonomic information to predict the functional potential of communities. PICRUSt2, a popular software designed for this purpose, has improved in its ability to predict phenotypes by incorporating new and updated databases, integrating amplicon sequence variants, optimizing pathway abundance predictions, and using reference phylogenies to improve predictions (Douglas et al., 2020). Tax4Fun2 offers an alternative to PICRUSt2 by supplementing 16S data with habitat-specific data, improving functional profile prediction with claimed greater accuracy than the original PICRUSt (Wemheuer et al., 2020). These functional predictions can be combined with an associated taxonomic profile to predict and quantify the contribution of specific taxa to a change in function relative to a control with the computational framework FishTaco (Manor and Borenstein, 2017). The purpose of this tool is to address the challenges of linking taxonomy to shifts in functional potential when using traditional statistical methods, like Wilcoxon rank-sum test or fold ratio. FishTaco reduces the complexity of interpreting metagenomics data, but it is dependent on the availability of a control sample to assess functional shifts and function predictions may be difficult to assign to host effects. Despite these advances, there are reasons to question the underlying principle that taxonomy can perfectly predict metabolic function. From a genetics perspective, inter-species genetic diversity may have a crucial role in specific metabolic pathways that cannot be assessed using marker genes (Ellegaard and Engel, 2016). Horizontal gene transfer occurs frequently within the microbiome (Groussin et al., 2021), introducing new functional potential to strains without altering their taxonomic classification. Furthermore, Chen et al. (2021) studied the genetic stability of individual gut microbiota over four years and found strain level differences that associate with disease states that arose over the course of the study. There is, also, evidence derived from studies using metabolomics to indicate that substantial metabolic differences exist among closely related strains, challenging the accuracy of phylogeny and genomic-based functional predictions (Han et al., 2021). These techniques are predictions and are intended to supplement 16S sequencing with additional information, but they are not replacements for more reliable shotgun metagenomic profiles (Douglas et al., 2020; Wemheuer et al., 2020), transcriptomics, and metabolomics.

The 16S gene also cannot be used to identify eukaryotic taxa and attempts must be made to also target 18S rRNA genes and internal transcribed spacers genes (Popovic and Parkinson, 2018), and there are no universal markers for viral genomes (Shkoporov et al., 2019). Growing interest in these understudied elements of the microbiome may not be suited to rRNA-based identification. These technical limitations may require alternative techniques, such as shotgun metagenomics or metatranscriptomics, but they do not prevent 16S sequencing from remaining a powerful tool for studying the microbiome. Guo et al., (2020; Table 2) provides an instructive example of how 16S rRNA amplicon sequencing can be used to narrow a specific investigation to a few taxonomic groups. The authors investigated the role of microbes in providing radioprotection to mice by taxonomically profiling mouse gut microbiota which survived up to 600 days following a high dose of total body gamma-ray irradiation. Using QIIME to generate a taxonomic profile, the abundances of Lachnospiraceae and Enterococcaceae were identified as potentially important for promoting intestinal repair following exposure to radiation. They further refined their search by inoculating germ-free mice with members of these bacterial families and reported that a cocktail of Lachnospiraceae provided greatest survival and reduced disease severity. Subsequent untargeted metabolomics analysis of fecal samples identified specific metabolites that are often associated with Lachnospiraceae, suggesting a causal link between microbe and radioprotection (Guo et al., 2020).

Table 2.

Examples of Microbiome Altering Xenobiotic Compounds

Toxicant/Xenobiotic	Methods	Observation	Significance	Reference
Radiation	16S rRNA gene sequencing LC-MS/MS	Lachnospiracaea and Enterococcaceae, produce propionate, indole-3-carboxaldehyde, and kynurenic acid, reducing damage and facilitating hematopoiesis and gastrointestinal repair.	Gut microbes protect the gastrointestinal tract from radiation exposure and facilitate recovery.	Guo et al. (2020)
Lead	16S rRNA gene sequencing Shotgun metagenomic sequencing GC-MS	Decreased and inhibited the development of phylogenetic diversity. Bile acids and vitamin E were diminished, while nitrogen and energy metabolism changed. Increased abundance of oxidation defense and detoxification genes.	Lead affects microbiota composition, favoring bacteria resistant to oxidative stress and heavy metals, and altering metabolite production.	Gao et al. (2017)
Benzo[a]pyrene	16S rRNA gene sequencing Metatranscriptomics sequencing GC-MS	No change to microbial composition. Four significantly altered metabolic pathways. Dose-dependent changes in seven volatile metabolites.	Chronic exposure alters the gut metabolome by altering microbial metabolism.	Defois et al. (2017)
Sucralose	16S rRNA gene sequencing. LC-MS/MS	Long-term sucralose exposure changed 14 genera and fecal metabolite levels which alter gene expression.	Sucralose alters microbiota composition and increases liver inflammation.	Bian et al. (2017)
271 orally administered drugs	Shotgun metagenomic sequencing LC-MS/MS	As few as one strain is needed to metabolize two-thirds of assayed drugs while multiple strains can produce dozens of metabolites from a single drug.	Microbial metabolism of drugs is robust and enabled by enzyme diversity.	Zimmermann et al. (2019)
Metformin	Shotgun metagenomic sequencing	Reduced abundance of sulfate-reducing bacteria and butyrate-producing bacteria associated with IBD and T2D. Increases in Enterorhabdus and Bacteroides associated with SCFA production.	Exposure alters microbiota composition with potentially beneficial outcomes.	Silamiķele et al. (2021)
Nicotine	16S rRNA gene sequencing. GC-MS	Nicotine exposure altered microbial functional genes corresponding with altered concentrations of glutamate, gamma-aminobutryic acid, uric acid, and xanthurenic acid.	Microbiome may alter host response to nicotine and demonstrates a gut-brain interaction.	Chi et al. (2017)
PCB 126 (Polychlorinated biphenyl)	16S rRNA gene sequencing LC-MS/MS	Exposure to PCB 126 significantly altered the microbiome, and was associated with increases of inflammatory markers (TNF-α, IL-6, and IL-18) and a decrease in SCFA.	Altered microbiota composition contributes to intestinal inflammation and metabolic disruption.	Petriello et al. (2018)
Glyphosate or Roundup MON 52276	16S rRNA gene sequencing Shotgun metagenomics LC-MS/MS	Altered cecal metabolites indicate an inhibition of the shikimate pathway and an increased response to oxidative stress. Changes in high abundance phyla were not observed, but four low abundance species were significantly altered.	Microbiome compositional changes are minimal and metabolomic changes may not affect host physiology.	Mesnage et al. (2021)

Whole Genome Sequencing

The declining cost of high-throughput sequencing and cloud computing are making whole genome sequencing more accessible. The ability to sequence all present DNA in a sample, as is possible with shotgun metagenomics, provides advantages that are mostly unchallenged by other techniques. Most notably, shotgun metagenomics improves the resolution of taxonomic classification beyond 16S-based metagenomics, allowing classification at the species and strain level (Breitwieser et al., 2019). This advantage was used by Silamiķele et al. (2021; Table 2) to investigate species and strain level differences in the mouse gut microbiota in response to doses of metformin, a treatment for type 2 diabetes, and high-fat diets. The authors report significant decreases in the abundance of species of sulfate-reducing bacteria that have been associated with inflammatory bowel diseases, and Flavonifractor plautii, a fecal biomarker for colorectal cancer. They also observed increased abundances of multiple Enterorhabdus species which are reported to provide obesity resistance, and several Bacteroides species known to produce beneficial short-chain fatty acids (Lee and Ko, 2014). The resolution from shotgun metagenomics led Silamiķele et al. (2021) to hypothesize that metformin metabolism may alter the composition of the gut microbiota and increase the abundance of potentially beneficial bacteria.

There are an ever-increasing number of tools for processing shotgun metagenomics data (Breitwieser et al., 2019), but the described tools represent, currently, the most widely used in microbiome research. As with 16S metagenomics, there are multiple approaches for taxonomic classification using shotgun metagenomic data. Tools like Kraken2 (Wood et al., 2019) or Centrifuge (Kim et al., 2016) use databases containing complete genomes as a reference and provide raw relative abundance outputs in terms of the proportion of sequences assigned to a given taxon out of the total number of sequences (sequence abundance; Lu et al., 2017; Wood et al., 2019; Sun et al., 2021). Profilers like Kaiju (Menzel et al., 2016), DIAMOND (Buchfink et al., 2015), or MMSeqs2 (Mirdita et al., 2021) also provide classifications in terms of sequence abundance, but compare reads against reference databases containing protein coding sequences (Menzel et al., 2016; Sun et al., 2021). The third approach, taken by MetaPhlAn2 (Segata et al., 2012) and mOTUs2 (Milanese et al., 2019), use clade-specific marker genes to provide taxonomic abundance, the proportion of genomes of a specific taxon to the total number of genomes detected (Milanese et al., 2019; Segata et al., 2012; Sun et al., 2021). Evidence supports each of these as valid approaches for classification, using both standardized data sets like Critical Assessment of Metagenome Interpretation (Sczyrba et al., 2017) data. However, it is argued that directly comparing sequence abundance to taxonomic abundance may produce misleading analysis, especially when depending on alpha and beta diversity measures, and should be avoided (Sun et al., 2021). Consequently, it may be beneficial to use a profiler based, in part, on the need to compare results with those of other studies. Although, more than just classification accuracy needs to be considered when choosing a profiler given that required computational resources (memory usage and CPU time) widely vary for these profilers (Ye et al., 2019).

Reconstructing metabolic pathways present among a community of microorganisms is essential for elucidating the mechanisms behind microbial responses to xenobiotic compounds. Advances in shotgun metagenomics and metatranscriptomics are enabling more accurate and descriptive functional profiling of the microbiome. In addition to taxonomic classification, shotgun metagenomics data are used to construct metabolic pathways and assign them to present taxa. Of two main strategies, HUMAnN2 (Franzosa et al., 2018) and MEGAN6 (Huson et al., 2011) use taxonomic classification and reference databases to identify proteins which are used to describe metabolic pathways. Leveraging taxonomic information reduces required computational resources and allows for a relatively rapid approach, however, it relies heavily on the information present in reference databases (Franzosa et al., 2018). The second approach leverages de novo assembly of metagenomics data with implementations in pipelines like metaSPAdes (Nurk et al., 2017) and MEGAHIT (Li et al., 2015). Assembled contigs are used in predicting and quantifying genes, which can be assembled into pathways or ontologies. This approach relies less on reference databases with a trade-off of greater computational cost and introduces error from assembly. Yet, metagenome assembled genomes may provide the only means to acquire draft and complete genomes of uncharacterized organisms which cannot be cultured in laboratory settings (Meziti et al., 2021; Parks et al., 2017). Although there is concern regarding the quality of these draft genomes, for example one study found that metagenome assembled genomes miss 25-50% of core and variable genes (Meziti et al., 2021). However, new binning techniques (Nissen et al., 2021) and long-read sequencing (Singleton et al., 2021) are producing high-quality metagenome assembled genomes.

Understanding the functional potential of gut microbiota provides an additional resource for examining how microbiota shape the microbiome. Gao et al. (2017) used shotgun metagenomics to further investigate taxonomic changes in the microbiome, identified using 16S rRNA amplicon sequencing, due to lead exposure in mice. The addition of functional pathway analysis indicated significant differences in nitrogen and energy metabolism, and genes related to mitigating or repairing oxidative damage. The authors used functional pathway analysis alongside metabolomics data to conclude that lead exposure perturbs gut microbes, altering important metabolites in the microbiome which may provide insights into lead toxicity for the host and into lead-induced diseases (Table 2). Using a similar approach, Chi et al. (2017; Table 2) used 16S rRNA gene sequencing to identify changes to the gut microbiota at family level due to nicotine exposure. The taxonomic changes were associated with sex-specific differences in the fecal metabolome and host response to nicotine exposure. The authors used shotgun metagenomics to investigate the phenotypic consequences of the taxonomic changes and reported sex-specific changes in carbohydrate metabolism, oxidative stress response, and DNA repair. The authors hypothesize that sex-specific differences in functional genes indicate a potential cause for sex-specific host responses to nicotine exposure. Furthermore, Chi et al. (2017) propose that the observed changes to carbohydrate metabolism may help to explain why nicotine exposure leads to greater weight loss in men compared to women. However, shotgun metagenomics can only predict phenotype based on genetic potential (Kuchina et al., 2021), so metatranscriptomics may be needed to provide evidence to demonstrate actual phenotypic change in response to a toxic compound.

Metatranscriptomics

Generating expression profiles are invaluable for understanding the response of the microbiota to xenobiotics. As an example, Defois et al. (2017) applied metatranscriptomics to study the short-term effect of benzo[a]pyrene, a common environmental contaminant, and reported altered microbial metabolism (Table 2). Upregulation of four metabolic pathways (cell wall compound metabolism, aromatic compound metabolism, vitamin and cofactor metabolism, and DNA repair and replication systems) indicate that the microbiota are responding to the presence of this pollutant, despite 16S rRNA sequencing showing no significant differences in microbiota composition. Metabolomic analysis supported this result by showing a dose-dependent change to seven volatile metabolites produced by bacteria. Taxonomic analysis alone would have been insufficient to identify the reported changes to the microbiome, but metatranscriptomics provided future directions to investigate changes to the gut metabolome.

Metatranscriptomics shares many of the strengths and challenges facing metagenomics, but with the added difficulties associated with analyzing RNA. Sequencing preparation may include a structural RNA removal or depletion step because bacterial ribosomal RNA is estimated to comprise 95% of total RNA (Peano et al., 2013). Failure to remove the ribosomal RNA would add unnecessary cost to sequencing runs and less useful information when coding RNA is of interest. Although, ribosomal RNA data can be removed in a pre-data processing step to simplify downstream analysis (Shakya et al., 2019). Most sequencing applications also require synthesizing cDNA by reverse transcription (Zhang et al., 2021), introducing technical biases (Minshall and Git, 2020). Direct sequencing of entire mRNA transcripts using long-read sequencing offers the potential for a transformative solution to this problem (Amarasinghe et al., 2020).

Like metagenomics, many individual tools have been developed to address each step of data processing, many of which have been discussed in detail in recent reviews (Niu et al., 2018; Shakya et al., 2019), however, use of software pipelines that incorporate these tools into a single package is more common. These pipelines take one of two major approaches to analysis: assemble reads into transcripts or analyze reads directly (Shakya et al., 2019). SqueezeMeta (Tamames and Puente-Sánchez, 2019) and IMP (Narayanasamy et al., 2016) take the former approach and assemble reads into larger contigs before analysis improving the ability to search reference databases, especially when paired with metagenomic data from the same community (Shakya et al., 2019; Zhang et al., 2021). In contrast, MetaTrans (Martinez et al., 2016), COMAN (Ni et al., 2016), SAMSA2 (Westreich et al., 2018), and HUMANn2 use read-based approaches for analysis. HUMANn2 uses filtered reads to search reference databases based on nucleotide sequence, while COMAN and SAMSA2 use a BLAST-like tool (DIAMOND; Buchfink et al., 2021) to search protein-based databases based on amino acid sequence. Taxonomic classification based on amino acid sequences allows for recognition of organisms that are more distantly related to the reference strain but is more prone to false positives than the nucleotide-based approach which cannot identify insufficiently conserved sequences (Shakya et al., 2019). MetaTrans is unique among read-based pipelines in its use of FragGeneScan to predict putative genes before mapping to reference protein databases for functional analysis, and by using rRNA sequences, which are typically removed, for taxonomic identification (Martinez et al., 2016). Alternative to these Illumina-based, short-read sequencing pipelines, SqueezeMeta (Tamames and Puente-Sánchez, 2019) is designed to process long-read sequence data. Long-read sequencing may dramatically simplify or eliminate the need for transcript assembly, improving discovery rates and lowering computational demand, however, the cost remains prohibitive for most metatranscriptomics applications.

Challenges for Sequencing Techniques

Despite the different technical and analytical strategies applied to 16S rRNA amplicon sequencing, shotgun metagenomics, and metatranscriptomics, they share many of the same technical, analytical, and reporting challenges. Fortunately, recent attention to these issues has produced a variety of solutions. Popular software pipelines, like HUMAnN2 or QIIME2, incorporate the required tools to process sequencing data, from quality control to assembly to taxonomic or functional classification. This packaging simplifies workflows, making microbiome analysis tools more accessible, and preserving the ability to adjust key parameters during data processing. Furthermore, the standardization of pipelines increases the ease of reporting parameters and addresses concerns of data reproducibility (O’Sullivan et al., 2021) in microbiome research. Both HUMAnN2 and QIIME2 are available within Galaxy (Afgan et al., 2018), an open computational platform to deploy software for internal or publicly available computational resources, with the benefit of being able to automatically record each data processing step for publication (Grüning et al., 2018). A version of this feature has recently been built into other available standalone deployments of QIIME2 (Bolyen et al., 2019) as well—expanding the options for streamlined method reporting. The ability to share explicit data processing workflows is critical for producing reproducible research, yet it does not ensure quality results. Each step in the workflow from extracting DNA or RNA to processing sequence data introduces potential sources of bias or error (O’Sullivan et al., 2021; Siegwald et al., 2019), and the inclusion of sufficient controls have been shown to improve the accuracy of results and validate results.

Standardized mock communities have been designed to be processed alongside experimental samples, permitting validation of workflow performance (Ducarmon et al., 2020), and assessing potential error due to DNA extraction bias (Costea et al., 2017), PCR bias (Gohl et al., 2016), or batch effects (Gohl et al., 2016). In addition, internal control DNA from known bacterial sources (spike-in controls; Hardwick et al., 2018) can be added to experimental samples to identify and adjust for error derived from sample or data processing. Yet, the inclusion of a mock community or spike-in controls may not be sufficient to validate the presence or absence of low abundance taxa. The near omnipresence of bacteria in the environment confounds our ability to process samples free of contamination. Contamination of laboratory reagents and kits is well documented (de Goffau et al., 2018; Olomu et al., 2020; Salter et al., 2014), and may become a problem when considering low-biomass samples or low abundance taxa (Salter et al., 2014). Reagent contamination dictates the inclusion of negative controls that will identify contamination and potential batch effects that are introduced (Kim et al., 2017).

Many publications demonstrate that their approach, method, or tool is highly effective at addressing a specific task, yet it is rare for a method to become universally applicable. As an example, the superiority of shotgun metagenomics over 16S could be argued based on improved taxonomic resolution and the ability to produce functional profiles. This may apply if intra-species diversity is substantial, or if the xenobiotic of interest selectively affects bacterial species; if this is not the case, shotgun metagenomics may produce unnecessary data with a significant financial cost. Furthermore, the ability to predict functional profiles based on 16S taxonomy, using tools like PICRUSt (Douglas et al., 2020), may be sufficient if the subject of the study depends on universal presence of a functional pathway within a specific clade. A functional profile derived from shotgun metagenomics may prove invaluable for understanding which specific organisms and pathways may be relevant to a specific study; although functional potential should not be mistaken for functional reality (Kuchina et al., 2021). Understanding the source of phenotypic changes may be best elucidated using metatranscriptomics; nevertheless, the difficulty in interpreting the data may make accurate reporting of alterations in expression challenging and explaining changes may be aided by a metagenomic profile (Shakya et al., 2019; Zhang et al., 2021). Greater precision can be achieved by pairing gene copy number with transcript abundance (Zhang et al., 2021), offering insights into whether a metatranscriptome change is due to changes in gene expression, or changes in taxonomy and gene copy number. Pairing these data sets also reduces the difficulty of producing reference-free metatranscriptome assemblies and allows the identification of novel protein-coding genes (Zhang et al., 2021). Similarly, the choice of data processing pipelines will change based on the intent of a specific study, available computer processing power, and bioinformatics expertise. Fortunately, multiple pipelines or tools may be applied to a dataset and the results compared to look for congruence and to assess which pipeline more appropriately suits the data or intent.

METABOLOMICS

In the context of toxicology and human health, the taxonomic and functional profiling of microbes tells an incomplete story. The realized metabolic consequences of exposing the microbiome to xenobiotics is essential to detailing the mechanisms at work, and the sequence-based profiling of microbes is currently unable to address other factors within the microbiome that affect metabolism. The profiling of small molecule, commonly referred to as metabolomics, provides a perspective of microbial metabolism considering external factors, such as nutrient availability, host factors, and microbe–microbe interactions. Whereas metagenomics and transcriptomics tells what could happen in the microbiome, metabolomics reports on what did happen. These orthogonal methods allow for moving beyond correlative causes to the detangling of complex metabolic networks that explain the causes behind microbial impacts on human health.

Complete explanations of these dynamics require an analysis of the metabolic contribution of the microbiota to the microbiome, which is made possible through global metabolite profiling. Often this depends on pairing separation techniques, including gas chromatography (GC), ion mobility separation, and liquid chromatography (LC) with detection technologies like mass spectrometry (MS), or the use of nuclear magnetic resonance spectroscopy (NMR). These technologies enable both a targeted approach using internal standards for quantifiable, high-confidence metabolite identification, and an untargeted approach for maximizing coverage of the metabolome (Ribbenstedt et al., 2018). Over the past two decades, advances in extraction methods, chromatography, detection technology, and data processing have rapidly expanded the scope of applications for metabolomics in both toxicology and microbiome research. Despite the many advances, the field faces several challenges to the ability to rapidly analyze and accurately identify millions of compounds—many microbially derived—that humans may encounter during their lifetime (Idle and Gonzalez, 2007). Fortunately, new technological advances continue to offer innovative solutions to challenges, and many resources are available to guide researchers in improving reporting quality, accuracy, and reproducibility. It is beyond the scope of this review to offer detailed technical discussion on available separation methods and instrument configurations, this topic has been thoroughly covered in recent reviews (Gika et al., 2019; Luan et al., 2019; Peisl et al., 2018; Vernocchi et al., 2016), and attention is given instead to potential applications and considerations for these techniques.

Mass Spectrometry-Based Metabolomics

MS has emerged as the leading detection method used in profiling the metabolites present in the microbiome due to its high resolution, sensitivity, speed, and ability to analyze a diverse range of metabolite classes compared with NMR. Paired separation techniques, such as GC, LC, or ion mobility, separate metabolites prior to ionization while providing retention time measurements to assist in metabolite identification. Most commonly, liquid chromatography is paired with mass spectrometry (LC-MS) to provide notable advantages in analyzing polar and semi-polar compounds (Theodoridis et al., 2012) of the microbiome. Reversed-phase chromatography, including C18 columns, excel at the separation of semi-polar compounds, while hydrophilic interaction liquid chromatography approaches have expanded the range of LC-MS to polar compounds such as sugars, amino acids, nucleotides, carboxylic acids, and phospholipids (Zeki et al., 2020). Despite improvements to LC-MS approaches, gas chromatography with mass spectrometry (GC-MS) is often preferred for the analysis of polar and volatile metabolites due to its efficiency, greater reproducibility, and more robust databases (Theodoridis et al., 2012; Zeki et al., 2020). Rather than serving as alternatives, the respective strengths of these separation techniques complement each other, and analysis using both methods may provide a more complete view of the metabolome (Gika et al., 2019; Zeki et al., 2020). However, retention time and m/z values are often insufficient for a confident feature annotation beyond a sum formula. Tandem MS (MSⁿ) is employed to provide structural information to aid feature annotation (Heiles, 2021). Most modern MSⁿ platforms rely on collision-induced dissociation to fragment mass-selected parent ions for subsequent detection (Johnson and Carlson, 2015). The fragmentation spectrum for a selected precursor ion provides a characteristic fingerprint based on fragment ion intensities. These fragment ions provide details about the structural qualities of a compound, allowing for more accurate identification with analytic standards or databases. Additional fragmentation steps may be included to differentiate the structures of similar isomers (Nash and Dunn, 2019), but fewer reference spectra are available in most databases. Structural information will remain an important metric for confident metabolite identification and may also help to interpret underlying biochemical mechanisms.

The three predominant high-resolution MS platforms (Xian et al., 2012) see frequent use in metabolomics, but do possess advantages that fit specific project needs. Time-of-flight (TOF; Allen and McWhinney, 2019) MS is based on the principle that the velocity of an accelerated ion is dependent on its mass. Ions are accelerated by high-voltage pulses and collide with a detector that reports mass spectra as a function of the time between ion acceleration and collision (Pitt, 2009). TOF analyzers provide the advantages of high acquisition speed, sensitivity, and mass resolution (Xian et al., 2012). Adding to these advantages, hybrid quadrupole-TOF (QTOF) mass analyzers combine high efficiency compound fragmentation of a quadrupole mass filter with the advantages of the TOF detector (Allen and McWhinney, 2019; Glish and Burinsky, 2008). Xue et al. (2019) leveraged the advantages of LC-MS/MS using a QTOF mass spectrometer coupled with an ultra-high pressure LC system to identify changes in microbial metabolites in the serum metabolome due to arsenic exposure. The authors investigated the impact of arsenic on a model in which mice are infected with Helicobacter trogontum to simulate a perturbed microbiome phenotype. Arsenic exposure reportedly altered 434 features for mice with a disrupted microbiome, while only 161 features changed for mice with the undisrupted microbiome. Identified features were largely involved in phospholipid, sphingolipid, fatty acid, cholesterol, and tryptophan metabolism, suggesting possible pathways in which microorganisms influence the host’s response to arsenic (Xue et al., 2019). Using MS, the authors were able to identify a potential role of the microbiome in mitigating arsenic exposure, but without supporting sequencing information are unable to identify relevant microbiota. While TOF-based instruments are disadvantaged by their low resolving power, relative to other high-resolution platforms, Fourier transformation-ion cyclotron resonance (FT-ICR) is distinguished by providing the highest resolution of all mass analyzers. FT-ICR measures the cyclotron frequency of ions trapped in a fixed magnetic field and uses Fourier transformations to convert frequency to m/z (Ghaste et al., 2016). High resolution and sensitivity allow the platform to distinguish peaks generated from complex samples with low metabolite concentrations. However, the high resolving power of FT-ICR comes at the cost of accessibility due to high financial costs, and the slowest acquisition rate of the high-resolution MS platforms (Ghaste et al., 2016). The slow acquisition speed also limits the ability to pair FT-ICR with fast separation techniques, like liquid chromatography. FT-ICR is often used to address specific challenges in distinguishing features, rather than as a standard instrument for metabolomics (Ghaste et al., 2016; Maier et al., 2017; Mangal et al., 2020). The Orbitrap fills the gap between TOF and FT-ICR-based instruments by providing high resolution and acquisition rate at an accessible financial cost. The Orbitrap analyzer belongs to the family of Fourier transformation MS platforms which provides spectra based on Fourier transformations of digitized image currents. Unlike FT-ICR, Orbitrap instruments rely on electrostatic fields, rather than magnetic fields, to trap ions in an orbit around a spindle-shaped electrode and an image current is recorded as the ion oscillates between the ends of the spindle (Hohenester et al., 2020). Orbitrap instruments, like TOF instruments, pair well with multiple separation techniques, offering additional flexibility relative to FT-ICR.

As with other techniques, the chromatography method and mass analyzer should fit the scope of the study, metabolites of interest, and resolution required. Although the most comprehensive studies would include multiple methods of data acquisition to compensate for the disadvantages of instrument configuration. This creates another challenge for metabolomics as combining multiple instrument configurations for a thorough survey of the metabolome reduces throughput and increases costs. As an example of his approach, Gomez et al. (2021) used both GC-MS and LC-MS/MS to investigate the effects on the microbiome of early life exposure to three environmental pollutants, BDE-47, tetrabromobisphenol, and bisphenol S. GC-MS-based metabolomics was used to quantify differences in short chain fatty acid production due to perturbations of the microbiome. The authors reported that exposure to bisphenol S diminished acetate, which was associated with changes in four bacterial genera, and may help to explain the chronic inflammation caused by bisphenol S. Furthermore, the authors leveraged the advantages of LC-MS/MS to quantify fecal total, secondary, and unconjugated bile acids, to validate sequencing-based predictions of increased bile acid production. Early life exposure to these three toxicants reportedly increased concentrations of each type of bile acid. Secondary unconjugated bile acids are known to be produced by gut microbiota and may act to suppress immune responses (Gomez et al., 2021). The analytical range provided by the two MS approaches supported bioinformatics-based predictions indicating that toxicant exposure produced long-term changes to the gut metabolome.

Imaging Mass Spectrometry

In addition to global metabolite profiling, MS is gaining popularity as a tool for spatially tracking specific metabolites through imaging mass spectrometry. The ionization of molecules, frequently by Matrix-Assisted Laser Desorption-Ionization, Secondary Ion MS or Desorption Electrospray Ionization, in a designated grid provides a mass spectrum at each pixel of the grid. This data are then used to reconstruct an image displaying the locations of detected features. Imaging mass spectrometry could be used to track microbial metabolic products within tissues to provide spatial insights into metabolic mechanisms, as is being done in pharmaceutical research (Schulz et al., 2019). Furthermore, analysis of a sample in consecutive sections can be assembled into a 3D representation (Vos et al., 2021), potentially allowing for investigations of micro-niches (Eckstein et al., 2020) within the microbiome. The different technical considerations and advantages of imaging mass spectrometry methods have been recently reviewed (Dong et al., 2016; Gilmore et al., 2019). Despite applications in plant biology (Maloof et al., 2020), pharmacology (Cornett et al., 2008), and cell biology (Brockmann et al., 2021; Pareek et al., 2020), this technique has yet to make notable contributions to microbiome research. However, future advances and novel methods may provide the ability to study the movement of toxic compounds in complex samples to make sense of interactions between toxin, host, and microbiota.

Nuclear Magnetic Resonance-Based Metabolomics

While MS will likely remain the predominant method to conduct metabolomics studies, NMR maintains advantages that are worth considering. NMR, primarily in the form of ¹H NMR, remains a highly reproducible, quantitative metabolomics technique with simpler sample preparation and data analysis. It is commonly accepted that NMR is a highly reproducible and repeatable technique (Dumas et al., 2006), and in part this has helped to generate large, reliable reference databases, including Human Metabolome Database (Wishart et al., 2018), Biological Magnetic Resonance Bank (Ulrich et al., 2008), and Birmingham Metabolite Library-NMR (Ludwig et al., 2012). NMR also excels at absolute quantification of metabolites due to direct proportionality of signal intensities to compound concentration when internal standards are included (Smolinska et al., 2012). While it is possible to apply NMR to untargeted metabolomics experiments, the underlying goal of measuring and identifying as many compounds as possible is hindered by the limited number of features that can be detected by NMR. However, while the relatively low number of features provides an incomplete picture of the metabolome, fewer features reduce the time and complexity involved in processing and analyzing NMR data. This may prove advantageous for certain applications or studies. A study conducted by Zhang et al. (2015) into the effects of tricholoroacetamide, a disinfection byproduct found in drinking water, on the microbiome provides an example for NMR use in microbiome research. ¹H NMR was used to evaluate changes to urine metabolic profiles of mice exposed to trichloroacetamide. Twenty structurally diverse metabolites were reported to be altered, many of which are thought to be produced by gut bacteria, including short chain fatty acids, indole derivatives, and choline metabolites. This data were paired with metagenomics data indicating changes to microbial composition, energy production, and amino acid metabolism. The authors conclude that trichloroacetamide exposure alters gut microbiota which in turn alters the host metabolome with potential negative consequences on host intestinal inflammation (Zhang et al., 2015). Despite technological developments reducing the deficiencies of NMR (Bingol, 2018), many future advances will depend on hybrid applications alongside MS. As an example, a technique termed SUMMIT MS/NMR (Bingol et al., 2015) uses accurate masses generated by high-resolution MS to predict potential structures and validation by experimental chemical shifts from 2D NMR to provide confident metabolite identification. Hybrid MS/NMR has served as a powerful tool for identifying novel metabolites (Bingol, 2018; Boiteau et al., 2018; Wang et al., 2017) and may help to fill in gaps left by any single technology.

Metabolomics Data Processing

Many of the greatest impediments to metabolomics is the need to process and analyze large data sets containing thousands of features—an impractical task with manual annotation and curation. Thorough data processing provides more accurate analysis by improving signal quality, reducing bias, and removing noise. Both MS- and NMR-based spectra undergo baseline correction to remove low frequency artifacts and high-frequency noise generated from the measurement instrument may be filtered out (Alonso et al., 2015). Peaks are chosen by algorithms (Tautenhahn et al., 2008) designed to smooth spectra and apply parameter thresholds to identify different peaks (Alonso et al., 2015). Spectra must then undergo spectral alignment to correct for shifts in peak position that may be caused by a sample’s chemical environment or differences in chromatography retention time (Smith et al., 2006). There are many commercial and freely available software to navigate steps, but popular (Spicer et al., 2017) choices are listed in Table 1. Some software options connect multiple tools into a workflow, allowing for pre-processing to feature annotation to statistical analysis in a single package. This is well developed for LC-MS data and implemented in popular workflows through Galaxy (Workflow4metabolomics or Galaxy-M), XCMS Online (Tautenhahn et al., 2012), MetaboAnalyst 5.0 (Pang et al., 2021), Metabolomics Analysis Visualization Engine (MAVEN; Clasquin et al., 2012), MS-DIAL 4 (Tsugawa et al., 2020), and MZmine2 (Pluskal et al., 2010). The assimilation of tools into a single workflow increases the ease of analyzing data while improving reproducibility of data processing. There is not a widely used complete workflow for NMR data, possibly due to frequent use of commercial software for data analysis. Even with access to open-source software and the maturity of some of these tools, data processing with different software may lead to extremely varied results. A recent study compared LC-MS data processing outputs from enviMass, MZmine2, Compound Discoverer, and XCMS Online and found only 10% overlap of features between the programs (Hohrenk et al., 2020). While the past 20 years have seen extraordinary developments in metabolomics because of advances in analytical platforms, future advances are needed in the form of expanded databases, new technologies to improve data processing, and thorough validation of new tools.

Metabolomics data processing and analysis is rapidly evolving as an area of bioinformatics and new software is regularly developed to address the field’s needs. The most current computational developments were recently reviewed (Misra, 2021), but examples will be listed to outline the most important areas in need of improved tools. A focus for metabolomics software development is on improving feature annotation for compounds not listed in available databases. A newly published tool, CANOPUS (Dührkop et al., 2021), uses fragmentation spectra generated from high-resolution LC-MS and neural networks to predict the compound classes of queried compounds. Compound classes can provide biological insights or narrow the scope of investigation without the need for complete annotations that require database entries or analytical standards. Similarly, Retip (Bonini et al., 2020) uses a library of experimentally derived LC retention times and multiple machine learning (ML) models to predict the retention time of compounds without database entries. In combination with MS/MS spectra, the predicted retention time may improve confidence in a feature annotation, although it does not replace the confidence gained experimentally derived measurements. While the vast majority of software is focused on LC-MS data processing (82% of tools published in 2020; Misra, 2021), the recently published SMART 2.0 (Reher et al., 2020) uses two dimensional NMR data to predict structures, cosine similarities, SMILES, and molecular weights. The tool was used to identify a cytotoxin from a complex extract generated from cyanobacteria, aiding subsequent isolation and characterization of a novel compound (Reher et al., 2020). The development of SMART 2.0 demonstrates potential applications and continued interest in NMR for conducting metabolomics. While new and developing software offer innovative solutions to standing problems in metabolomics, rigorous evaluations of the performance and reproducibility of these tools are needed prior to wide adoption.

Challenges for Metabolomics

Advancements in analytical platforms have enabled more detailed and expansive metabolomics studies, however, the capabilities of these platforms are limited by the ability to identify detected metabolites (Uppal et al., 2016). Commonly a period develops between the development of technology that generates large amounts of data and the organization of that data for future use. Metabolomics is currently in this period where data are generated in large quantities but is not widely available and databases are relatively sparse. As a comparison to sequencing, NCBI’s Sequence Read Archive (Leinonen et al., 2011) contains 333 290 studies and 1 681 081 individual bacterial samples, while the Metabolomics Workbench (Sud et al., 2016) contains 667 studies relating to humans, including non-microbiome studies, and HMDB contains approximately 23 000 detected metabolites at the time of writing of this review. Future contributions to and funding of public data repositories and databases will significantly ease the difficulty further developing databases and software to automate data processing. The importance of these contributions cannot be understated.

Reproducibility is a major challenge for metabolomics studies, contributing to the broader challenge of reproducibility within microbiome research. Errors in sample handling, labor-intensive sample preparation, intrinsic differences between instruments, and feature annotation are major sources of variation which may lead different research groups to different results. Some of these issues can only be addressed with diligence and accurate reporting of methods, while others have more robust solutions that should be put into practice. In the case of feature identification, the difficulty lies in the reality that the vast majority of the millions of estimated metabolites that a human may come into contact with are unknown (Idle and Gonzalez, 2007) and even fewer compounds are available as analytical standards (Liu et al., 2020). Consequently, a reader may have legitimate reasons to doubt the confidence of a putative feature annotation if they are not provided with feature properties (chromatography retention time, m/z, mass fragmentation pattern, or NMR chemical shift) in a publication. Reporting standards, like those outlined by the Metabolomics Standards Initiative (Sumner et al., 2007), provide guidelines for defining and reporting metabolites. Metabolomics Standards Initiative suggests four levels of confidence with each subsequent level requiring more rigorous experimental and analytical evidence, providing a framework to can be used to bolster reader confidence in metabolite identification. Critiques (Creek et al., 2014; Malinowska and Viant, 2019) and calls for updates (Spicer et al., 2017) made been made regarding the Metabolomics Standards Initiative, but at the present, it is the most widely known initiative to report confidence in identifying metabolites and should not be lightly dismissed. Unlike other challenges facing metabolomics studies, reproducibility will not be addressed solely with technological advancements, but must be addressed with diligence, quality controls, and commitment from investigators.

The greatest impediments to future advances in metabolomics can be removed through the wide adoption of data sharing. Funding agencies, publishers, and the community (Haug et al., 2017; Rocca-Serra et al., 2015) have all made calls to support the development of the field by making raw and processed data available. Public availability of data sets expands databases, provides resources for software validation, enables new studies, and bolsters reproducibility. Several data sharing initiatives and repositories have been created to address these calls, and have subsequently recognized their position as central hubs of metabolomics research. As central hubs, they have expanded their aims to include the training of new researchers and to foster scientific collaboration. The Metabolomics Workbench is a National Institutes of Health sponsored repository for promoting international data sharing, and for facilitating future development of the field by providing a platform to host training materials, protocols, and analytical tools. Researchers deposit or access raw and processed metabolomics data collected with either MS or NMR along with accompanying metadata. Depositors are encouraged to include detailed protocols with instrument parameters, sample collection, and extraction methods alongside data, increasing study transparency and providing a methodology resource for researchers to draw upon. MetaboLights (Haug et al., 2020) is another leading platform of international metabolomics data sharing maintained by the European Bioinformatics Institute (EMBL-EBI) with similar goals to Metabolomics Workbench. The repository holds data submitted from MS and NMR-based metabolomics studies and undergoes manual curation to ensure complete and high-quality submissions prior to becoming publicly available. MetaboLights also holds a goal of providing training resources for new or experienced members of the metabolomics community, including lectures and workshops. Global Natural Products Social Molecular Networking (Wang et al., 2016) is another public tandem MS data repository and spectral library but seeks to serve the field with additional unique capabilities. Central to Global Natural Products Social Molecular Networking is the idea that the rapidly increasing amount of MS/MS spectra continuously improves the ability to identify metabolites and it can maintain a “living” repository by automatically reanalyzing public data sets. Users can subscribe to specific data sets and will be alerted when changes are made, including new or changed metabolite identifications. This works to maintain the relevance of data sets over time as new metabolite identification may offer new insights or routes of investigation. The Global Natural Products Social Molecular Networking repository is complemented by a suite of analysis tools, including molecular networking tools for visualizing relationships between compounds. Furthermore, the platform offers the ability to view other researchers who are tracking changes to specific data sets with the intended goal of recognizing other community members that are interested in similar research questions. These connections offer a possible avenue for developing collaborative relationships within the field of metabolomics. The data, resources, and tools provided by these public repositories demonstrate the value of data sharing. Collaboration is a fundamental principle of scientific research and contributions of data to public repositories advance the field and the researchers within it.

METAPROTEOMICS

Although the microbiome is predominantly studied using sequencing and metabolomics, metaproteomics is another valuable resource which was not highlighted in this review. Metaproteomics offers the ability to fill gaps in knowledge that these other two techniques cannot directly address. Researchers are currently using metaproteomics to profile total protein compositions of microbiome samples in a culture-independent method. Thorough reviews of metaproteomics have been recently published (Sajulga et al., 2020; Salvato et al., 2021; Zhang and Figeys, 2019), but a summary will be provided to highlight another forward-looking experimental approach that is providing novel insights into the microbiome.

Metaproteomics uses high-resolution mass spectrometry to identify proteins present in a sample of the microbiome. A protein profile of the microbiome can be used to provide analysis of taxonomy, function, and metabolic pathways. Yet, it is a technique often understood when employed alongside other culture-independent techniques, like sequencing and metabolomics, and is largely dependent on these techniques for functional and taxonomic analysis. For example, protein profiles can be used to taxonomically identify bacteria present in a sample, but successful approaches often depend on using metagenomics to provide reference protein sequences or to narrow the scope of database searches (Tanca et al., 2016). Mills et al. (2019) used a strategy of integrating shotgun metagenomics with metaproteomics to investigate the relationship between genes and protein in a patient with Crohn’s disease. The authors report comparable taxonomic analysis between both techniques, while metaproteomics identified functional associations that were absent in the metagenomic data. Similarly, metaproteomics data can be used for functional analysis, however, this is challenged by deficient gene annotations and functional characterizations of proteins (Heintz-Buschart and Wilmes, 2018). The integration of metabolomics with metaproteomics may provide additional interpretations of how microbes create a metabolome that benefits or detracts from host health. Ke et al. (2019) treated mice fed a high-fat diet with strains of Bifidobacterium animalis and Lactobacillus paracasei and used metabolomics and metaproteomics to assess the effect of these bacteria on weight gain. The treatments reduced mouse body weight gain and alleviated metrics of metabolic disease. Analysis of metabolomics data identified the production of short-chain fatty acid and reduction of bile acid pools as the result of the treatments, while metaproteomics identified alterations in carbohydrate, amino acid, and energy metabolism as the result. When integrated, these techniques provide a more expansive investigation into perturbations of the microbiome and offer new avenues of research.

Metaproteomics offers a means to gather data that is distinct from sequencing or metabolomics techniques, while integration with these other techniques may work to explain the link between genome and metabolome. However, this field has not matured to the same degree of sequencing and metabolomics. Available databases contain less reference material and fewer datasets, and the available range of software resources is more limited. With its current limitations, metaproteomics does not match the accessibility of sequencing or metabolomics. Yet, the application of current metaproteomic techniques to specific research questions and future technological developments are worth consideration.

INTEGRATING TECHNIQUES

Linking the taxa and metabolites of the microbiome is an area of intense interest and great challenge. Profiling the metabolome may identify compounds responsible for host responses, but it may be difficult to alter the metabolome without altering the microbiota. Reference genome databases are not exhaustive, genome annotations are incomplete, and metabolic interactions may obfuscate the production of target metabolites. These limitations led early studies of the microbiome to often conclude with analysis and correlations but did not offer thorough investigations of causes or mechanisms. As technologies and methods improve, there are expectations that the complex interactions of the microbiome can be analyzed using statistical methods for the purpose of developing testable hypotheses. These hypotheses are assessed in simplified animal and cell models, firmly rooting the field in traditional scientific practice.

Statistical Analysis

Statistical predictions do not demonstrate the causes of biological mechanisms, but they are essential for narrowing the scope of experimental investigation. Testable hypotheses are generated from multi-variate statistics which can effectively integrate metabolomics, sequencing, and other data sets. However, integrating these data sets is confounded by the inherently high-dimensionality (Hongzhe, 2015), sparsity (Jiang et al., 2019), and compositionality (Combettes and Müller, 2021; Gloor et al., 2017) of the data. Metabolomics and sequencing generate high-dimensional data sets with hundreds or thousands of features, far surpassing the sample size, and making it difficult to describe the relationship between predictor variables and response variables. Excessive zeros and high frequencies of low observations of taxa—described as sparse data—produce a heavily skewed distribution which contributes to poor model fit (Jiang et al., 2019). Furthermore, Illumina sequencing generates compositional data due to the design of the technology. The total sequence count for a sample, referred to as sequencing depth or library size, is based on the fixed capacity of the sequencing instrument and not on biological considerations (Jiang et al., 2019; Quinn et al., 2018). This means that sequence reads are proportionally related to each other, and statistical methods that do not consider proportionality produce high false discovery rates (Quinn et al., 2018). Therefore, methods designed for compositional data, commonly referred to as compositional data analysis (CoDA) methods, should be used for analysis (Combettes and Müller, 2021; Gloor et al., 2016, 2017; Jiang et al., 2019). The compositional nature of sequencing data is often underappreciated or misunderstood. Many published studies employed statistical models that do not take compositionality into consideration, despite the availability and ease of applying compositional data analysis alternatives. For example, is has been long understood that Pearson correlations are prone to false positive correlations when applied to compositional data (Aitchison, 1982; Friedman and Alm, 2012), however, it is still commonly used. SparCC (Friedman and Alm, 2012) and SpiecEasi (Kurtz et al., 2015) are accepted alternatives for generating correlations within compositional data. As another example, common distance or dissimilarity matrices used for ordination and clustering, such as UniFrac, Bray-Curtis, and Jenson-Shannon divergence, do not consider compositionality without proper transformation of the data. PhILR (Silverman et al., 2017) and Aitchison distance (Aitchison et al., 2000) provide Log-Ratio transformations that allow for accurate analysis. Additionally, attempts to correct for compositionality and differences in sequencing depth by random sub-sampling without replacement, known as rarefying or rarefaction, reduces precision (McMurdie and Holmes, 2014) and introduces additional biases (Willis, 2019). Comprehensive lists of alternative data transformations, distance matrices, and statistical techniques have been recently published (Gloor et al., 2016; Hongzhe, 2015; Jiang et al., 2019).

Successful models for integrating metabolomics and genomics data sets will effectively address the characteristics of the data to accurately predict relationships between microbes, metabolites, and host responses. Due to the highly variable nature of microbiome data, there is not a single model that will always produce accurate results when applied to microbiome data. Therefore, it is beneficial to apply and validate different models to find which best fits the data, rather than force data to fit a model. Linear regression algorithms have been designed to consider the constraints of microbiome data (Bodein et al., 2019; Combettes and Müller, 2021), and provides the unique advantage of being widely interpretable by researchers in most disciplines of biology and chemistry. Alternatively, sparse Partial Least Squares (sPLS) regression serves as the basis of the popular (Barber et al., 2021; Jones et al., 2021; Rosario et al., 2021) R package, mixOmics (Rohart et al., 2017) and is used to discriminate between sample groups and identify the discriminating features. mixOmics was designed to provide robust models capable of integrating thousands of features from sparse, high-dimensional data sets generated with genomics and metabolomics techniques. A notable drawback, however, is the reliance on linear combinations of variables despite the likelihood of some non-linear relationships.

ML algorithms are becoming a leading tool for identifying patterns and relationships among features of the microbiome. Many of these models take advantage of the size of available data sets by pairing traditional statistical techniques with algorithms designed to adapt models and improve performance. ML algorithms are divided into two categories: unsupervised and supervised. Unsupervised ML works to identify structures and patterns with data, commonly through dimensional reduction. Dimensional reduction is the principle behind the widely used principal component analysis and principal coordinate analysis (Namkung, 2020), to simplify the high-dimensional data of microbiome research into an easily visualized form. Advances in supervised learning are largely responsible for increased interest in ML because of its ability to predict relationships between microbiome features and host traits. Unlike traditional regression models which are designed as explanatory, the goal of supervised learning is to build models from a set of categorized data before predicting the category of unlabeled data (Knights et al., 2011). Commonly used supervised learning techniques for microbiome research have been recently reviewed (Moreno-Indias et al., 2021; Namkung, 2020; Zhou and Gallins, 2019), but random forest (Asnicar et al., 2021; Chen and Ishwaran, 2012) will be highlighted due to its popularity, ease of application, and flexibility. Random forest is a non-linear model commonly applied to microbiome research to predict important features for a specific host response. Random forest is based on the construction of hundreds to thousands of decision trees with classifiers based on a random selection, with replacement, of samples (bootstrapping) and aggregating multiple trees by averaging predictive values (bagging) (Chen and Ishwaran, 2012; Namkung, 2020). The algorithm then randomly selects a subset of variables to create a classification tree to use for predictions. Like all data driven techniques, random forest models benefit from large sample sizes because of the high-dimensionality of the data and the model must be trained and cross-validated before its application, often resulting in poor performance for small (n < 50) sample sizes (Topçuoğlu et al., 2020). In part, random forest is popular due to the ease of applying it to microbiome data through freely available R or Python libraries (Chen and Ishwaran, 2012; Zhou and Gallins, 2019), and its resistance to overfitting. There are, however, drawbacks with this algorithm as it commonly struggles with data sparsity and the noise generated from many irrelevant features (Knights et al., 2011). It is also not inherently interpretable, rather, it provides a ranking of importance variables that influence the response trait (Chen and Ishwaran, 2012). Like many cutting-edge techniques, reporting standards for ML models are lax and lack transparency, validation, and detailed methodology. These problems manifest in models that serve as black boxes which cannot be interpreted by those outside of it. Researchers should explain why a specific model was chosen, the parameters used, and demonstrate the performance of the model (Namkung, 2020; Topçuoğlu et al., 2020; Zhou and Gallins, 2019). As the field matures, models and their interpretations will become more complex, transparency will remain crucial to conveying results and opening the black box to other researchers.

Further development of ML techniques may address some of the current weaknesses with available algorithms. Deep learning is a subfield within ML and is driven by the design of deep neural networks (Eraslan et al., 2019; Wainberg et al., 2018). Deep neural networks are constructed networks of successively interconnected nodes, called artificial neurons. The relatively simple computational output of one layer of nodes is fed into a subsequent layer, allowing for increasingly complex computation through each subsequent layer. This allows for the modeling of outcomes with complex dependencies. The practical result of this technique is a non-linear model capable of processing high-dimensional, spare data sets with less bias in model design (Zhu et al., 2020). While early in its application to microbiome research, multiple tools are available (Oh and Zhang, 2020; Sharma and Xu, 2021), and deep neural networks have been applied in studies of the microbiome (Benjamino et al., 2018). Regardless of the advances made to ML algorithms, these models do not undertake the mechanistic hypothesis testing that biological research is based on (Breiman, 2001; Eraslan et al., 2019). It can, however, reduce the scope of experimentation from unworkable to practical.

Validating Hypotheses

This emphasizes the need for an orthogonal technique to both add confidence and further explore a mechanism—metabolomics provides a complement to metagenomics. Metabolomics can provide analysis of samples from germ-free mice colonized with a single bacterial strain to provide evidence that the metabolic predictions from sequencing data hold true in biologically relevant experimental conditions. There is, however, a developing appreciation for the need for a more complex model system. A gut microbiota consisting of a single bacterial strain is not reflective of the complex metabolic interactions produced by hundreds of taxa that span multiple kingdoms and phlya. Attempts are being made to produce more robust models by incorporating strains of interest into synthetic communities prior to inoculating germ-free mice. The inclusion of other organisms may demonstrate that competitive interactions, nutrient availability, or microbial cross-feeding will not abrogate the production of specific metabolites. A consensus has not been reached on a standardized synthetic community composition, although Altered Schaedler Flora (Biggs et al., 2017; Lyte et al., 2019) have been used for decades, and commercial products are available. This synthetic community is criticized as an overly simplified metabolic representation of the gut microbiome (Rohde et al., 2007), but an alternative composition has not been widely accepted. The lack of an accepted standardized synthetic community with defined strains prevents the comparison between studies and hinders reproducibility. Regardless, the use of germ-free mice is currently essential to demonstrating a causative link between specific taxa, production of metabolites, and a host response.

CONCLUSIONS AND FUTURE PERSPECTIVES

To progress the field of microbiome research, complex interactions of biological and biochemical interactions need to be dramatically simplified, however, as the field has developed it has become clear that many of these simplifications have removed important nuance from discussions. These convenient simplifications affect everything from data collection through analysis. For example, it is often ignored that sequencing runs are variable—due to the nature of random sampling—and replicate runs may reveal both zero and non-zero counts (Marioni et al., 2008), yet it is rare for technical sequencing replicates to be included. This is an unavoidable product of the expense that would come from sequencing multiple technical replicates for each experimental sample, but when combined with the technical limitations of sequencing, including sample preparation, platforms, and data processing, it is difficult to make confident conclusions based on sequencing data alone. Other seemingly small technical details are often overlooked, but can affect the results of experiments. For example, the time of day when samples collected from animals may affect metabolite concentrations and microbiome composition (Jones et al., 2021). This concern logically follows what is known about digestion and nutrient absorption but may be ignored during study design or not reported in the methods of publications. The aggregate of these details (metadata), including subject information, sample collection methods, and data handling, contextualizes the data from a sample and provides essential information required to evaluate a study. The collection and publishing of complete metadata is fundamental to reproducing and analyzing a study (Huttenhower et al., 2014; Langille et al., 2018; Ryan et al., 2021; Vangay et al., 2021).

The importance of data sharing initiatives in the fields of metabolomics and metagenomics were discussed in this review, but the need for generalized data sharing principles and standards should also be recognized. Generalized principles are intended to promote data sharing across research disciplines and to allow for easier integration or analysis of data sets. The implementation of standards by large organizations is a direct approach to ensuring these standards are adopted by researchers. Organizations, like EMBL-EBI, publishers, like Springer Nature Journals and PLOS, and funding agencies, like the NIH and NSF, have adopted the FAIR (Wilkinson et al., 2016) principles for scientific data management. FAIR emphasizes four principles, Findability, Accessibility, Interoperability, and Reusability, to provide guidance for ensuring that published data are transparent, reproducible, and reusable. It is intended that these principles will reduce the difficulty in accessing, contextualizing, and using published data without direct communication with the original investigators. Data management projects (Martini et al., 2022; Paini et al., 2022; Watford et al., 2019; Wood-Charlson et al., 2020) in both fields are presently making commitments to the FAIR principles and researchers should consider these principles and others to ensure that their data remains accessible to facilitate new analyses.

It is common for microbiome studies, including those relevant to toxicology, to simplify the entire microbiota down to the role of a single bacteria, its ability to metabolize a xenobiotic, and the effect on mice or cultured cells. These studies have provided invaluable insights, but often they address what could happen rather than what happens in biologically relevant conditions. In many instances, these studies disregard interspecies genetic diversity and lack investigation into the universality or stability of genetic elements with the bacterial species of interest. It is well known that genetic instability is essential to bacterial evolution and sequencing technologies have provided insights into the genetic diversity present within species (Ellegaard and Engel, 2016; Pasolli et al., 2019), and into how genomes within a species change over time within a host (Caporaso et al., 2011; Chen et al., 2021). Additionally, conditions in a micro-niche or interactions with other organisms may influence the metabolic potential of any one species (Coyte and Rakoff-Nahoum, 2019). Introducing a single strain to a germ-free mouse reflects, in many ways, in vitro experiments. Both approaches generate useful data by providing a simplified model that is practical to apply, however, the explanatory potential of these methods is limited by their artificial conditions. Moreover, these conditions are often limited to the investigation of bacteria, despite the growing evidence of the contributions that eukaryotes (Beghini et al., 2017; Pasolli et al., 2019) and viruses (Gregory et al., 2020; Liang and Bushman, 2021; Shkoporov et al., 2019) make to the microbiome. Finally, perturbations of single organisms, or the entire microbial community, due to toxic exposure are insufficient to suggest an impact on the host. The technologies described in this review alongside animal models make it possible to establish this link, for example, through fecal transfers to germ-free mice (Ridaura et al., 2013). Thus, it is critical to design studies that will address whether the observed change to the microbiota significantly influence the host. This remains a challenging task and there is a need for new technologies to further address this important issue.

Lastly, every toxicology study, including those focused on the microbiome, should consider one of the field’s most fundamental adages, “the dose makes the poison.” Relevant conclusions can only come from evaluating realistic doses delivered in a way that considers differences in acute compared with long-term exposures. Regular exposures may apply selective pressures that shape the microbiome over time (Lozano et al., 2018; Shao and Zhu, 2020). In the same way that size of a dose may alter outcomes, so too can the size of the microbiota. Microbiota are often analyzed in terms of relative abundance or absolute read counts, but neither of these metrics provide an absolute quantification of bacteria within the microbiome (Shanahan and Hill, 2019). It is not possible to accurately compare microbiota or determine the cause of a change to the metabolome without quantifying the present organisms. While there is ongoing debate regarding the most accurate method for quantifying microbiota (Galazzo et al., 2020; Jian et al., 2021), quantitative PCR remains an accepted and accessible method for this purpose (Jian et al., 2020). These details need to be addressed to enable rigorous investigation of the complex interactions between host, microbiota, and xenobiotics. Like ongoing efforts to move microbiome studies beyond taxonomic associations, future work should not shy away from addressing these complexities.

FUNDING

This work was supported by the National Institutes of Health grants; R01 ES028288 (A.D.P.), U01 DK119702 (A.D.P.), and R35ES028244 (G.H.P.). This work was also supported by the USDA National Institute of Food and Federal Appropriations under Project PEN047702 and Accession number 1009993.

CONFLICT OF INTEREST

The authors declare they have no conflicts of interest.