Tools for Analysis of the Microbiome

Abstract

Over the past decade, it has become exceedingly clear that the microbiome is a critical factor in human health and disease, and thus should be investigated to develop innovative treatment strategies. The field of metagenomics has come a long way in leveraging the advances of next generation sequencing technologies resulting in the capability to identify and quantify all microorganisms present in human specimens. However, the field of metagenomics if still in its infancy, specifically in regards to the limitations in computational analysis, statistical assessments, standardization, and validation due to vast variability in the cohorts themselves, experimental design, and bioinformatic workflows. This review summarizes the methods, technologies, computational tools and model systems for characterizing and studying the microbiome. We also discuss important considerations investigators must make when interrogating the involvement of the microbiome in health and disease in order to establish robust results and mechanistic insights before moving into therapeutic design and intervention.

Introduction

What is the microbiome and why is it important?

As modern microbiology and the development of next generation sequencing technologies have matured, research has increasingly focused on the complex microbial communities that interact with the host to influence disease processes ^[1]. Recent studies have focused on the composition of the communities, or the “microbiota”, which is a group of microscopic organisms in any given region, whether it be a specific body site, habitat, or ecosystem ^[2]. Other investigators have taken this a step further and expanded their exploration to the “microbiome” which is the collective genomes and gene products of the microbiota residing within a host or environment ^[3]. Classically, however, these terms have been used interchangeably. These investigations have facilitated the creation of the field of metagenomics, which is the examination of all genetic material recovered directly from environmental or living samples ^[4]. Metagenomic based experimentation has enabled the ability to bypass in vitro studies or cultivation in order to characterize the composition and functionality of entire microbial communities ^[5]. The field of metagenomics and microbiome analyses have received growing scientific attention due to large research consortiums such as the Human Microbiome Project (HMP) ^[3], the European MetaHIT ^[6], and the Integrative Human Microbiome Project (iHMP) ^[7].

Classic infectious disease and pathology have historically focused on one pathogen equating to one disease ^[8]. However, we now understand that imbalances in the microbiota are also associated with many different disease states. Often, reduction in microbial diversity and outgrowth of specific species can induce negative effects like inflammation or infection ^[9]. Reports have shown evidence for the potential involvement of the microbiome in almost all walks of health-related complications to include obesity ^[10], diabetes ^[11], cardiovascular disease ^[12], oncogenesis ^[13], cancer therapy response ^[14], Alzheimer’s ^[15], and other neurodegenerative conditions ^[16].

The focus of this review is to describe what “tools” are at our disposal if one wants to begin to characterize the impact of the microbiome on evolution of a disease, or the effect of a disease or intervention on the microbiome. One must consider, however, that we are in the infancy of metagenomics. Microbiome studies produce vast amounts of data which necessitate sophisticated computational tools and the technologies utilized are rapidly evolving. Additionally, many of the mathematical tools available provide assessments of association, but not causation. Thus, investigators need to exercise caution that microbiome characterization, data analyses, and modeling is only a small piece of the discovery process and should be used to complement classic in vitro techniques and in vivo model experimentation in order to establish cause and effect.

Sequencing Methods, Technological Advances and Bioinformatic Tools for Studying the Microbiome

There are a number of different technologies available to study the microbiome. For this overview, we will first discuss the fundamental methods, and then the primary technologies available for use with each method (Figure 1).

Figure 1.

Technologies for studying the microbiome

Marker Gene Analysis

Targeted sequencing methods are the most common methods used and encompass 16S ribosomal RNA gene sequencing for bacteria, and internal transcribed spacer (ITS) region sequencing for fungi, among other, less common target genes. While highly conserved, these genes have diverged over time and provide a unique barcode that can be assigned to specific taxonomies and can also be counted to identify the frequency of each member within the microbial community. One challenge that arises in the analyses of these marker genes is defining and identifying a unique sequence. Both the 16S gene and ITS genes contain hypervariable regions that can change over a short period of time and can even differ within a single cell. Additionally, individual taxa can share identical gene sequences (e.g. Escherichia and Shigella) complicating the definition of a unique sequence and differentiation of individual taxa ^[17]. Moreover, each individual gene is amplified with PCR and each PCR product is sequenced a single time, leading to multiple opportunities for errors that can impact the identification of a unique sequence ^[18]. This has led to the adoption of operational taxonomic units (OTUs), a method for binning sequences using a threshold of divergence (most often 97% or 99%) ^[19–21]. The choice of this threshold is often arbitrary as these cutoffs do not match biologically relevant cutoffs and can change based upon the choice of variable regions sequenced ^[22]. This OTU binning allows one to control for biological and technical variations and each OTU can be treated as a distinct category that can be analyzed with or without taxonomic information ^[23]. Typically, however, data is analyzed using taxonomic information. Taxonomy can be assigned via machine learning methods such as the RDP classifier ^[24], a naive Bayesian classifier, or mapping to popular reference databases, such as Greengenes ^[25] and SILVA ^[26]. Standard microbiome analysis packages, such as Mothur ^[27], QIIME ^[28], and DADA2 ^[29] provide an interface for taxonomic assignment.

Current 16S and ITS sequencing relies most frequently on the sequencing length allowed by the Illumina MiSeq (Illumina, San Diego, CA), frequently leveraging the 2×300 sequencing length to cover as many variable regions as possible or gain the highest accuracy possible. This sequencing method utilizes region specific primers, such as V1-V3 or V4, to sequence from both the forward primer and the reverse primer, generating a full amplicon for each taxa within the microbial community ^{[30, 31]}. There is a current push to move to full length 16S and ITS sequencing methods with either PacBio or Oxford Nanopore sequencing technologies, but the databases available for each marker gene lag behind the technological advances and limit the utilities of these methods ^[32].

Shotgun Metagenomics

Marker gene sequencing methodologies have revealed incredible insights into the role of the microbiome in health and disease, but only focus on a small portion of microbial genomes. Shotgun metagenomics approaches consist of methods that utilize untargeted sequencing methods to capture all microbial genomes present within a sample ^{[33, 34]}. The capture of the full repertoire of genetic information from a microbiome sample allows the study of bacteria, fungi, DNA viruses, and other microbes but is dependent on reference genomes and scientific knowledge ^[35–40]. This information can be used similarly to amplicon sequence-based methods to identify which taxa are present and the relative abundance of each ^{[41, 42]}, or to identify gene coding sequences and mine the functional potential of the microbial community ^[43].

Thus far there is no consensus regarding the best assembly methods for shotgun metagenomic data. Metagenomic shotgun assemblies are either performed de novo, based on reference genomes, or a hybrid of both. Much like whole genome sequencing, short read data can be joined via overlapping reads and used to construct contigs (Overlap, Layout, Consensus assembly), however, the computational demands of this method are impractical ^[44]. Thus, many de novo assembly methods instead use the de Bruijn graph approach to assemble sequencing reads ^[45]. Commonly used software tools for de novo assembly include MetaVelvet ^[46], IDBAUD ^[47], metaSPAdes ^[48], and MEGAHIT ^[49]. In reference-guided metagenomic assembly, such as MetaCompass, sequencing reads are mapped against reference databases in order to reconstruct contigs ^[44]. As stated above, reference-based assemblers are restricted by the availability of reference genomes and quality of the database.

Often researchers perform read-based profiling of representative or discriminative genes (or markers) from unassembled shotgun metagenomics reads and compare those to reference databases in order to assign taxonomy or annotate genes. Taxonomic binning can utilize similar DNA compositions or nucleotide patterns like k-mer lengths, GC content, or gene homology ^[50]. Kraken, for example, utilizes unique k-mer distributions across sequences in order to assign taxonomy ^[51]. In contrast, MetaPhlAn2 uses unique clade-specific marker genes to differentiate microbial taxa and estimate their relative abundance ^[52].

Metatranscriptomics, Metabolomics, and Metaproteomics

Metatranscriptomic methods leverage similar analytical ideas to shotgun metagenomics but specifically capture the RNA transcribed from microbial cells, allowing assessments of the expression activities of these organisms ^[53]. Shotgun metagenomics and metatranscriptomic methods primarily rely upon Illumina sequencing methodologies, most often the HiSeq or NovaSeq families of instruments due to the high throughput and low cost per base. However, there has also been a movement toward PacBio and Oxford Nanopore sequencing technologies to leverage longer read lengths that aid in gene calling and genetic mapping to reference genomes. A standard workflow involves the isolation of the total RNA from the microbiome sample, RNA enrichment, fragmentation, cDNA synthesis, and preparation of transcriptome libraries for sequencing ^[54]. Typically, RNA sequence reads are mapped to different genomes and pathways (KEGG) ^[55] to identify the taxonomy of transcriptionally active organisms as well as the function of their expressed genes. Bioinformatics programs like SOAPdenovo ^[56], have been used for alignment and assembly of metatranscriptomic data from microbiome samples. Comparisons are made between different groups or variables to determine which pathways are up and down regulated or during different health and disease conditions ^[57].

Metabolomics analyses focus on profiling the metabolites microbiota produce and how these products interact with both microbiota and host metabolism ^{[58, 59]}. These methods often quantify small molecules such as antibiotics, antibiotic biproducts, and host and/or bacterial metabolism intermediates. Metabolomics often utilize mass spectrometry to identify known metabolites ^[59]. Metaproteomics also use mass spectrometry, but focus instead on identifying and quantifying the proteins present within a microbiome ^{[60, 61]}. Metaproteomics and metabolomics are both rapidly developing technologies for studying the microbiome.

Characterizing the Microbiome

Microbial Diversity Measurements

In general, microbiome differences are typically evaluated by comparing either (or both) alpha and beta diversity metrics. Alpha diversity metrics quantify within sample diversity and can be compared across groups ^[62]. For example, it is commonplace to compare the mean species diversity between samples from a diseased versus a control group. Regularly used alpha diversity metrics are species richness estimators such as observed OTUs and Chao1 index ^[63], as well as Shannon and Inverse Simpson indices, which measure both species richness and evenness ^[64]. Another less frequent way to measure diversity are phylogenetic richness estimators, such as Faith’s phylogenetic diversity ^[65]. Estimators of both richness and evenness, such as Shannon and Inverse Simpson, are more considered more robust as they are less sensitive to the number of sequences per sample ^[65]. The Shannon index is more influenced by rare OTUs, while Inverse Simpson is more influenced by the dominant/abundant OTUs ^[66].

Beta diversity compares between sample diversity, and is often calculated by comparing feature dissimilarity, resulting in a distance matrix between all pairs of samples ^[67]. One conventional calculation for beta diversity is the Bray–Curtis dissimilarity, which is a quantitative measure that accounts for taxa abundance when comparing two communities ^[68]. In addition to taxa abundances, Weighted Unifrac distance also takes into consideration phylogenetic relatedness when measuring the differences in two communities ^[69]. Unweighted Unifrac distance, on the other hand, is a qualitative measure which only considers the presence and absence of taxa. Both Unifrac measurements require a phylogenetic tree as the scores are derived by calculating total branch distances between shared and unshared bacteria on a phylogenetic tree ^[69]. The Jaccard index, or similarity coefficient, is another qualitative measurement that does not consider relative abundances, but feature presence/absence ^[65]. By evaluating taxa presence, absence, and abundances one can investigate the extent of the difference between the compositions of communities between samples. If possible, the addition of relatedness allows for the potential to assess evolutionary divergence. It is important to consider that although taxa can be distinct, related organisms are likely to perform similar functions. Software for calculating alpha and beta diversity are included in commonly used bioinformatics pipelines such as QIIME ^[28], Mothur ^[27], as well as VEGAN and phloseq ^[70] packages in R.

Functional analysis

In addition to the taxonomic composition of a microbiome, researchers can identify differences in metabolic function between different microbial populations. Using 16S sequencing data, one can predict a functional profile by using programs such as PICRUSt ^[71] or Tax4Fun ^[72]. Using the relative abundance of taxa within the community, these programs can predict the gene content potential functionality based on the reference genome for each taxa present. However, these methods are a rough approximation, as they do not take into consideration actual protein expression and are highly dependent on the reference genomes and their annotations.

Shotgun and metatranscriptome approaches can also allow for functional analysis. Once a metagenome is assembled, gene predictions are made using tools such as MetaGeneMark and Glimmer-MG. Following coding gene identification, functional annotation is carried out using computationally demanding protein sequence homology-based searches (typically UBLAST and USEARCH based ^[73]) against databases of orthologues (e.g. EggNOG ^[74] or COG ^[75]), enzymes (e.g. KEGG ^[55]), or protein domains and families (e.g. Pfam ^[76], TIGRFAMs ^[77], or InterPro ^[78]). Pathway enrichment analysis, clustering, and scoring can be performed using programs such as pathfinder ^[79], or a metabolic network can be derived as well using programs such as KEGGscape ^[80].

Due to the vast number of tools and computational demands, a number of publicly available automated pipelines have been created which include quality-filtering, gene calling, functional annotation, and basic statistics and visualization all on one platform, such as MG-RAST ^[81] and MEGAN-CE ^[82]. Additionally, programs like HUMAnN2 can be used to determine the presence/absence and abundance of microbial pathways and gene families in a community using short read data directly ^[83]. The results are generally represented in the form of a table, heatmap, bar plot, or butterfly plot containing functional pathway abundance levels relative to a reference group in order to visualize variation between different communities.

As discussed above, more precise examinations of microbiota functionality are possible using metabolomics, metatranscriptomics, and metaproteomics. Direct measurements of fecal (or other sample type) metabolites can be measured via liquid chromatography and high resolution/tandem mass spectrometry (MS/MS) ^[84]. Many metabolites, however, are highly volatile, thus collection and proper storage of samples are critical for reproducibility and concordance. Although the metabolite content of the sample can be measured, this method cannot differentiate human vs. microbial derived products, nor can it distinguish which microorganism the metabolite is derived from ^[85]. Metatranscriptomic sequencing offers characterization of which RNAs are transcribed, but does not reflect a perfect representation of functionality as protein expression also depends on translation and post-translational modifications ^[53]. Metaproteomics also employs a chemistry-based approach to characterize proteins being produced by the microbiota, with the advantage being these measurements can be used to assess the “expressed” proteins of microbial community members ^[85].

Data Visualization and Statistical Methods for Differential Group Comparisons

In most microbiome studies, the approach to analysis is to look for differential microbial diversity, taxa abundance, or functional components (e.g., genes or pathways) between the comparison groups (i.e., treatment versus group control). Statistical methods for microbiome analysis are an ever-evolving field due to the inherent complexity of microbiome datasets. The employment of common statistical methods are often difficult because microbiome data sets are high-dimensional as they can potentially have thousands of taxonomic units, zero-inflated due to the majority of taxa being rare or differences in sequencing depth, and most data output are compositional ^[86]. Moreover, the main limitations of microbiome studies in general are a lack of statistical power and the inability to control for confounding factors that are intrinsic to clinical cohort studies.

Most investigators start by visualizing the data in order to discover potential associations or markers within their dataset which can then be further tested via more rigorous statistical methodologies. Due to the complexity of microbiome data, visualization methods often use dimension-reduction-based ordination methods, such as principal coordinate analysis (PCoA) or principal component analysis (PCA) ^[87]. These methods reduce the distance matrices into two- or three-dimensional visual representations of sample distances. These samples then can be easily labelled (e.g, using color, shape, etc.) by different categories in order to overlay important clinical metadata. This allows the investigator to visualize potential clustering by clinical variables in an unsupervised way.

One can then apply more sophisticated statistical evaluation to determine if the clustering is biologically significant. When looking at differences in the community composition as a whole, ANOSIM has been used to assess significant clustering differences by comparing within- and between-group similarity using distance metrics. ANOSIM confirms or rejects the null hypothesis that the average similarity between samples of one group is the same as the average similarity between samples of a different group ^[88]. PERMANOVA is a non-parametric permutation test which performs a multivariate analysis of variance based on distance matrices to test the overall difference in microbiome community structure between different clusters or groups ^{[89, 90]}. ANCOM is a similar method which detects differentially abundant taxa between microbial populations based on log-ratios ^[91].

Heat maps are another popular method of visualizing microbiome data to detect potential clusters or differences between groups ^[87]. The most generic heat maps compare taxa abundance between samples or the presence/absence of specific gene families. Often, hierarchical clustering can be combined with the heatmaps in order to further group samples with similar bacterial profiles into branches of a dendogram. Further, clinical metadata can also be overlaid onto the heatmap in order to discover potential clinical cofactors associated with specific bacterial profiles.

Conventional statistical approaches, such as the t-test, Wilcoxon rank-sum test, ANOVA, or the Kruskal-Wallis test are often used to compare simpler features between groups such as to determine differences in alpha diversity or abundances of single known specific taxa associated with disease phenotypes ^{[88, 92]}. However when comparing lower level taxonomic differences, such as genus, species, or OTUs, these traditional tests are prone to false positives due to the large number of variables if not corrected for false discovery ^[92]. One option is to use linear discriminant analysis of effect sizes (LEfSe), which is a widely used method developed specifically for microbiome data ^[93]. This method first calculates the Kruskal-Wallis rank sum p-value to detect significant differentially abundant features between groups, then performs linear discriminant analysis to determine the effect size of those specific attributes. Ideally, all associations should be experimentally tested directly either in vitro or in vivo to confirm the detected associations.

Graphical Networks and Machine Learning

Network analysis a popular approach for visualizing microbiome community interactions. Correlation networks are not only used to illustrate perturbations to the community structure in different scenarios, but can also be used to see how environmental factors, metabolites, clinical traits, or other microbes (i.e. interkingdom interactions) interact within the microbial community. Graphical networks are often used to compare interactions in different states (e.g. diseased vs. healthy state) or to show which organisms co-occur or mutually exclude each other.

Generally, networks can be inferred using pairwise relationships where networks are based on similarity or correlation coefficients between pairwise variables. Commonly used programs to infer correlation networks for microbiome data are SparCC (Sparse Correlations for Compositional data) ^[94], CCLasso (Correlation inference for Compositional data through Lasso) ^[95], and SPEIC-EASI (Sparse and Compositionally Robust Inference of Microbial Ecological Networks) ^[96]. At the current, SPEIC-EASI is the most widely used model as it appears to be the most robust estimating interactions using algorithms for both sparse neighborhood and inverse covariance selection after first CLR transforming the count data. Regression-related methods can also be used to predict the abundance of one species from the abundance of combinations of other sets species, such as sparse regression ^[97], Dirichlet-multinomial regression ^[98], and generalized boosted linear models (GBLMs) ^[97]. Another tactic is contingent on the presence–absence patterns of taxa in association with different phenotypes/outcomes, also known as association rule mining ^[99].

Machine learning algorithms are emerging as a widely accepted tool in the analyses of microbiome datasets, as they can be used on many different types of high-dimensional data and are especially useful for feature selection in multi-feature datasets ^[65]. Random Forest ^[100], for example, is an ensemble machine learning technique for classification and regression that is often applied to identify important taxa and clinical covariates that can distinguish different phenotypes/ categories or predict specific outcomes. Although less accurate and stable, methods such as CART analyses (Classification and Regression Trees) can more interpretable and thus clinically actionable as decision trees allow the investigator to have insights into which variables are important at what cut off, and in what order ^[101]. One must consider however that these models must always be cross-validated either via sample and replacement, or by enrolling separate cohorts for training, testing, and validation.

Model systems for studying the microbiome

Animal models

As previously discussed, the aforementioned analyses only derive association or correlation of microbiome features with disease phenotypes, but to not determine causation. Thus, interactions between the microbiome and the host must be tested and extrapolated from model systems in order to gain mechanistic insights. Animal models are extensively utilized to study the human microbiome as animals allow for ease of experimental manipulation and control of specific variables, scalability, and reproducibility that is mostly unattainable in human studies. Gnotobiotic, or germ-free mice, are frequently used to test either the effects of the absence of the microbiome or specific communities or strains of interest can be administered to the animal and evaluated for effects on the host ^[102]. For example, “humanized” murine models are produced when a human fecal source is used to colonize the mouse ^[103]. Other less expensive animal models include zebrafish, fruit flies, and Caenorhabditis elegans which can also be used to test whether the microbiome composition and function correlate with such host variables (i.e. age, genotype, phenotype, diet etc.), or experimental exposures ^[104]. Of course, important caveats for animal models include evolutionary distance which results in relevant differences in anatomy, physiology, and the microbiome composition itself in relation to the human host ^[105].

In vitro and ex vivo models

Rather than using model organisms, an investigator may choose to study microbiome interactions in vitro or ex vivo ^[106]. The main experimental systems currently used to study microbial interaction with the host include co-culturing microorganisms with primary host tissues or cell lines; microfluidic culture with engineered tissue; and intestinal enteroids or organoids ^[107]. Additionally, in order to understand specific microbial changes, one may choose to use an in vitro bioreactor device ^[108], such as a simulator of the human intestinal microbial ecosystem (SHIME) ^[109]. These simulator devices, are continuous culture systems that mimic the human digestive tract microbiome and allow the control of nutrient availability, pH, and other environmental conditions ^[110]. More advanced systems even control for compartmentalization. Although the use of in vitro primary cells and established cell lines allows for simplicity and reproducibility, these models are limited in their translatability in that they are a single cell type and only a few organisms can we tested at once ^[106]. Ex vivo systems allow for the ability to model cross-talk between cell types not generally found in in vitro systems. Moreover, systems like InTESTine™ system by TNO ^[111] and enteroids/organoids ^[112] allow for modeling of bacterial-host interactions in live compartmentalized and physiologically relevant tissue. However many ex vivo systems have not been validated, and remain costly.

Important Future Considerations for Microbiome Investigation

The current tools available to study the microbiome are maturing rapidly and the future is bright. As more researchers begin to investigate the role of the microbiome in human health and disease, more focus is brought to technological advancements as well as developing robust statistical methods and analytical tools.

One current limitation gaining significant attention is how to estimate the sample size necessary to address a specific research question. Current methods exist to estimate the sample size needed for PERMANOVA based analyses ^[89], as well as for comparisons leveraging the Dirichlet-Multinomial distribution ^[98], but methods to identify the number of samples needed to test hypotheses with many of the methods listed above remain undeveloped. This is partially due to the lack of established metrics for determining effect size in terms of changes in the microbiome. We can consider an effect size in two ways for microbiome studies: how much change must be observed to be biologically relevant, and how much difference in the microbial consortium must exist to differentiate two or more groups? Work is ongoing to establish the source of both technical and biological variation in the microbiome, and the more studies we have assessing topics such as the diet’s impact on the gut, the more insight we have into the consistency and scale of differences observed ^{[113, 114]}. Both of these are needed to inform effect size estimates used within sample size calculations.

In addition to the need for better sample size estimation, there is a current need for advanced statistical modeling tools, particularly ones to analyze data from longitudinal studies. One of the hallmarks of the human microbiome is its dynamic nature that can be transformed and altered over time ^{[115, 116]}, but this may potentially be in direct contrast to the observation of a core microbiota that resists perturbation ^[117–119]. Better longitudinal methods are required to truly interrogate this variability and elucidate the complex relationships between taxa, the host, and the environment over time. Methods have recently been released to visualize longitudinal data and generate estimates adapted from microbial ecology ^[120], and to generate dynamic Bayesian networks ^[121]. Additional advances in longitudinal modeling are forthcoming and will continue to expand the repertoire of analytical tools available to scientists.

While a general consensus in bioinformatics and biostatistics methods has been reached for 16S targeted sequencing-based methods, there has been a recent batch of publications and methods attempting to control and adjust for the technological variability inherent within amplicon based methods ^{[22, 29]}. These methods focus on statistical modeling of the errors of Illumina based sequencing to correct individual reads, purporting to generate species or strain specific groupings, similar to OTUs. Early assessments of these methods using known mock communities show the true complexity of microbial communities is overestimated by these methods, but the potential for error correction is promising ^[29]. While the utility and impact of such methods is to be determined, it is important to consider such potential advancements as we continue to explore the impact of the microbiome in health and disease.

Contrary to 16S targeted sequencing methods, there is still a need for universal and comprehensive analytical pipelines for shotgun metagenomics and metatranscriptomics. In particular, while there are available tools that leverage reference genomes for alignment ^[122], utilize custom databases of marker genes ^[52], perform metagenomic assemblies ^[48], and utilize unique k-mer patterns to identify taxa ^[51], a consensus has not been reached within the microbiome community as to which method(s) is superior and which pipeline(s) best implements these methods. Until this consensus is reached, the barrier for interrogating the metagenome may be too high for many researchers to adopt the tools in their own studies.

As we continue to remove the technological barriers, and reduce the cost of generating metagenomic, metatranscriptomic, metabolomic, and metaproteomic data, there will continue to be a push into the next phase of analytical approaches dealing with integrated ‘omics data. Current methods for integrating these data types rely upon linking them through a common database, such as UniProt ^[123] for metagenomics data and metaproteomic data, or ordinating them in multidimensional space ^{[124, 125]}. This integrated omics approach will also be integral in further attempts to mine the microbiome for antimicrobial resistance determinants ^[126], virulence genes that impact the outcomes of infection and facilitate bacterial recalcitrance to treatment ^[127], and assessments of the roles of accessory genes and genes of unknown function within the microbiome ^{[128, 129]}.

Conclusion

In summary, we stand on the precipice of a rapidly developing research area that holds incredible promise in understanding and defining microbial interactions at the human interface. However, under a “bedside, to bench, and back to bedside” theory of investigation where studies often start with observational retrospective or prospective cohort studies, researchers must take special care in their experimental design to systematically include all steps in the examination process in order to make proper advancements in the field (Figure 2). This includes not only microbiome characterization and associations with clinical metadata using patient populations, but proper statistical analysis and modeling methods which take into consideration the eccentricities of metagenomics, validation of findings using other cohorts, mechanistic discovery in vitro, ex vivo, or in small animal models. Only then can we come full circle back to the patient using interventional strategies with microbiome-based therapeutics

Figure 2.

A “bedside, to bench, and back to bedside” approach to microbiome investigation

Future Unmet Needs and Implications for the Clinician.

• Universal analytical pipelines for shotgun metagenomics and metatranscriptomics still remain elusive among the research community, thus researchers from multiple disciplines and institutions must work together to come to a consensus.

• Comprehensive methods for multi-omic integration continue to be ambiguous, yet investigations combining metagenomics, metabolomics, transcriptomics, and host genomics are crucial to moving the field forward.

• In order to bring microbiome research into the therapeutic realm, investigators need to move beyond clinical association studies by validating their models in other clinical cohorts and understanding the mechanisms of causation in vitro, ex vivo, and animal model systems.