High-Resolution Characterization of the Human Microbiome

Abstract

The human microbiome plays an important and increasingly recognized role in human health. Studies of the microbiome typically employ targeted sequencing of the 16S rRNA gene, whole metagenome shotgun sequencing, or other meta-omic technologies to characterize the microbiome's composition, activity, and dynamics. Processing, analyzing, and interpreting these data involve numerous computational tools that aim to filter, cluster, annotate, and quantify the obtained data and ultimately provide an accurate and interpretable profile of the microbiome's taxonomy, functional capacity, and behavior. These tools, however, are often limited in resolution and accuracy and may fail to capture many biologically- and clinically-relevant microbiome features, such as strain-level variation or nuanced functional response to perturbation. Over the past few years, extensive efforts have been invested toward addressing these challenges and developing novel computational methods for accurate and high-resolution characterization of microbiome data. These methods aim to quantify strain-level composition and variation, detect and characterize rare microbiome species, link specific genes to individual taxa, and more accurately characterize the functional capacity and dynamics of the microbiome. These methods and the ability to produce detailed and precise microbiome information are clearly essential for informing microbiome-based personalized therapies. In this review, we survey these methods, highlighting the challenges each method sets out to address and briefly describing methodological approaches.

Introduction

Recent marked advances in sequencing technologies have been followed by an explosion of studies utilizing these technologies to explore a wide range of microbial communities, including those that inhabit the human body. Such studies apply targeted sequencing of the 16S rRNA gene as well as whole metagenome shotgun sequencing to characterize the human microbiome in numerous settings. Analyses of these sequencing data commonly use an assortment of clustering, binning, annotation, and assembly algorithms to ultimately profile the composition of species in each sample, the set of genes they collectively encode, or the genome sequence of specific member species (Figure 1). Combined, these efforts to map the human microbiome in health and in disease have led to an increased appreciation for the important role of the microbiome in human well-being^1–5.

Figure 1. Schemes of microbiome analysis.

Metagenomic data, as well as other meta-omic data, can be processed and analyzed in various ways to address a diverse set of questions concerning the microbiome's composition, capacity, and function.

Nevertheless, common computational metagenomic analysis methods are often limited in resolution and may fail to resolve nuanced, yet important and potentially clinically-relevant details concerning the composition of species and genes in the microbiome. Standard 16S rRNA surveys, for example, are often limited to a genus-level taxonomic identification⁶, can fail to distinguish closely related taxonomic groups, and cannot always unambiguously discriminate rare, low-abundance taxa from noise⁷. Shotgun metagenomic analyses may similarly fail to identify the taxonomic origins of a gene of interest or to produce accurate and unbiased estimates of gene families’ abundances^8,9.

Clearly, however, given the complexity of the human microbiome, accurate and high-resolution mapping of the microbiome is crucial for gaining a principled understanding of community behavior, function, and ultimately its impact on the host¹⁰. For example, accurately profiling strain-level microbiome composition is vital for tracking ecological trends over time, such as the spread of bacterial vaginosis-associated strains between sexual partners¹¹. Discerning subtle genomic variation between closely related strains of the same species may also have important clinical implications, as in the case of Propionibacterium acnes, which displays extensive strain variation in the skin microbiome with potential impact on various skin conditions¹². Likewise, Escherichia coli has well-characterized variation in toxin production, which results in high pathogenicity for a subset of strains, while other strains are commonly found in healthy gut microbiomes¹³. Careful differentiation of strains can also inform clinical decision making by, for example, providing valuable insights as to whether a patient will respond to the heart failure drug digoxin¹⁴. Accurate detection of low abundance species is similarly essential as such rare species may still play important roles in various biological processes. Indeed, even species present at less than 0.01% abundance in oral microbial communities can play a key role in causing oral inflammatory disease¹⁵. A high-quality, unbiased, and rigorous characterization of the metagenome's gene content is equally important for pinpointing disease-associated shifts in the functional capacity of the microbiome⁹.

Moreover, many molecular processes that play important roles in the microbiome's activity and dynamics go beyond the microbiome's taxonomic and genic composition and accordingly cannot be profiled through metagenome sequencing. For example, oligosaccharides found in breast milk can change microbial gene expression and production of physiologically relevant microbial metabolites in the infant gut without affecting the abundance of most species¹⁶. Exploring such processes may require detailed information about transcript and protein variation, metabolite concentrations, and spatial distribution. Indeed, new ‘omics’ technologies can now comprehensively quantify such community features, but computational methods available to analyze the resulting sizable and novel datasets are largely still in early stages and may be limited in resolution.

In this review, we accordingly describe an array of recent computational methods and analytical approaches that set out to address these challenges and to provide high-resolution, multi-omic, systematic characterizations of the microbiome at multiple levels (Table 1). While some of these approaches have primarily been applied to environmental microbial communities, all are broadly applicable and potentially useful in the context of the human microbiome and its health impacts. We first discuss taxonomic analysis of the microbiome, focusing on methods for detecting strain-level variation within each member species. We specifically describe methods that utilize targeted 16S rRNA or whole metagenome sequencing data for strain-level profiling, identification, and tracking, either de novo or based on existing reference genomes. We also describe recent methods for assembling the genomes of novel strains directly from metagenomic data. We next discuss methods for improved functional characterization of the microbiome, including accurate detection of the various gene families encoded by the metagenome and precise quantification of their abundances, and for linking taxonomic and functional profiles. Lastly, we describe several recent frameworks for analyzing and integrating other microbiome-derived high-throughput ‘omic’ datasets and for profiling additional facets of the microbiome's composition and activity.

Table 1.

Tool	Synopsis	Ref.	Website
16S rRNA analysis
M-pick	Identifies taxonomic clusters based on modularity analysis of a sequence read graph	24	http://plaza.ufl.edu/xywang/Mpick.htm
Swarm	Uses iterative linkage clustering to identify natural subpopulations	27	https://github.com/torognes/swarm
Minimum entropy decomposition	Hierarchically clusters sequences by iteratively subdividing based on sequence entropy	30	http://merenlab.org/2014/11/04/med/
Oligotyping	Identifies subpopulations within predefined OTUs by identifying and clustering the most informative nucleotide positions	31	http://merenlab.org/projects/oligotyping/
Oclust	Hierarchically clusters PacBio CCR reads into OTUs	33	https://github.com/oscar-franzen/oclust/
CopyRighter	Correct 16S rRNA data for copy number variation	40	https://github.com/fangly/AmpliCopyrighter
metagenomeSeq	R package for normalization and differential abundance analysis	42	http://cbcb.umd.edu/software/metagenomeSeq/
phyloSeq	R package for processing, normalization, differential abundance analysis, and visualization	43	https://joey711.github.io/phyloseq/
RAIDA	Differential abundance analysis using ratios between taxa	44	http://cals.arizona.edu/~anling/sbg/software.htm
Strain-level characterization using shotgun metagenomic data
WG-FAST	Identified strains from low coverage genome datasets	63	https://github.com/jasonsahl/wgfast
Sigma	Taxonomic analysis of metagenome data at the strain level, including variant calling and statistical uncertainty calculations	62	http://sigma.omicsbio.org
Constrains	Identifies strains from metagenomic sequence data and reconstructs their phylogeny	64	https://bitbucket.org/luo-chengwei/constrains
PathoScope	Identifies the proportion of reads from individual microbial strains in metagenomic sequencing data	70	https://sourceforge.net/projects/pathoscope/
-	Large-scale characterization of strain-level copy number variation	66	http://elbo.gs.washington.edu/download.html
phyloCNV	Profiles species abundance; identified, characterizes, and analyzes strains based on SNPs and copy number variation	65	https://github.com/snayfach/PhyloCNV
Assembling reference genomes from shotgun metagenomic data
MEGAHIT	Assembles metagenomic short reads with low memory use	78	https://github.com/voutcn/megahit
MetaVelvet	Assembles metagenomic short reads into contigs	79	http://metavelvet.dna.bio.keio.ac.jp/
CONCOCT	Binning using gaussian mixture models and sequence composition and abundance	90	https://github.com/BinPro/CONCOCT
GroopM	Automated genome recovery from metagenomes	87	http://ecogenomics.github.io/GroopM/
MetaBAT	Bins genomes based on tetranucleotide frequency and abundance	89	https://bitbucket.org/berkeleylab/metabat
MaxBin	Bins assembled genomes using EM algorithm	88	http://downloads.jbei.org/data/microbial_communities/MaxBin/MaxBin.html
ABAWACA	Binning based on mono, di, and tri-nucleotide frequency and abundance, using hierarchical clustering	91	https://github.com/CK7/abawaca
CheckM	Assessing quality of putative genomes	94	http://ecogenomics.github.io/CheckM/
MetaDecon	Metagenomic deconvolution and reconstruction of species-specific genomic content	93	http://elbo.gs.washington.edu/software_metadecon.html
Specialized functional annotation HMM databases
FOAM	Annotation of KEGG orthology groups	119	http://cbb.pnnl.gov/portal/software/FOAM.html
Resfams	Annotation of antibiotic resistance genes	108	http://www.dantaslab.org/resfams
dbCAN	Annotation of carbohydrate-active enzyme (CAZyme) genes	125	http://csbl.bmb.uga.edu/dbCAN/annotate.php
Functional annotation and quantification of shotgun metagenomic data
ShortBRED	Representative marker-based protein family profiling	126	https://huttenhower.sph.harvard.edu/shortbred
MUSiCC	Accurate normalization of functional profiles	9	http://elbo.gs.washington.edu/software_musicc.html
MicrobeCensus	Accurate normalization of functional profiles	129	https://github.com/snayfach/MicrobeCensus
Combined taxonomic and functional annotation of shotgun metagenomic data
LMAT	k-mer based taxonomic assignment of metagenomic reads	133	https://sourceforge.net/projects/lmat/
Kraken	k-mer based taxonomic assignment of metagenomic reads	134	http://ccb.jhu.edu/software/kraken/
TAC-ELM	Neural network-based taxonomic classification of metagenomic reads	135	http://cs.gmu.edu/~mlbio/TAC-ELM/
MetAnnotate	Combined taxonomic and functional annotation with HMMs and alignment to HMM-family protein sequences	138	http://metannotate.uwaterloo.ca
SeMeta	Multiple-level clustering of reads followed by cluster assignment to taxa through representative read alignment	140	http://it.hcmute.edu.vn/bioinfo/metapro/SeMeta.html
MetaCluster-TA	Clusters assembled contigs and assigns taxonomy by alignment	139	http://i.cs.hku.hk/~alse/MetaCluster/index.html
Meta-omic analysis tools
TAG	Assembles a metatranscriptome incorporating information from a metagenome assembly	153	http://omics.informatics.indiana.edu/TAG/
Anvi'o	Interactive processing and visualization of metagenomes and metatranscriptomes	156	http://merenlab.org/projects/anvio/
Pipasic	Produces quantitative strain-specific peptide assignments using a sequence similarity correction	163	https://sourceforge.net/projects/pipasic/
BacSpace	Processing and quantitative analysis of microbial community FISH imaging data	174	https://bitbucket.org/kchuanglab/bacspace/downloads
MIMOSA	Metabolic model-based integration of microbiome taxonomic and metabolomic profiles	172	http://elbo.gs.washington.edu/software_MIMOSA.html

High-Resolution Characterization of the Microbiome's Taxonomic Composition

One of the most common and relatively accessible starting points for human microbiome analysis is taxonomic profiling. Specifically, by sequencing and analyzing taxonomy-associated marker genes, researchers can readily identify the various species present in a given microbiome sample and estimate the relative abundances of each species¹⁷. The study of such taxonomic profiles and the way they vary across individuals or between cohorts can provide numerous insights into the link between the microbiome ecology and the host's health. Such studies can, for example, pinpoint specific species with known virulence factors or community-wide dysbiotic features as biomarkers of disease^18,19. As noted above, however, taxonomic profiling is often limited in resolution and may therefore hinder our ability to detect more fine-grained determinants of disease. Below, we describe several new and exciting developments in the analysis of both marker gene data and whole metagenomes that aim to provide a more detailed, high-resolution map of the microbiome's taxonomy.

High-resolution and accurate analysis of 16S rRNA data

To date, the most prevalent form of comprehensive microbiome taxonomic data is produced via targeted amplification and sequencing of the 16S rRNA gene, a commonly used phylogenetic marker²⁰. The analysis of such 16S rRNA sequencing data typically involves clustering of the obtained sequences (usually based on sequence overall percent similarity) into Operational Taxonomic Units (OTUs), and determining the relative abundance of each OTU in the sample. The taxonomy of each OTU can then be inferred by clustering reads with reference sequences of known taxonomy or by a classifier algorithm that predicts each OTU's (or each read's) likely taxonomy^21,22.

This clustering-based approach is efficient, widely-used, and well-established; yet several challenges remain in terms of accurate and precise taxonomic quantification at the species or strain level. First, a measure of the overall percent similarity between two 16S rRNA sequences may not fully capture the variation present in the sequenced region or taxonomic divergence this variation represents. Indeed, the number and nature of polymorphisms and of true subpopulations included within a single, similarity-based OTU cluster can vary greatly across OTUs²³. A number of recently introduced algorithms aim to account for such variation using graph-based clustering approaches and grouping sequences based on local base differences between reads rather than by overall percent similarity^24,25. These algorithms have proved successful in identifying higher-resolution sequence clusters and more accurately describing the population structure in each sample. One example of such an algorithm, termed Swarm^26,27, first performs exact linkage clustering to group reads that have one nucleotide differences to any other read in the same cluster, and then refines each cluster based on read abundance distributions. This approach has been successfully applied to characterize fine-scale taxonomic profiles of bacteria and protists in several environments^28,29. An alternative method for addressing the limitation of similarity-based clustering, termed minimum entropy decomposition, generates a hierarchy of read groupings by iteratively subdividing the dataset into groups based on the entropy explained by each division³⁰.

Moreover, the clusters produced by any OTU picking method may also vary in homogeneity and within-cluster diversity across the various samples. Describing this within-cluster variation may lead to sub-OTU level taxonomic insights such as sharing of an OTU subpopulation across samples One computational approach to capture this variation (termed Oligotyping) uses Shannon entropy calculations to detect the most informative nucleotide variation and to correctly identify subpopulations within predefined OTU clusters³¹. This method relies on a combination of strategies to deemphasize likely sequencing errors compared to true strain variation and has been successfully applied to track strains of Gardnerella vaginalis shared between sexual partners¹¹ and to study population dynamics in the oral microbiome and in sewage^30,32. Another approach for extracting more detailed taxonomic information from 16S rRNA reads relies on longer-read sequencing technologies (most notably, PacBio Single Molecule Real Time sequencing) to obtain sequence data from more variable regions of this gene. Two recently introduced pipelines process and cluster PacBio circular consensus sequencing reads^33,34, accounting for the specific characteristics of this different sequencing platform.

Notably, as methodologies for characterizing 16S rRNA datasets have proliferated, so have studies comparing and evaluating these approaches^22,35–37. These studies, however, have not necessarily reached a clear consensus on the superior approach, but have rather demonstrated that the choice of algorithm can have substantial impact on subsequent analyses, and that the best choice of method likely depends on the community being analyzed and the sequencing technology used.

Once sequences have been grouped into OTUs, OTU abundances can be analyzed and compared across taxa or samples. However, accurate measurements of 16S rRNA read count may not necessarily accurately mirror the abundances of the various taxa in the community. First, since the copy number of the 16S rRNA gene varies across microbial genomes, 16S-based surveys may overestimate the abundances of taxa with multiple copies of this gene. Using reference genome information to normalize for this variation can adjust and improve estimates of the relative abundance of different taxa in the same sample^38–40. PCR amplification can also introduce bias into abundance comparisons between taxa, since ribosomal genes from some taxa may amplify poorly with commonly-used primer sets⁴¹. This limitation also prevents the comparison of taxonomic abundances across different datasets generated using different primers. Lastly, the relative abundance of reads assigned to a given OTU across samples can be skewed by changes in the absolute abundance of another OTU, a phenomenon known as compositional bias. A number of tools have been introduced to correct for this bias, primarily by adopting techniques developed to address a similar problem in RNA-Seq experiments^42–44, or alternatively to account for this effect in analyzing relative abundance values⁴⁵. Failure to address this bias can result, for example, in the identification of spurious correlations between the abundances of different OTUs, limiting our ability to robustly analyze co-occurrence relationships between different taxa⁴⁶. Combined, these various biases render the relationship between the relative abundances of 16S rRNA reads and true taxonomic abundances extremely complex. One recent study set out to comprehensively characterize the joint impact of the various factors influencing this relationship by using synthetic mock communities of vaginal microbiome taxa and fitting regression models that predict true abundance of a given taxon as a function of both 16S rRNA read count and several taxon-specific bias correction terms⁴⁷. While such a detailed approach can be helpful for interpreting and analyzing 16S rRNA datasets of well-studied taxa, the precise relationship between 16S read counts and true community taxonomic structure for many microbiome studies remains to be characterized.

Resolving strain-level taxonomy from shotgun metagenomic data

While 16S rRNA-based surveys can provide important insights into the taxonomic composition of a given microbiome sample, their ability to resolve strain-level genetic diversity is inherently limited. In fact, substantial genotypic variation can exist in the absence of noticeable 16S rRNA sequence divergence⁴⁸. This variation can impact the capacity and behavior of a species and ultimately impact community-level activity. For example, some species, such as E. coli, have extremely marked variation in gene content^48,49, which can influence the strain's pathogenicity or ecological niche^50,51. Moreover, the concept of a bacterial species is in fact somewhat subjective and may not be captured well by the level of divergence in a ribosomal gene sequence⁶. It is therefore often informative to go beyond species-level resolution and to characterize the composition of strains (i.e., within-species taxonomic divisions) and strain-level variation within the microbiome.

Unfortunately, however, traditional methods for detecting, characterizing, and tracking strain-level diversity rely on sequencing^52–56 or applying microarrays^57,58 to cultured isolates, and are therefore not readily applicable in a microbiome setting where many microbial taxa cannot be easily isolated or cultured^59,60. Moreover, efforts to isolate all strains of interest in a given microbiome sample and sequence their genomes can be extremely resource-intensive⁶¹. An increasingly feasible alternative is to decipher strain-level diversity directly from shotgun metagenomic data using a plethora of novel and sophisticated computational techniques. Indeed, by identifying within-species genetic variation directly from metagenomic samples, a more comprehensive set of strains can be characterized in a high-throughput manner from a single sequencing experiment. This approach has been successfully applied, for example, to detect pathogenic strains of E. coli in clinical samples or for biosurveillance^62,63, to identify novel strain-level dynamics in the infant gut⁶⁴, to confirm the retention of personal strains over time⁶⁵, and to demonstrate extensive, widespread, and clinically-relevant strain-level variation in the gut microbiome⁶⁶.

Notably, strain-level variation can manifest in two ways: single nucleotide polymorphisms (SNPs) within shared genomic content, and variation in the presence (or copy number) of complete genes or specific segments of the genome. Most recently developed metagenomics-based SNP analysis methods take advantage of reference genome collections to estimate community diversity, detect strains of interest, or find shared strains between different metagenomic samples. These methods may use full genomes^62,63,65,67 or marker genes known to contain loci with strain-identifying SNPs⁶⁴. Since some reference genomes may be extremely similar to each other, the first step in many of these methods is to cluster genomes by similarity and select a representative genome for each cluster to which metagenomic reads can be aligned. This step speeds up the alignment process and reduces the likelihood of reads from the same strain being mapped across different but closely related references. One method, Sigma⁶², first uses such reference genome alignments to estimate relative abundances of taxa and then relies on the obtained estimates to refine the read mapping assignments. ConStrains⁶⁴ takes a more efficient approach of mapping reads only to previously-identified informative marker genes, which facilitates strain profile comparisons between samples (e.g. for tracking strain dynamics in the infant microbiome^68,69). Another tool, PathoScope⁷⁰, identifies strains using reference genomes and estimates the relative share of different strains from the same species in a given sample, and is particularly optimized to detect low abundance strains from clinical samples.

The detection of strain-level variation in gene copy number (CNVs) or in gene content is a more specialized application of shotgun metagenomic-based taxonomic classification that focuses specifically on functional variation. Indeed, CNVs are an important source of functional difference between strains, with many variable genes involved in metabolism^71,72, membrane and transport proteins^56–58, and virulence^55,71. Identifying which genes are present, absent, or vary in copy number across the various strains in a microbiome sample is therefore a crucial task that has been address by several recent studies. Most of these studies rely on mapping short metagenomic reads to some set of reference genomes (using a variety of read-mapping strategies), aiming to detect genomic regions for which the observed coverage varies from our expectation. Analysis of data from the Human Microbiome Project, for example, used a similar approach, mapping reads directly to a reference genome of Streptococcus mitis and demonstrating strain-level variation in the presence/absence of various genomic elements of this species¹. More recently, a first large-scale analysis of strain-level copy number variation was introduced, using universal single-copy genes to translate coverage measurements into copy number estimates and inferring the copy number of thousands of genes across dozens of species and in >100 samples⁶⁶. Comparing copy number estimates across samples, this study has demonstrated extensive and widespread strain variation in the gut, including variation associated with obesity and Inflammatory bowel disease. Several later studies used a similar approach for detecting strain-level variation but focused mostly on genes’ presence/absences rather than on variation in copy number⁷³. Other studies extending this approach have first constructed pan-genomes (as inventories of all genes known to occur in any strain of a particular species) and map reads to these pan genomes^65,74. An alternative approach to directly mapping short reads to reference genome is to first assemble metagenomic reads into contigs, identify predicted genes in these contigs, and then align those to a reference^71,72. These longer query sequences may improve strain-specific gene identification but may be more limited in scale.

Assembling reference genomes from metagenomes

Detailed identification of strains and species can be more informative if combined with information about the gene content of each genome. Genome content information represents a mechanistic link between the taxonomy of a given microbial organism and its functional capacity, and, more generally, between community ecology and community-wide activity. Indeed, many prevalent gut species have been isolated and sequenced, yet many microbial taxa (and many strains) still lack any reference sequence⁷⁵. Assembling complete genomes directly from shotgun metagenomic reads is therefore a crucial (though clearly nontrivial) task. A recent study, for example, has demonstrated the utility of assembling shotgun reads for linking taxonomic and functional dynamics after a dietary intervention for patients with Prader-Willi syndrome⁷⁶. The past several years have, however, witnessed substantial progress in the quality and number of genomes recovered and assembled from metagenomes⁷⁷.

Assembling genomes from metagenomes commonly involves two steps. First, shotgun metagenomic reads are assembled into contigs and then the obtained contigs are grouped into multiple bins such that each bin ideally includes contigs from the same taxon. The assembly step can be performed using numerous assemblers that have been optimized for assembling metagenomic reads such as MEGAHIT⁷⁸, MetaVelvet-SL⁷⁹, Ray Meta⁸⁰, or IDBA-UD⁸¹. The binning step often relies on nucleotide composition, exploiting the relationship between phylogenetic relatedness and similarly in various sequence features such as GC content or k-mer frequency⁸². Such nucleotide composition-based methods are prevalent, well-established, and several different implementations are available^83,84. More recently, a different strategy for binning was introduced, which utilizes the fact that different reads from the same species will tend to co-vary in abundance across samples^72,85. This approach was later refined by using the obtained differentially abundant bins of contigs to re-assemble the reads⁸⁶. Finally, over the past few years, several exciting methods that integrate both the nucleotide composition-based and the differential abundance-based approaches have been published, including GroopM⁸⁷, MaxBin⁸⁸, MetaBAT⁸⁹, CONCOCT⁹⁰, and ABAWACA⁹¹. Another recently introduced method, termed Latent Strain Analysis (LSA), bins genomes using single value decomposition, enabling it to assemble genomes from very large datasets and thus identify rare species not found with other methods⁹². Alternative approaches bin reads or whole genes without assembly, for example by utilizing the expected co-variation between the abundance of various genomic elements in the metagenome and the abundance of the OTU from which they originated to deconvolve the metagenome into taxon-specific genomic data⁹³. When considering these various binning methods, it should be noted that nucleotide composition-based methods have the advantage of being applicable even when only a single metagenomic sample is available, whereas differential abundance-based methods require multiple (and ideally a large number of) samples. When multiple samples are available, however, recent methods that combine both nucleotide composition and differential abundance will likely perform best. A comprehensive comparison of the performance of these many different binning algorithms has not yet been presented, though tools for validating the quality and completeness of assembled genomes are available (see, for example, CheckM⁹⁴ and METAQUAST⁹⁵).

While the methods above relied solely on metagenomic short read data, new molecular technologies hold promise for improving metagenome-based genome assembly. For example, combining short read sequencing with Hi-C data (which provide information about the physical proximity of the different sequences) has shown to improve contig binning in synthetic mixtures of microbes^96–98. Long read and single-molecule sequencing can similarly help to link sequences from the same genome. For example, PacBio reads have been combined with short reads to reconstruct high-quality, closed genomes from the skin microbiome⁹⁹, and synthetic long reads have been successfully used to improve assembly quality^100,101. These approaches require additional experimental and computational steps, but may significantly improve the ability to recover quality genomes from complex community samples, and are particularly promising for recovering genomes of rare species¹⁰¹.

High-Resolution Characterization of the Microbiome's Functional Capacity

Taxonomic analyses can be extremely useful for detecting disease-associated shifts in community composition and for characterizing states of ecological dysbiosis. Some research questions, however, may be best addressed by considering the aggregate functional potential of the microbiome, regardless of the individual species that carry a specific gene or perform a specific function. Identifying which gene families are encoded in a metagenome provides insight into the capacity of the community as a whole and allows for comparison of the functional potential of a given sample to that of another sample or another environment¹⁰². It can facilitate, for example, the identification of novel metabolic functions^103,104, disease-associated shifts in the microbiome's metabolic capacity^2,105, functional profile variations due to environmental fluctuations^106,107 or antibiotic resistance genes^108,109. In such settings, researchers commonly take a gene-centric approach, treating the community as a single supra-organism^110–112 and profiling the set of genes collectively encoded by the metagenome. To this end, these studies directly annotate each read in the metagenome (or each gene identified in assembled contigs) with a functional category. Importantly, this approach is particularly useful when the community harbors many poorly characterized species with no reference genome. In this section, we describe recent developments in functional annotation of metagenomic samples that aim to provide a more nuanced, targeted, and accurate quantification of an individual microbiome's functional capabilities.

Accurate annotation and quantification of the metagenome's functional profile

Functional annotation of shotgun metagenomic reads can be accomplished by a variety of recently introduced frameworks^113–118, and is typically based on mapping these reads to genes or protein domains with known functional classifications. Read mapping is done either by aligning each read to a reference database of gene or protein sequences or by using probabilistic models (such as Hidden Markov Models; HMMs) to evaluate the likelihood that a given read belongs to a specific protein family or domain.

The general annotation approach provides a useful broad overview of the functional profile of a community, but may have a high false positive rate due to the large reference databases used. Such false positives may represent, for example, reads originating from genes that in fact have no closely related references in the database but that still map to genes with which they share regions of homology even though they may not perform the same function. To address this shortcoming, recent efforts have produced tailored reference databases that cover specific classes of proteins, in the hope that such specialized databases could improve the specificity and accuracy of functional annotation. Specifically, while large databases of protein-based HMMs exist, several specialized HMM databases have been recently introduced for metagenomic annotation. For example, FOAM¹¹⁹ is a database designed to identify genes matching KEGG Orthology groups^116,120 that can aid in characterizing the metabolic potential of communities^121,122. Resfams¹⁰⁸, on the other hand, was developed to recognize the structure of antibiotic resistance genes and has been used to study the human gut resistomes of different cultures^123,124. Yet another database, dbCAN¹²⁵, specifically targets carbohydrate-active enzymes. A related method, ShortBRED¹²⁶, similarly quantifies a specialized set of proteins of interest, but uses alignment-based annotations rather than HMMs for a more efficient and general approach that allows for customized user-defined reference databases. These metagenomics-specific and specialized databases are a key component for accurate annotation of complex metagenomic samples. Much progress has also been made in methods for read alignment, focusing primarily on speeding up the alignment process^117,118 or providing efficient web-based annotation tools^114,116.

Notably, however, even when shotgun reads are aligned to an appropriate database the resulting calculated functional profile can be markedly impacted by various factors, including experimental and computational biases and the protocol used to annotate each read based on the obtained alignments. Sample processing and library preparation can, for example, bias the predicted functional profile of a metagenomic sample¹²⁷. A recent study systematically evaluated such homology-based annotation practices and demonstrated that variation introduced by computational protocol selection could completely mask true biological variation between samples, suggesting goal-specific best-practice guidelines for metagenomic annotation¹²⁸. Moreover, once the samples’ functional profiles have been determined, rigorous normalization and calibration of samples are still required to allow accurate comparison across samples (e.g., to identify disease-associated functional shifts). A couple of recent studies, however, have demonstrated that the commonly used compositional normalization (i.e., using the relative abundance of each gene family within the metagenome) introduces marked biases both across and within microbiome samples^9,129. These studies have further presented novel methods (termed MUSiCC⁹ and MicrobeCensus¹²⁹) that use universal single-copy genes to calibrate measurements of gene abundances and to correct these biases. Use of these methods should improve the accuracy and statistical power of future comparative functional analyses.

Integrated characterization of function and taxonomy

As noted above, methods for characterizing both the microbiome's taxonomic profile and its functional capacity have advanced rapidly over the past few years. Yet, a remaining important challenge is the integrated analysis of these two aspects of the microbiome and the determination of which taxa provide which functions. Such information will not only allow us to fill in gaps in the availability of reference genomes but is also a crucial first step in the development and design of targeted microbiome manipulations that could modulate the community's function.

A simple approach to associate taxa with functional potential is to annotate reads (or partial assemblies) with both taxonomy and function using any of the methods discussed above. Determining the taxon of origin for the many reads in a metagenomic sample, however, can be both computationally expensive and methodologically challenging due to the short length of shotgun reads and varying distribution of taxonomy-distinguishing loci across genomes. To address the latter issue, early tools such as MEGAN¹³⁰ and MTR¹³¹ used a lowest common ancestor (LCA) approach that assigns a read the highest-resolution taxonomic classification that is shared by all sequences to which the read aligned. LCA classifications are clearly limited in resolution, leaving a large fraction of reads with only a course-grained taxonomic assignment or none at all¹³². To improve the precision of taxonomic assignment of shotgun metagenomic reads, several recently introduced tools have incorporated information on k-mer frequency profiles in reference genome databases, though how those profiles are used varies greatly between tools. LMAT¹³³ and Kraken¹³⁴ both assign taxonomy based on identified LCA taxa for k-mers in each query sequence. Other methods train models on the k-mer profiles associated with each taxon, using a variety of machine learning approaches including neural networks (TAC-ELM¹³⁵), naïve Bayes classifiers (RITA¹³⁶), or linear models-based methods¹³⁷. TAC-ELM also incorporates data on GC content and RITA combines BLAST-based reference alignments. Comparisons between Kraken and the linear model-based method above suggest that while exact k-mer matching methods like LMAT and Kraken are more accurate when query sequences originate from reference genomes, they may produce overly specific classifications for sequences from genomes absent from the reference database¹³⁷. Moreover, Kraken requires fairly long (31 amino acid) k-mer matches, which may potentially reject many short reads due to insufficient data. These observations suggest that exact k-mer matching methods are most appropriate when a metagenome is dominated by well-characterized taxa and consists of sufficiently long reads, whereas machine learning approaches are superior for samples with more novel or unclassified microbes.

A few alternative methods for taxonomic assignment of shotgun reads use more specialized techniques. MetAnnotate¹³⁸ first uses an HMM approach to functionally annotate metagenomic reads, then determines taxonomy based on comparisons to the homologs of the matching protein family. Notably, this approach combines both functional annotation and taxonomic assignment into a single pipeline. Another tool, MetaCluster-TA¹³⁹, partially assembles reads, clusters the resulting contigs, and then assigns the LCA taxonomy given cluster alignments to genomes. SeMeta¹⁴⁰ similarly groups reads that contain overlapping sequence, clusters those groups by k-mer profiles, and then assigns each cluster with a taxonomic classification using an LCA approach for representative reads. These clustering-based methods aim to leverage groups of reads to obtain a broader genomic context for taxonomic classification (in contrast to the k-mer approaches that classify single reads), but may still produce low-resolution or incorrect taxonomic assignments if clustering of reads is incorrect. Together, these novel techniques allow more detailed and accurate functional profiling of microbiome samples, which will ultimately aid in understanding the human microbiome's functional capacity, dynamics, and impact on the host.

Characterization of Other Microbiome Facets via Meta-Omic Assays

While deep genomic characterization of microbial communities has rapidly advanced our understanding of community structure and function, many community features cannot be captured by metagenomic assays. For example, the oral microbiome undergoes a dramatic shift in metabolism in response to carbohydrate consumption without any taxonomic group shifting substantially in abundance¹⁴¹. Likewise, communities with very different taxonomic profiles may in fact have similar functional metabolic profiles¹⁴². To study such processes in detail and to characterize these additional facets of the microbiome's activity, researchers utilize comprehensive ‘meta-omics’ technologies (including metatranscriptomics, metaproteomics, and metabolomics) that can systematically characterize community-wide gene expression, protein abundance, and metabolite concentration over time or in response to perturbations. In fact, multiple reviews have recently called for an integrative approach that combines and compares these ‘omic’ assays to identify and characterize the underlying biological mechanisms in the microbiome^143–146. However, analyzing each of these omic datasets presents substantial bioinformatic challenges that have only been partially addressed to date. As in metagenomics, accurate and high-resolution quantification of the measured elements and accounting for various regularities, biases, and dependencies in the data are key for realizing the full potential of these exciting high-throughput datasets. These meta-omic assays, and the unique challenges each one presents, are discussed below.

A metatranscriptomic assay generally involves reverse transcription and cDNA sequencing of RNA material isolated from a microbiome sample. Such measurements of gene expression at the community level can provide important information on how different species respond to each other and to environmental changes such as antibiotic treatment¹⁴⁷ or dietary perturbations¹⁴⁸. This technology was further used to characterize gene expression patterns in a diverse range of communities^148–152. A typical analysis of such metatranscriptomic data consists of transcript assembly, annotation with functional and/or taxonomic information, normalization, and testing for differential expression between sample groups. None of these processing and analysis steps is necessarily simple or straightforward. The assembly of metatranscriptomic data can be performed by any transcript assembler, but it may be useful to leverage reference information from an associated metagenome. For example, a recently developed method applied a de Bruijn graph-based approach to incorporate information on metagenome assembly quality and completeness to improve subsequent transcript assembly¹⁵³. Assembled transcripts can be annotated for taxonomy and function using any of the metagenome annotation tools described above. However, as in the case of metagenomic assembly, fully assembled transcripts may not always be easy to obtain or informative (although, as an alternative, an assembly-free metatranscriptome-specific annotation pipeline is also available¹⁵⁴). Moreover, a recent simulation study recommended that a reasonably unbiased analysis could be achieved by both assembling transcripts and including unassembled transcripts in subsequent clustering and annotation¹⁵⁵. Notably, even after the metatranscriptome has been processed and the number of reads associated with each gene and/or taxon has been calculated, evaluating and exploring such data is a daunting task due to the potentially thousands of taxa, each with thousands of expressed genes, that are represented by these data. To address this challenge, an interactive tool (termed Anvi'o) has been recently introduced, implementing several metatranscriptomic and metagenomic processing algorithms and producing clear visualizations of assemblies and profiles at the species-, gene-, contig-, and sample-level¹⁵⁶.

Statistically sound normalization and rigorous quantitative comparisons of such complex metatranscriptomic datasets are a further challenge. The abundance of reads from a given transcript in a metatranscriptome depends on multiple factors, including the expression level of that transcript in its resident species, the abundance of that species in the community, and various biases associated with RNA-Seq experiments (such as compositional bias and batch effects). Extensive simulation and evaluation of such RNA-Seq biases and the development of rigorous methods for addressing them have produced useful tools for analyzing single-organism RNA-Seq experiments and correcting RNA-Seq-associated biases^157–159, some of which have already been applied in the microbial community setting. In contrast, however, methods for differentiating between transcript abundance changes occurring due to gene regulation in a given taxon versus those occurring due to ecological shifts are still lacking and are an important area for future research.

Similarly, while metaproteomic assays present a powerful opportunity to understand protein-level regulation in complex communities, the analysis of such data presents a plethora of challenges, including both the traditional obstacles associated with proteomics-based experiments and additional complications associated with assaying a mixed community of microbes. Such studies generally use tandem mass spectrometry to quantify peptide fragments, and then identify the source proteins of each peptide by searching against a reference database of theoretical or previously collected spectra. Since a peptide typically cannot be identified unless it is found in the reference database, the choice of database and search parameters can have a substantial impact on the obtained results. Indeed, this effect was convincingly shown in a recent study comparing peptide identifications in a human intestinal metaproteomic dataset with a classic single-organism proteomic dataset using several different metaproteomic databases and search strategies¹⁶⁰. An efficient way to narrow the search space and identify uncharacterized proteins is to use a database of theoretical spectra constructed from associated metagenome sequencing reads to search the obtained peptides¹⁶¹. In a recent study, for example, a strain-resolved metagenome was used to analyze a longitudinal metaproteomic dataset from the gut of a preterm infant¹⁶². Further difficulty associated with analyzing community-wide proteomic data arises because a given peptide may match homologous proteins across multiple taxa. Pipasic is a recently developed tool that addresses this challenge by correcting for the amount of similarity in peptide sequences from different strains¹⁶³. Moreover, proteins at the community level display an enormous dynamic range of abundances, and it therefore cannot be reliably determined whether a peptide not detected in a given sample by an untargeted assay is indeed completely absent or present but at a very low abundance. This incompleteness restricts the utility of metaproteomics for community metabolism modeling, though this limitation may be ultimately mitigated by improving technology. As a promising example, one recent study was able to use metaproteomic data to construct and compare detailed metabolic models of two naphthalene-degrading bacterial communities¹⁶⁴.

Importantly, while genes and proteins vary across taxa, metabolites are, at least in principle, universal. Accordingly, in contrast to metatranscriptomics and metaproteomics, the processing and analysis of community-wide metabolomic data can rely on standard approaches for single-organism metabolomics with essentially no modifications. For untargeted mass spectrometry metabolomics, these analyses typically involve normalization and putative identification of metabolites by searching either for matches in a spectral library or for known compounds with matching mass and chromatographic elution profiles¹⁶⁵. The greater challenge, however, lies in the interpretation of these datasets, and in linking the observed variation in biomolecule abundances with other data on community structure and function. Statistical associations between disease, metabolite concentrations, and microbial species abundances have been observed in case-control studies of Crohn's disease, colorectal cancer, and C. dificile infection among other conditions, but the mechanistic nature of these links remains unclear^166–169. A few studies have further used metabolic pathway information to quantify the link between shifts in the metagenome and functionally related metabolome variation^170,171. Moreover, a recent study has introduced a novel computational framework, MIMOSA, for metabolic model-based integration of community taxonomic and metabolomic data and for evaluating whether variation in the metabolome can be explained mechanistically by variation in the community's taxonomic profile¹⁷². Such methods are crucial for gaining a principled, systems-level understanding of how changes in community ecology impact community metabolism and behavior.

Lastly, while not strictly an ‘omics’ assay, high-resolution imaging of microbial communities and the study of community spatial distributions is another area of rapid technology and bioinformatic development. Spatial factors can affect microbial community nutrient availability, communication, and biofilm formation, among other processes¹⁷³. Methods for quantifying the distribution of microbes in a community and relating it to associated omic data are therefore clearly needed. One increasingly popular technique is fluorescent in situ hybridization (FISH) with primers specific to various bacterial taxa of interest, combined with high-resolution microscopy^174,175. A recently developed tool, called BacSpace, systematically processes and analyzes such data by filtering out non-microbial fluorescence and calculating and aggregating distances between different microbial cells and environmental landmarks¹⁷⁴. Another approach to examine microbial biogeography on a larger scale involves mapping and visualizing many metabolomic and taxonomic profiles via a 3D model of a community site. This strategy has been applied to communities growing on solid culture¹⁷⁶, as well as to the human skin microbiome¹⁷⁷. Computational and quantitative methods along these lines are crucial for incorporating information of spatial heterogeneity into a more complete mechanistic and quantitative understanding of the microbiome.

Conclusion

The growing appreciation for the scientific and clinical importance of the human microbiome has given rise to an explosion of microbiome studies. These studies now routinely generate, assemble, and explore high-dimensional meta-omic data at an unprecedented scale. Above, we have broadly outlined the most common types of approaches and computational tools available for processing and analyzing such data, with emphasis on several areas in which increasingly higher resolution and precision can be gained from computational analysis of microbiome data (Figure 1). Fortunately, such tools are regularly distributed as open-source software that can be applied to datasets from a wide range of studies (Table 1). It is important to note, however, that these methods (and likely many other methods that will be developed to address these challenges in coming years) are ultimately limited by the large numbers of genes of unknown function and yet-uncharacterized taxa present in the microbiome. Developing efficient, cost-effective, and rigorous methods to demystify these hidden layers of microbiome diversity is therefore necessary to realize the full potential of microbiome research. Nevertheless, the resolution and scale of microbial community profiling will likely continue to improve with future technology development. These technologies will provide an increasingly more detailed view of the structure and function of the microbiome's subpopulations and even single cells across time and space, the behavior of such subpopulations, and the way they interact with one another and with the host. These advances will contribute to the growing field of personalized medicine, with applications ranging from precise identification of pathogenic strains for targeted treatment, through careful monitoring of dysbiotic microbial communities in disease, to personalized and rational design of microbiome manipulations.