Amplicon-based sequencing approaches

The 16S ribosomal RNA (rRNA) gene provides a highly suitable target for bacterial classification by DNA sequencing. A description of this method, as it applies to the skin microbiome, has recently been described in the Research Techniques Made Simple series (Jo, 2016). In brief, this region of the bacterial genome consists of conserved and hypervariable regions and, in particular for the skin microbiome, the V1-V3 region was found to yield accurate results for taxonomic classification (Meisel et al., 2016). The V4 primers that are commonly used in studying microbiota from the gastrointestinal tract (Caporaso et al., 2012) require some minor modifications to capture the highly prevalent and abundant skin microbe Cutibacterium acnes (Zeeuwen et al., 2017). For the analysis of fungal communities, regions of DNA between the 18S, 5.8S, and 28S rRNA genes, termed internal transcribed spacers (ITS), contain both hypervariable regions and conserved regions for taxonomic identification and primer annealing, respectively. The ITS sequence resides in a much broader phylogenetic population, and is thought of as a more “universal” barcode for fungi, but the variation also comes with less accuracy and specificity in taxonomic identification (Schoch et al., 2012). Additionally, fungi are less studied and thus, their phylogenetic placement through computational methods and expert-based curation of phylogenetic relationships are lacking, which can be a limiting factor.

Most investigators rely on institutional cores or commercial operations to perform the sequencing. Here, we will focus more on the computational “pipeline” of analysis, that starts with the input of raw sequence data and ends with statistical analysis and graphical representation of the microbial communities (Figure 1). The current recommended pipeline tools are QIIME2, mothur, and HmmUFOtu (Caporaso et al., 2010, Kuczynski et al., 2011, Schloss et al., 2009, Zheng et al., 2018). The first step in the pipeline is pre-processing in which sequencing errors are eliminated. The next step is grouping of DNA sequences into operational taxonomic units (OTUs). This grouping is based on similarity or sequences that are “close” based on a defined sequence distance metric (Rossello-Mora and Amann, 2001). As it is highly likely that all microbes within an environment are not known, OTUs have become the standard for cataloguing the microbiome (Rossello-Mora and Amann, 2001, Schloss and Handelsman, 2005).

This process of ‘OTU picking’ can be achieved in two ways; by matching sample sequences to a database of reference sequence (such as Greengenes (DeSantis et al., 2006)) or alternatively, the sequences can be clustered into de novo OTUs using no references. Once clustered, these OTUs are mapped to known sequences to determine the taxonomic composition of the sample (Caporaso et al., 2010, Zheng et al., 2018). At this stage in the pipeline, the output is an OTU table (formatted as a delimited text file or BIOM file). This table includes all OTUs identified, their “abundance,” or number of reads, found in each sample, and usually the taxonomy assigned to each OTU. The level of taxonomic classification varies in accuracy, and is dependent on the region of the 16S rRNA gene sequenced and the identify of microbes in the sample. Results given at the species level should be interpreted cautiously unless customized (eg. skin-specific) databases are being used (Conlan et al., 2012, Meisel et al., 2016).

Once taxonomic assignment is complete the typical next step in the pipeline is to examine the diversity of the microbiome both within and between different samples, termed alpha and beta diversity, respectively. Most of the tools for this analysis were developed by the field of ecology and have been adapted to microbial community ecology. The pipeline tools QIIME2 or mothur have built-in plugins or programs to perform these analyses directly, but many alternative tools exist particularly for users with statistical and bioinformatics backgrounds. A large collection of these downstream tools, such as those in the “vegan package” (Oksanen et al., 2018), can be installed into an R environment. R is an open-source computer language/environment designed for statistical analyses and graphical presentation of data (Team, 2017). A widely used R tool is phyloseq, which offers an intuitive suite of functions to aggregate data, perform statistical analysis, and graph the results (McMurdie and Holmes, 2013).

Statistical analysis & graphical presentation

Typically, microbiome data is non-parametric; the distribution of data (OTUs) are unknown and thus assumptions about the distribution should not be made when selecting statistical tests. Consequently, non-parametric statistics need to be employed. For example, in place of t-tests, an appropriate choice is the Mann-Whitney/Wilxocon rank-sums test. Instead of applying an ANOVA test across >2 groups, the Kruskal-Wallis one-way analysis of variance test can be used, and the Spearman rank correlation coefficient should be used rather than the Pearson when examining co-occurrence of OTUs and/or taxa. Additionally, microbiome data is inherently multidimensional, and thus requires specialized tools. One of these is UniFrac (Lozupone and Knight, 2005), which utilizes a distance matrix that incorporates phylogenetic distances in comparing dissimilarity of microbial communities between two or more samples (Beta-diversity).

Multi-dimensional data can be challenging to display visually; three-dimensional graphs can be difficult to interpret and four-dimensions and above cannot be drawn. To overcome this, a procedure termed principal components analysis (PCA) can be utilized. This is a statistical technique that transforms large sets of observations, into a set of uncorrelated variables termed principle components, that emphasize the major differences in the data. The first principle component has the largest variation in the data, the second principle component has the next largest variation and is unrelated to the first. The first and second principle components are then plotted as a two-dimensional graph. Other methods that perform this task are non-metric multidimensional scaling (nMDS) of which principal coordinates analysis (PCoA) is a subtype. The details of these methods are beyond the scope of this review, but they do permit statistical analysis to be performed on the data in the form of a permutational multivariate ANOVA, or PERMANOVA. These statistical and graphical tools are available in the aforementioned vegan R package. Many other statistical methods have been adopted for more specific analyses and graphing of microbiome data, including defining community types through Dirichlet multinomial clustering and identifying “biomarkers” by testing multiple decision tree models, in a process known as random forests, both of which can be performed in the R environment.

Shotgun metagenomic sequencing

Shotgun metagenomic sequencing, or the untargeted sequencing of all microbial genomes present in a specimen, is considerably richer in providing information than amplicon-based profiling approaches. Unlike amplicon-based sequencing, where specific primers are targeted to regions of ribosomal RNA genes, DNA is prepared for shotgun metagenomics by random fragmentation, addition of barcoded sequencing tags, and limited cycle amplification (Figure 1). Since shotgun metagenomics captures a greater variety of gene content in a sample, multi-kingdom compositions at strain-level resolution (as depicted in Figure 2, adapted from (Oh et al., 2014), as well as functional profiles for communities are captured. Shotgun metagenomics have provided key insights into the skin microbiome in atopic dermatitis, including the role of strain-level variation of Staphylococcus aureus (Byrd et al., 2017), and mechanistic understanding of how microbial metabolic pathways are altered to enhance ammonia production and increase skin pH (Chng et al., 2016).

Two different analytical approaches are used for shotgun metagenomic datasets: assembly-based and read-based profiling (for a comprehensive discussion, the authors recommend (Quince et al., 2017). While read-based, assembly-free profiling is faster and mitigates issues with assembly, it relies upon reference genomes at the expense of uncharacterized microbes that have no references available. A popular tool to generate taxonomic profiles without assembly is MetaPhlan, which maps shotgun reads to reference marker genes (Segata et al., 2012). These data may then be used to derive the alpha and beta diversity metrics previously described. Functional profiles can be produced using the HUMAnN tool (Abubucker et al., 2012) or similar, that takes the DNA reads and maps them against universal gene-protein databases. This allows identification of the proteins encoded by the DNA and functional pathway linkage of the proteins.

This work was supported by awards from the National Institutes of Health (R01-NR-015639, R01-AR-006663 and R00-AR-060873 to E.A.G.), Burroughs Wellcome Fund PATH Award (E.A.G.), and the Linda Pechenik Montague Investigator Award (E.A.G.). This research was also supported by the Penn Skin Biology and Disease Resource-based Center (P30-AR-069589). C. B.-M. and L.F. were supported by the Dermatology Research Training Grant, T32-AR-007465.