The first step, once the samples are collected, is the extraction of total DNA, which will then be used as template for PCR amplification of the marker gene. After the DNA is extracted and purified, the workflow for most 16S studies becomes roughly the same. Fecal samples are composed of microorganisms that differ in characteristics, such as size and cell wall composition, and that are present in different proportions. This can make the purification of a DNA sample that accurately represents the original community (i.e., that keeps all species and their abundances at the same relative proportions) a challenge, as different sample handling and DNA extraction protocols can yield samples with different bacterial ratios. It has been shown, for instance, that frozen fecal samples yield a higher amount of DNA from Gram-positive than from Gram-negative bacteria, probably due to the effect that the freeze–thaw cycle can have over the Gram-positive cell wall (46).

Differences in gut microbial community patterns can also arise due to the principles of the genetic material extraction protocols, causing the over or underrepresentation of the same microbial group in DNA extracted from subsamples of the same source (47). Some DNA extraction kits use bead-containing lysing matrices and vigorous shaking steps that contribute to the disruption of the cell wall, whereas others rely on chemical lysis (48). Several studies have consistently demonstrated that protocols that involve a bead-beating step yield higher quantities of bacterial DNA, and, most importantly, these samples tend to be a more comprehensive representation of the microbial community, regardless of the source material and analysis method (49–51). The differences between subsamples extracted with different kits can even be statistically significant, which is why it has been suggested that data from studies using different extraction methods should not be compared (52). Opportunely, studies are increasingly using similar DNA extraction protocols. For instance, the PowerSoil^® DNA isolation kit (MoBio) has become popular because it performs well in a wide variety of samples, including human feces. Although using the same extraction protocol does not guarantee accurate representation of the microbial community under study, it allows comparison among studies.

Another issue with DNA extraction is that, due to the non-specificity of marker gene and metagenomic assays, they are highly sensitive to contamination with foreign microbial DNA. The presence of bacterial DNA from sources other than the original sample can alter the outcome of the analysis in a way that it no longer mirrors the original community it is supposed to reflect. Contamination sources may include the PCR reagents (53, 54), ultra pure water (55, 56), and, even, the DNA extraction reagents (57, 58). The genetic material extracted from samples with low biomass is more prone to being drowned by contaminant DNA (59, 60), and the contamination profile varies between laboratories, extraction kits, and batches from the same kit (60). Procedures to reduce the effect of contamination include the maximization of starting biomass from which DNA is extracted, the randomization of the order in which samples are to be processed, the collection, processing and sequencing of technical controls of the reagents to be used (storage media, DNA extraction kits, and PCR kits), the recording of the kit lots as additional metadata, and the quantification of negative-control sequences (60).

Today, there is no standard procedure on how to deal with sequences showing up in technical controls. One suggestion would be to compare the abundance in real samples and controls: if an OTU has similar relative abundance in samples and controls, it is likely a contaminant; otherwise, it probably is not. This approach has the drawback that the threshold in which the abundance of an OTU is considered a contaminant is subjective (61). Another method involves the removal of OTUs whose abundance is negatively correlated with amplicon concentration, as it is assumed that the signal from contaminant sequences in low biomass samples is less likely to be drowned by the signal of real data (61). In any case, it is necessary to be aware of taxa that are present in negative controls, taxa statistically associated with a particular batch of reagents, and taxa biologically unexpected in the treated samples.

In marker gene studies, total DNA is used as template for the PCR amplification of the barcode region. As in single-template PCR, the efficiency of multi-template PCR is influenced by the GC content of the target region (62), the DNA concentration (63), and the thermocycling conditions (64). However, because of the multiorganismal origin of the gut microbiome, a series of particular difficulties and artifacts, such as primer mismatches, gene copy number variation (CNV), chimeras, heteroduplex, and skewed template-to-product ratios, are encountered and can distort the diversity measures. Primer selection, CNV normalization, and chimeric sequence removal are discussed below; for a detailed discussion of reagents and PCR conditions in multi-template assays, see Ref. (65).

Next-generation sequencing technologies refer to various strategies that rely on a combination of template preparation, sequencing and imaging, and genome alignment and assembly methods (83). The major advance offered by NGS is the ability to produce an enormous volume of data cheaply and fast. The transition from Sanger to NGS has opened new horizons in the gut-microbiome field by making it possible to collect millions of sequences, spanning hundreds of samples (80). A good example of this is the Human Microbiome Project, which used NGS to characterize the diversity of bacteria, archaea, and viruses that inhabit various areas of the human body in several hundreds healthy individuals (84). In the last decade, the throughput of NGS technologies has dramatically increased, and the operation cost has reduced, which, in turn, has boosted its use in microbial studies. However, the major drawback of all NGS technologies is that they raise concerns regarding the quality of data.

When sequencing genomes, multiple reads are used to construct a consensus and the error rate, defined as the number of errors per total base call (25), is, thus, reduced since each nucleotide in the original sequence is called several times by different reads. Such approach cannot be used when sequencing marker gene amplicons, such as the 16S rRNA, because each individual read is considered an identifier of an independent organism (e.g., a bacterium), and it is not possible to assemble the amplicon sequences (34); hence, the reduction of the error rate by other means becomes imperative.

One strategy to determine how many errors are introduced at each NGS run consists of sequencing a synthetic mixture of genomic DNA (mock community), comprising several known bacterial species, along with the samples. Reads are compared with a reference database of the marker gene, and errors are identified in pairwise alignments of each experimentally generated sequence relative to the closest reference sequence (25, 32, 34). Sequencing mock communities to assess the error rate of each individual amplicon sequencing run should become a standard step in microbial community analysis (see http://www.hmpdacc.org/HMMC/) (25, 29).

Currently, Roche 454 GS-FLX, Illumina MiSeq, Ion Torrent PGM, and PacBio SMRT are the most used platforms for the study of the gut microbiome (35, 85–87). However, each technology performs differently in the trade-off between read length, sequence throughput, and error rate (Table 1). As mentioned above, since hypervariable regions correlate differently with the whole 16S rRNA gene (88, 89), it is arguably better to sequence shorter reads at greater depths and with lower error rates (e.g., Illumina, Ion Torrent) than longer reads with higher error rates (e.g., PacBio) (34). The former allows the detection of low-abundance microorganisms (90, 91) and the avoidance of unnecessary greater computing times due to the description of non-existent organisms caused by artifactual sequences. Although increased read length usually improves classification, platforms such as PacBio are currently limited by their high sequencing error and low yield of sequencing data relative to the other platforms (34).

Up to this point, procedures in the archetypical workflow described in Figure 1 take place in vitro. Hereafter, treatment of raw DNA sequences occurs in silico. To reduce sequencing error rates, it has become mandatory to apply stringent sequence curation and denoising algorithms. Inadequate cleaning of reads can have many negative effects including limited ability to identify chimeras and inflation of α and β diversity metrics (92). Low-quality sequences, artifacts, and contamination can compromise the downstream analyses and, thus, must be removed from the dataset.

The first step is the removal of reads with ambiguous base calls (N) in the barcode or in the marker gene amplicon, as it is not possible to determine the true nucleotide sequence (93). On the other hand, mismatches in the primers and barcodes are usually allowed up to a certain number; the removal of sequences with less than three mismatches has little effect on the reduction of the error rate (93). Emulsion-PCR-based platforms (e.g., 454, Ion Torrent) are known for producing homopolymer-associated indel errors (33); these artifacts have been shown to account for a large proportion of errors in benchmark studies using mock communities and to be associated with low-quality scores (92). Therefore, reads with homopolymers longer than eight nucleotides should be culled (25).

In addition, in most sequencing platforms (e.g., 454, Illumina, Ion Torrent), quality scores reduce in a lengthwise fashion, and it is possible to identify breakpoints where the quality criteria are not met. Sequences can be trimmed to those breakpoints to reduce the overall error rate. Two trimming approaches have been widely used: a “hard cutoff” method trims the sequences at the first nucleotide with a quality score below a given threshold (94); this minimizes the error rate but also reduces the average sequence length. Another method, called “sliding window,” calculates the average quality score within a sequence window (or substring) and trims when the average quality score within that window drops below a threshold; the latter method has the advantage that reduces the overall error rate without reducing the average sequence length (25). Reads with anomalous lengths (well above or below the expected value for a given technology) are also removed, as they likely represent PCR or sequencing errors, or become not informative as a result of the quality trimming (93).

The use of a pre-clustering algorithm has also been shown to reduce the number of sequences that are the result of sequencing errors and to predict with higher accuracy the number of expected OTUs in template preparations of known taxonomic composition (95). It assumes that rare sequences are more likely to derive from abundant sequences and can, therefore, be merged if they are within a specified similarity threshold. This threshold must always be lower than the value used for OTU clustering, usually 1% (25).

Also, contaminant sequences must be removed from the dataset. Due to the nature of the 16S rRNA gene, mitochondria, chloroplast (96), and other eukaryotic sequences are likely to be amplified and should be identified and discarded, along with sequences unclassified at the domain level; according to the scope of the study and the primers used, bacterial or archaeal sequences would also be needed to get removed.

Sequences composed of two or more parents are named chimeras. Chimeras are a serious concern in studies of the gut microbiome because they can lead to the description of non-existent organisms and inflate diversity metrics. This kind of artifact arises from errors during PCR, and several factors influence its appearance, such as DNA damage (97), the amplification of highly similar sequences (98), a high number of cycles, and short elongation times (99). This suggests that prematurely terminated amplicons that anneal to a homologous template to prime the next PCR cycle are likely to be the major cause of chimera formation.

The detection of chimeras in libraries of 16S amplicons is particularly challenging, as sequences are short and highly similar. There are multiple algorithms designed to detect and remove chimeric sequences (100–107), which follow the same basic principle: substrings or fragments of the query sequence are compared to a set of reference sequences in order to establish if the said substrings match different references. Once a chimera is identified, it is removed from the dataset. Some algorithms use allegedly chimera-free 16S sequence databases as reference, including Chimera Slayer (105) and DECIPHER (108). Others [e.g., Perseus (106), UCHIME (107)] use a database-free approach that assumes that the most abundant sequences from the query dataset are unlikely to be chimeric and can, therefore, be used as reference. Database dependency influences the ability of different algorithms to identify and remove chimeras (109). Database-independent algorithms have the advantage of being able to detect them even if the studied community is poorly described (25). In contrast, database-dependent algorithms rely on reference collections that only contain gene sequences from cultured bacteria and are not expected to perform as well on samples that contain sequences from yet uncultured organisms (24), something very common in studies of the gut microbiome. Thus, the use of algorithms that do not rely on databases should be preferred in order to minimize the inflation of diversity caused by chimeras, especially when dealing with poorly characterized gut microbial communities.

A problem that arises when studying the gut microbiome is the difference in the number of copies of the 16S rRNA gene among species, which can range from a single copy up to 15 (134). This variation can lead to erroneous abundance assessment; at equal number of cells, taxa with few copies of the 16S rRNA gene have lower amplicon counts than taxa with more copies of the gene. Therefore, CNV can result in over or underestimation of microbial abundance. CNV has not deserved full attention; yet, it is of utmost importance since it can result in a biased description of the microbial community. Indeed, it has been suggested that bacterial diversity could be overestimated by a factor of 3 due to 16S CNV (135).

In microorganisms with known 16S rRNA gene copy number, CNV could be corrected by weighting read counts by the inverse of its gene copy number. However, the problem is more difficult to deal with in cases where the gene copy number is unknown. A possible solution in these cases is to use the value of a closely related organism (136). Another possibility is to place 16S reads on a phylogenetic tree and calculate gene copy number using phylogenetically independent contrasts (137, 138). While these methods have been shown to improve the measures of diversity and abundance of microbial communities, they rely on databases of 16S and sequenced genomes, which, as with phylotype-based clustering, lack information of uncultured and poorly cultured organisms. In cases of poorly studied deep evolutionary lineages (say, rare phyla), CNV correction is definitely an unsolved issue.

Although CNV can move away estimates of diversity from reality, it must be noted that researchers usually want to compare these estimates between treatments (e.g., obese vs. lean, vaginal delivery vs. C-section, probiotic vs. placebo). In other words, we look for relative changes in the abundance of OTUs A, B, and C; even if they would be badly estimated due to the assumption that they only have one 16S rRNA gene each, what is important is to see how populations change under different tested conditions. The take-home message from CNV is that we should emphasize more comparisons of the same OTU among samples than comparisons among OTUs within samples.

In order to guarantee reproducibility of results, we encourage researchers and journals to explicitly include and require the following technical information in microbiome publications: (I) DNA extraction method, including the type of extraction kit if one was used and modifications to the standard protocol proposed by the manufacturer; (II) description of how DNA contamination was controlled for (e.g., DNA extraction of negative controls); (III) 16S rRNA hypervariable region targeted including the nucleotide sequences of the primers used; (IV) sequencing technology employed; (V) description of sequencing error-rate assessment (e.g., was a mock community sequenced in parallel with the samples?); and (VI) description of in silico analyses (culling of dubious sequences, removal of chimeras, OTU clustering and taxonomy assignment, copy number variation correction), including the code or command lines with parameters used if appropriate.