Normalization of compositional data/making it comparable across studies
A major challenge in the analysis and interpretation of amplicon sequencing
data remains the relative nature of the data, which may not reflect actual microbial abundances (25). Numerous
normalization approaches have been considered to account for differences
in rRNA gene counts across taxa
(26). Data
normalization remains even more critical in amplicon sequencing of
eukaryotes. At each step of the sample handling and sequencing process,
a subset of the sample is selected, and additional bias has the
potential to be introduced
(27). Numerous tools for statistical analysis have been introduced to circumvent challenges associated with highly compositional relative abundance data. We here suggest such data-driven approaches to address the concerns of normalization, false-discovery rates and the compositional nature of sequencing.
One of the first approaches to analyze amplicon sequencing data is to
remove potential sequencing errors. Doing so contributes to the
elimination of chimeras and other sequencing artifacts that tend to
falsely boost diversity levels \cite{Edgar2011,Haas2011}. The use
of amplicon sequencing variants (ASVs), instead of operational taxonomic
units (OTUs) attempts to overcome this issue by assigning a greater
probability of a true biological sequence being more abundant than an
error-containing sequence
(30). To that end,
bioinformatic tools such as DADA2
(31) and Deblur
(32) attempt to
use sequencing error profiles to resolve amplicon sequencing data into
ASVs. Furthermore, even though there are some caveats associated with
the use of ASVs that might require previous considerations, they have an
intrinsic biological meaning as a DNA sequence, as opposed to OTUs.
Additionally, they make the merging of datasets possible, even when the
sequencing primer pairs are different
(30).
Another relevant step when analyzing sequencing data is to account for
the different sequencing effort across samples (i.e. different library
sizes) that can result in a substantially different number of recovered
reads even within sample replicates. Ways to tackle this issue include
total library size normalization and rarefaction, although recent
literature has advised against the latter
(33).
Bioinformatic tools such as DeSeq2 and EdgeR that were originally built
for differential gene expression analyses of RNA-seq data, now extend to
amplicon-based studies. These packages provide ways to normalize count tables using
the “relative log expression” (RLE) and the “Trimmed Mean of
M-values” (TMM) normalization approaches respectively
(34, 35). Both
methods are applied on a raw or a low-abundance filtered count table and
have performed well in both real and simulated datasets and outperform
rarefaction-based approaches
(33). Other
alternatives to account for the compositional aspect of sequencing data
include center log (CLR), isometric log (ILR) or additive log (ALR)
ratios transformations of a count data matrix
(36, 37).
After data normalization, traditional amplicon sequencing data analyses
include the generation of distance matrices for ordination, clustering,
and variance partitioning analyses. Commonly used distance metrics
include Bray-Curtis, Jaccard and Unifrac (weighted and unweighted) that
– regardless of their value in other fields - also do not take into
account the compositional nature of sequencing data. The Aitchison
distance - defined as the Euclidian distance on top of a center-log
transformed count matrix – is a viable compositional alternative
(36) on top of
which ordinations (e.g. PCA biplots) can be performed. Additionally, the
Philr transform metric has been introduced as compositional alternative
to the weighted Unifrac, that carries phylogenetic information
(38). Most of the
above mentioned compositional options are implemented in R packages and
include publicly available tutorials.
As a consequence of all the above-mentioned limitations, we recommend a critical evaluation of the different data analyses tools in light of the intrinsic nature of each experimental setup (see section “Ecological interpretations from amplicon sequencing data”).
Data-driven approaches to more quantitative sequencing studies
Or: Experimental approaches to more quantitative sequencing studies
In addition to the limitations imposed by sequencing technology and the
nature of the sequencing data, another aspect that prevents amplicon
sequencing data analyses from being fully quantitative is the potential
multiple copies of marker genes per organism. The 16S rRNA gene copy
number per microbial cell can vary between 1-18 and can additionally
show variation within different strains of the same species
(39, 40;
Lavrinienko et al., 2020). Therefore, relying solely on the number and
diversity of 16S rRNA gene sequences can lead to inaccurate estimates of
abundance and diversity of microbial communities. Correcting for 16S
rRNA gene copy numbers in sequencing surveys remains challenging,
particularly for soil. Introducing an internal spike-in can be a useful
tool towards more quantitative amplicon data analyses and there are a
few studies that applied this technique in soil
(41, 42). There
are however, important considerations: i) the choice of spike should
neither fall on members of the existing microbial community, nor it
should be in concentrations that would shift the sequencing effort
towards it; ii) the timing of addition of the spike (before or after
nucleic acids extractions) will dictate the kind of retrieved
information: while adding the spike after extraction can provide good
estimates of sequencing biases, it does not take extraction efficiency
into account (42).
A recent study combined amplicon sequencing, a synthetic DNA spike of
known concentration on the samples prior extraction, and qPCR
quantifications to back calculate the number of copies before extraction
after taking into account the extraction yield. The ratio of each OTU
against the initial concentration of 16S rRNA genes was used to
calculate more accurate abundance levels of each OTU after taking
extraction efficiency into account
(43). As a
consequence of all the above-mentioned limitations, we recommend a
critical evaluation of the different data analyses tools in light of the
intrinsic nature of each experimental setup (see section “Ecological
interpretations from amplicon sequencing data”).