Data pre-processing and quality checking.

Data analysis using CoMA requires input files in FASTQ format (paired-end or single-end, either uncompressed or compressed with gzip), which is the standard NGS file format provided by all common sequencing platforms. In case of paired-end reads, forward and reverse reads are first merged at their overlapping region, using the open-source tool “Paired-end assembler for DNA sequences” (PANDAseq [52]). This step is followed by a QC step (an individual quality report is created for each sample) and a subsequent sequence selection step, both accomplished by the “Preprocessing and Information of Sequence data” (PRINSEQ [53]) utility. Amplification primers, sequencing adapters and multiplex identifiers are removed (trimming) and reads are selected based on Phred quality score (PQS), amplicon length and ambiguous bases. The selection parameters are different for each specific primer pair and can be adjusted based on the information gained from QC. A second QC step, checking the success of trimming and sequence selection, finally completes the data pre-processing workflow.

OTU clustering and taxonomic classification.

All steps within this program section are either part of the “Less OTU scripts” (LotuS [39]) pipeline or the “simple demultiplexer” tool (sdm [39]). First, an error-correction step is applied, followed by the removal of artificial chimeric sequences, resulting from cross-hybridization of DNA fragments during library preparation. Subsequently, clustering to OTUs is done with the UPARSE-OTU algorithm [54] at a 97% similarity level, following a de-novo assembling approach. OTU sequences are then aligned/classified with a desired gene reference database using either “Basic Local Alignment Search Tool” (BLAST [55]), “Local Aligner for Massive Biological Data” (Lambda [56]) or the “Ribosomal database project” (RDP [57, 58]) classifier. CoMA supports several popular taxonomic databases, including SILVA, Greengenes, RDP, UNITE, HITdb [59], beetax [60], and PR2, as well as any custom database provided by the user. A custom database needs to be composed of two individual files: a FASTA file with all sequences and a TAX file with the accompanying taxonomic information. The correct structure for both files is provided in the CoMA manual. It is possible to use multiple databases, where the first serves as primary- and the other(s) as backup database(s). OTUs that cannot be assigned to the primary database are then diverted to the backup database(s). This workflow section is completed by the generation of universally applicable output files (tab-delimited OTU-table, BIOM file, NEWICK tree), allowing for specific subsequent data analysis or individual visualization steps by the user if desired.

Post-processing of data.

Data post-processing starts with an initial removal step, where very rare OTUs can be dropped from analysis (filter_otus_from_otu_table.py, QIIME). OTUs can be omitted due to a low number of total reads on the one hand or due to rare occurrence within the samples on the other hand. To assess whether the sequencing effort (i.e. number of reads per sample) was sufficient, rarefaction curves are computed using Mothur’s rarefaction.single command. The applied metric can be selected by the user out of the following possibilities: OTUs, Chao1 richness estimator, Shannon-Wiener diversity index, Simpson’s diversity index and Good’s coverage for OTUs. The output is provided either as text file containing all raw data or as line plot. Based on rarefaction curve analysis, the dataset can be subsampled to a unique read count for each sample to bring the data onto a common scale [61]. This step is done using the single_rarefaction.py script provided by the QIIME platform. The pseudo-random number generator used here for subsampling is an implementation of the Mersenne twister [62]. To facilitate use, CoMA provides the read count for each sample, sorted in increasing order. The next step allows for renaming sample names; all further results are presented using the newly assigned names. Finally, metadata can be assigned to the samples. These metadata can be used in all upcoming steps in order to group the samples.

Data visualization and statistical appraisal.

First, summary files, containing the most abundant taxa of each sample at all taxonomic levels (kingdom, phylum, class, order, family, genus, and species) are created. The user can compute a report for bacteria, archaea, fungi and eukaryotes, as well as a general summary. Moreover, a summary report for a specific taxon can be created. During the next step, publication-ready and highly esthetic bar charts and heatmaps can be created for each taxonomic level. The user can choose between plots for bacteria, archaea, fungi, eukaryotes and prokaryotes. Alternatively, plots can be computed for a specific taxon, including all descendent taxa at the lower taxonomic levels. In addition, this step allows the user to set a threshold, based on relative abundance, for taxa to be included in the plots and a decision whether unassigned taxa will be included or excluded from the depictions. Moreover, the user can choose between various popular image file formats (EPS, JPEG, PDF, PNG, PS, RAW, SVG, TIFF) and adjust the pixel density if a raster graphic format was selected. Subsequently, Venn diagrams can be created to compare detected taxa among given groups for each taxonomic level. Within the next step, alpha diversity (or within-sample diversity) can be calculated using various different calculators (observed OTUs, Shannon-Wiener index, Simpson’s index, Pielou’s evenness index, Good’s coverage of counts, Chao1 richness estimator, Faith’s phylogenetic diversity). Every metric has different strengths and limitations—technical discussion of each metric is readily available online and in ecology textbooks. The output is provided either as raw-data text file or as publication-ready bar chart. Finally, CoMA offers two different approaches for analyzing beta diversity: ordination and hierarchical cluster analysis. Ordination can be done using various different metrics (Minkowski, Euclidean, Manhattan, Cosine, Jaccard, Canberra, Chebyshev, Braycurtis, Dice, Weighted UniFrac, Unweighted UniFrac). After calculating of a distance matrix, data are visualized in either two- or three-dimensional Principal coordinates analysis (PCoA) plots. To quantify the strength of the grouping, statistical tests (ANOSIM, PERMANOVA; 999 permutations) are applied. For cluster analysis, the user can choose between the following linkage methods for the cluster analysis: Single, Complete, Average (UPGMA), Weighted (WPGMA), Centroid (UPGMC), Median (WPGMC) and Ward. These methods are used to compute the distance d(s,t) between two clusters s and t, using a bottom-up approach. In addition, the user can select a suitable metric to determine the distance between two individual data points: Euclidean, Cosine, Cityblock, Correlation, Jaccard, Braycurtis and Dice. Results from cluster analysis are presented as dendrogram plots.

Mock data analysis with CoMA.

Mock-13 forward and reverse reads were merged in a first step and the quality of the joined reads was then determined. Based on QC analysis, reads were selected due to their sequence length (240–265 bp) and their mean PQS (> 15). The reads were not trimmed since primers and other sequence appendices were already cut off in the raw input files. After the second QC step, reads were clustered into OTUs (> 97% similarity) and representative sequences were aligned to the SILVA database (SSU 132) using BLAST. The same procedure was applied for mock-16, however with two minor modifications. Before quality selection, primers were trimmed off from the forward (19 bp) and reverse end (20 bp). In addition, the PQS selection was more rigorous (> 30) since the overall quality of the reads was higher in mock-16 compared with mock-13.

A slightly different protocol was applied for mock-26 since the dataset was still multiplexed and contained single-end reads. Demultiplexing was done with the “demux emp-single” command within the QIIME2 toolkit since CoMA does not support multiplexed input files in its current version. Selected files (Mock.81, Mock.92, Mock.99) were then subjected to the CoMA pipeline, starting with the first QC step. Reads with a maximum length of 500 bp were then selected. A minimum sequence length was not adjusted since ITS reads typically show a wide distribution in sequence length and thus a loss of information should be avoided. Since the overall quality level of the sequences was high, a minimum average PQS > 30 was adjusted. After the second QC step and the OTU clustering, reads were aligned to the UNITE database (version 8) using the BLAST algorithm.

Mock data analysis with Mothur.

Data analysis of mock-13 and mock-16 with Mothur (v. 1.39.5) was done according to the standard operating procedure (SOP) for Illumina MiSeq data, provided at the Mothur homepage (https://www.mothur.org/ wiki/MiSeq_SOP, [32]). Input forward and reverse reads were merged with “make.contigs”. After checking the quality with “summary.seqs”, ambiguous bases and reads with a length of > 270 bp (mock-13) and > 292 bp (mock-16) were removed using “screen.seqs”. A table of unique sequences was created with “unique.seqs” and a prevailing counting table, providing the occurrences of the unique sequences in each sample, was created using “count.seqs”. To prepare the SILVA (version 132) database, “pcr.seqs” was applied on the database file with the following limit settings: start = 13,862, end = 23,444. Unique sequences were aligned to the prepared database using the “align.seqs” tool and outliers were removed with “screen.seqs” (start = 8, end = 9,578 (mock-13) or 9,582 (mock-16), maxhomop = 8). After a sequence filtering step with “filter.seqs” (vertical = T, trump =.), reads were again checked for unique sequences using the “unique.seqs” tool to exclude potential sequence redundancies introduced in course of the trimming step. For de-noising of data, sequences were pre-clustered with “pre.cluster” (diffs = 2) and chimeras were removed with “chimera.uchime” and a subsequent “remove.seqs” step. To get rid of sequencing errors, sequences were classified with “classify.seqs” and unknown reads (at kingdom level) as well as reads assigned to chloroplasts, mitochondria or eukaryota were removed with “remove.lineage”. Sequences were then clustered using “cluster.split” (splitmethod = classify, taxlevel = 4, cutoff = 0.03) and a shared file was created with “make.shared”. After taxonomic assignment (“classify.otu”), a BIOM file was created (“make.biom”) and then transformed to a tab-delimited OTU-table using the independent “biom-convert” tool (http://biom-format.org/).

For mock-26, selected files (Mock.81, Mock.92, Mock.99) were imported using “make.file” and “fastq.info”. Low-quality reads were then excluded (“trim.seqs”; qaverage = 15) and all individual files were merged to a combined FASTA file (“merge.files”, “make.group”). After summarizing all data with “summary.seqs”, reads with > 10 ambiguous bases were removed. A greater tolerance of ambiguous bases compared with mock-13/mock-16 was required since all sequences constantly contained a considerable number of ambiguous bases. A sequence length selection was omitted due to the high inconsistency among the ITS data. After finding the unique sequences (“unique.seqs”) and creating a count table (“count.seqs”), analysis was continued as described for the other mock communities, starting with the pre clustering step (“pre.cluster”). The newest Mothur-optimized UNITE release (version 7) was used as reference database for sequence alignment.

Mock data analysis with QIIME.

Data analysis of mock-13 and mock-16 with QIIME (v. 1.8.0) was done according to the recommendations for Illumina data on the QIIME homepage (http://qiime.org/tutorials/processing_illumina_data.html, [33]). First, paired-end input reads (all files in FASTQ format) were merged using the “multiple_join_paired_ends.py” script within the QIIME toolkit. This is an alternative version of the “join_paired_ends.py” script and allows the merging of already-demultiplexed FASTQ files. Quality filtering was done with the “multiple_split_libraries_fastq.py” script and chimeras were removed with “usearch61” using the “identify_chimeric_seqs.py” and “filter_fasta.py” scripts. Reads were then clustered into OTUs with the “usearch61” method of “pick_otus.py” and representative (= most abundant) sequences for each OTU were picked using “pick_rep_set.py”. Representatives were then assigned to the QIIME-optimized Silva SSU database (release 132) with the “assign_taxonomy.py” script (assignment method: uclust) and the OTU-table was constructed using “make_otu_table.py”. The created BIOM file was converted to a tab-delimited OTU-table with the external “biom-convert” tool as described in the previous chapter.

Selected input files for mock-26 (Mock.81, Mock.92, Mock.99) underwent, in a first step, a quality selection (“split_libraries_fastq.py”), where reads with an average PQS < 15 were removed. All upcoming steps were done as described above for the other mock communities. The newest QIIME-optimized UNITE release (version 8, dynamic construction) served as reference database for sequence alignment (“assign_taxonomy.py”).

Mock data analysis with QIIME2.

Data analysis with QIIME2 (v. 2019.10) followed the “Moving pictures” tutorial, accessible on the program’s webpage (https://docs.qiime2.org/2019.10/tutorials/moving-pictures/). However, mock-13 and mock-16 required a slightly alternative procedure due to their paired-end data structure, and the files from all mock communities (mock-13, mock-16, mock-26) needed to be imported in a different manner since the data were already demultiplexed.

First, mock-13 and mock-16 input files were imported using “qiime tools import” with the settings “SampleData[PairedEndSequencesWithQuality]” as data type and “CasavaOneEightSingleLanePer- SampleDirFmt” as input format. After summarizing and quality-checking the imported data (“qiime demux summarize”), “qiime dada2 denoise-paired” was applied. This central step calls the DADA2 pipeline and provides several functions, including sequence trimming and truncation, quality filtering, removal of phiX reads, chimera removal, and the construction of a feature table (suitable filtering settings resulted from the previous quality-checking step). In contrast to all other pipelines used for this study, DADA2 provides ASVs, which are created by grouping unique sequences. ASVs are equivalent to OTUs at a 100% cutoff level and are discussed to replace customary OTUs (97% cutoff) in the future. Therefore, QIIME2 always means QIIME2-DADA2 within the context of this publication since the pipeline also offers other options for sequence clustering (e.g. Deblur) and results may considerably differ. For mock-16, 19 and 20 bp were trimmed off from the forward and reverse end, respectively. No trimming was applied for mock-13 since all sequence appendices were already removed in the raw input files. Thereafter, taxonomic information was assigned to the representative sequences using a pre-trained Naive Bayes classifier (Silva 132, 99% OTUs from the 515F/806R inner primer region, available at https://docs.qiime2.org/2019.10/data-resources/) and the “qiime feature-classifier classify-sklearn” tool. After exporting the table- and taxonomy QIIME2 artifact with “qiime tools export” and the external “biom-convert” tool, a tab-delimited ASV-table file was created.

For mock-26, selected input files (Mock.81, Mock.92, Mock.99) were imported with “qiime tools import” (data type: “SampleData[SequencesWithQuality]”, input format: “CasavaOneEightSingle-LanePerSampleDirFmt”). Sequence preparation and ASV-table construction was again done with the DADA2 pipeline; however, in this case with the single-end version of the plugin (“qiime dada2 denoise-single”). No trimming or sequence truncation was applied. Representative sequences were then taxonomically assigned with a pre-trained UNITE classifier and the “qiime feature-classifier classify-sklearn” plugin. The classifier was trained on the newest UNITE database release (version 8, dynamic construction), and the reads were not trimmed to the ITS primer sites as suggested by the QIIME2 documentation. The final tab-delimited ASV-table file was created as previously described for mock-13 and mock-16.

Soil sampling.

Samples were taken in Trins, Tyrol (Austria; 47°04′59” N, 11°25′00” E) on November 28, 2016. The annual average temperature is 4°C and the annual precipitation 770 mm [67]. Samples were taken at the end of the vegetation period, two months after the last fertilization with manure (swamp samples). Details on the sampling sites are given in Table 2.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Sites that were selected for soil analysis in order to demonstrate the functionality of CoMA.

https://doi.org/10.1371/journal.pone.0243241.t002

From each site, four field replicates were taken within an area of approximately 100 m². Each field replicate was composed of nine cores (0–10 cm depth, Ø 2 cm) taken in a regular systematic grid strategy (within a square of 1.4 x 1.4 m). Immediately after the sampling, soil samples were sieved (< 2 mm) and stored at -20°C until use. Soil physico-chemical characterization (S1 Table) was done as described by Fernández-Delgado Juárez et al. [69].

DNA extraction and NGS.

Total DNA was extracted using the NucleoSpin Soil kit (Macherey-Nagel, Düren, Germany) according to the manufacturer's protocol with minor adaptations: 0.25 g (fresh weight) of twelve soil samples (stored at -20°C) were lysed in Lysis Buffer 1 (SL1) at room temperature using a horizontal shaker (MM 2000, Retsch, Haan, Germany) at an amplitude of 80% for 5 min. Elution was done in 2x40 μL of Buffer SE. DNA extracts were checked on a 1.5% agarose gel and DNA concentration was determined via QuantiFluor dsDNA dye measurement using the Quantus fluorometer (Promega, Mannheim, Germany). Extracts were stored in low-DNA binding tubes (Axygen, Corning, USA) at -20°C until sequencing. Samples were then subjected to NGS amplicon-sequencing (Microsynth AG, Balgach, Switzerland) on an Illumina MiSeq device using a 250 bp paired-end (v2) approach targeting the V4 region of the 16S SSU rRNA gene. The applied primer pair was 515f (5’-GTGCCAGCMGCCGCGGTAA-3’) and 806r (5’-GGACTACHVGGGTWTCTAAT-3’), as described by Caporaso et al. [70] and recommended for soil in the Earth Microbiome Project [71].

Soil data analysis with CoMA.

Data analysis with CoMA (quality selection, OTU clustering, taxonomic assignment) was done as previously described for mock-13 but with a PQS cutoff ≥ 30. Rare OTUs with a sum of reads < 50 within all samples were excluded (corresponding ~0.005% of all reads as previously described in other studies [72, 73]) and data was subsampled to a depth of 92,539 after checking rarefaction curves (S1 Fig). This corresponds to the lowest read count among all samples and all analysis pipelines (CoMA, Mothur, QIIME). At this depth, the rarefaction curves of all samples reached a steady state and thus the sequencing effort was assumed to suffice. Diversity analysis was done using the four provided diversity indices: number of OTUs, Shannon-Wiener, Chao1 and Simpson. For cluster analysis, UPGMA (CoMA input: “average”) was chosen as method and Braycurtis distance as metric. This combination is suitable for grouping microbial communities and therefore often used in microbial ecology studies [74].

Soil data analysis with Mothur.

Soil samples were analyzed with Mothur following the same protocol and settings as described for mock-13. Removal of rare OTUs (sum of reads < 50) as well as data subsampling (92,539 reads) was done with CoMA as described above.

Soil data analysis with QIIME.

Soil data analysis was done following the same protocol and settings as described previously for mock-13 and mock-16. OTUs removal and subsampling was again done with CoMA using the same settings.

Benchmarking with mock communities.

The benchmark test showed that all tested pipelines (CoMA, Mothur, QIIME, QIIME2) were able to depict the taxonomy in each of the three mock communities in a proper way. Irrespective of the dataset, taxa were found with a high accuracy and a relatively low error rate. However, when looking closer, differences in performance were found among the tools depending on the dataset. While the CoMA output was closely related to that of QIIME2 in mock-13 and mock-26, a considerable distance between these two tools was seen for mock-16, where CoMA was almost identical with QIIME. This indicates that the accuracy of an analysis highly depends on the dataset and that no general conclusion for the best performing analysis tool can be drawn when not considering sample specificities. Decisive characteristics of input data may include the applied primer system (e.g. 16S rRNA, 18S rRNA, ITS, functional genes), the variable region (e.g. V3 or V4 for 16S rRNA), the sequence length, as well as the taxonomic structure of the sample (e.g. species richness, evenness). Even though a general conclusion is difficult, CoMA seemed to be the overall most consistent platform in this test set since it did not show any severe inaccuracies for any of the three datasets, while the other tools did so for at least one mock community.

When comparing the OTU-based platforms (CoMA, Mothur, QIIME) with QIIME2 computing ASVs, no clear trend was observed. For mock-26, the fungal ITS dataset, QIIME2 performed particularly well and was the most accurate tool for this dataset in terms of mean per-taxon deviation. Overall, however, QIIME2 was ranked last, what may be explained with the averagely lowest genus detection rate (< 83%) among all platforms within the test set, especially seen for mock-16. Here, QIIME2 was not able to reveal almost 25% (12 out of 46) of all bacterial/archaeal genera. These results evoke the question whether ASVs are truly preferable over OTUs, albeit it needs to be emphasized that this study was not designed to compare OTUs and ASVs in all its aspects and hence no general conclusion can be drawn. Callahan et al. [78] suggested a general replacement of OTUs in favor of ASVs for all marker-gene analyses and argued with a higher reproducibility, a finer sequence resolution and the advantage of comparability between different studies. On the other hand, existing sequencing technologies often lack precision and hence an accurate resolving of exact sequences appears doubtful. Choosing the clustering strategy is a tradeoff between lumping taxa on the one hand (OTUs) and splitting them on the other hand (ASVs), and the decision seems to depend on the analyzed dataset, particularly in terms of the applied primer system and the length of the inner primer region. This is also supported by Edgar [79], who tested various OTU cutoff levels and found optima of 99% for full length 16S rRNA, and 100% for short hypervariable regions (e.g. V4 of the 16S rRNA). For the next years, we expect a co-existence of OTUs (probably with a 99% cutoff instead of 97%) and ASVs, always depending on the data to analyze. However, as soon as the sequencing technologies are reaching the next level in terms of precision, ASVs may be favorable and the field should clearly move in this direction.

Irrespective of the dataset, the results of all pipelines were more closely related to each other than to the expected values (with one exception: QIIME at mock-26). This indicates that the inaccuracies in this benchmark test were, to a certain degree, caused by the sequencing process rather than by the applied analysis tool. Several other authors who reported a systematic distortion of NGS data [80, 81] drew the same conclusion. Biases occur all over the process, starting with the extraction of DNA, the PCR, and the sequencing procedure itself [82]. According to the authors, the steps of DNA extraction and PCR are particularly susceptible to errors, which further accumulate until the end of the sequencing process. This results in a severe bias towards some strains, while others are neglected [83]. These aspects always need to be considered when using amplicon sequencing and we strongly suggest to include samples with a known composition (= mock community) to the analysis in order to estimate the degree of introduced inaccuracy.

In all test categories, CoMA satisfied with a good performance in the benchmark test based on mock communities. It seems to provide proper results, irrespective of the characteristics of the analyzed dataset. The CoMA pipeline is therefore suggested as highly accurate alternative to the established NGS analysis platforms for any kind of input data. Future CoMA releases will also include the possibility to select the OTU cutoff level (currently: 97%) and to generate ASVs rather than OTUs (using the UNOISE3 algorithm [84]), making the pipeline more flexible and congruent with different opinions in the field no matter in which direction sequence analysis is heading in the future.

Platform comparison using real soil data.

Comparing the three soils, all of the three pipelines used for the soil dataset (CoMA, Mothur, QIIME) generally showed the same trends concerning microbial diversity. The overall level of diversity, however, was different among these analysis tools, which can be explained by the way diversity and richness are typically calculated. Most algorithms, also those in CoMA, are based on OTUs. Therefore, OTU clustering and subsequent data post-processing steps are crucial and may significantly affect diversity calculations, as well as all other upstream analyses. Generally, there are three different strategies for clustering sequences into OTUs: closed-reference OTU picking, de novo assembling and open-reference OTU picking [72]. Closed-reference OTU picking first assigns sequences to a taxonomic database and thereafter clusters them based on the taxonomic classification. This process is fast, but some problems arise when taxonomic assignment is ambiguous or when no similar entry can be found in the reference database [85]. De novo assembling reverses this approach and clusters sequences in a first step, followed by the taxonomic assignment. This strategy is more robust, especially when analyzing complex datasets, but also tends to be more intense in terms of computational effort [86]. Open-reference clustering manifests itself a hybrid of the above approaches. Similar to closed-reference OTU picking, sequences are first assigned to a reference database and unclassified reads are thereafter assembled de novo. This third option combines advantages of both approaches; however, it leads to two different OTU definitions within one single strategy [87].

CoMA as well as Mothur and QIIME all used de novo assembling for clustering of the sequencing data on a 97% similarity level. However, each tool implements individual algorithms or rather complete tool packages. CoMA and QIIME use USEARCH (in fact UPARSE [54]) for OTU clustering by using different versions of this algorithm (CoMA: USEARCH 10.0.240, QIIME: USEARCH 6.1.544), though. Mothur uses OptiClust [86] instead. Assuming that both USEARCH versions used in CoMA and QIIME are largely comparable, particularly OptiClust may have caused such differences with regard to Mothur. This is supported by our data where CoMA and QIIME were generally more closely related to each other than to Mothur. Beyond the fundamental OTU clustering, each pipeline includes various steps for quality filtering, including data demultiplexing, data denoising and chimera removal. All these steps vary among the tools and may significantly affect the results of data analyses. Subsequent removal of rare OTUs as well as data subsampling would evoke different outputs as well. However, this can be disregarded for the present comparison since all OTU-tables were treated with the same algorithms.

Data analysis with Mothur resulted in different microbial compositions in the three soil habitats even though most of the key microbes were found with all of the three pipelines. Despite the different algorithms for OTU clustering and data quality filtering, the output of these analyses might be strongly influenced by the aligning tool/mechanism and the used reference database for taxonomic assignment. In all cases, alignment was done via different tools: BLAST in CoMA, RDP in Mothur and UCLUST in QIIME, following the recommendations in the respective SOPs. However, it should be borne in mind that all of the three pipelines also offer alternative aligners and as such, we are not addressing this aspect in detail. Taking a closer look at the results revealed that most of the conflicting taxa showed particularly low abundances. For the composition of low-abundant taxa, removal of rare reads has a particularly high impact. The same taxon may be slightly above the cutoff in one analysis, whereas it is slightly below in the other. Consequently, a taxon that was in fact determined with both tools may appear just in one of them, resulting in a decreased coverage. We assume that the applied reference database is another important driving factor for the varying results between CoMA/QIIME and Mothur. Both, CoMA and QIIME used the newest distribution of SILVA (SLV_132). Taxonomic classification with Mothur, however, was done using the current release of the RDP trainset (Version 16) as described in the SOP. This explains the differences found in terms of unclassified taxa since SILVA includes significantly more sequences than RDP [88], resulting in a better coverage. Nevertheless, this does not explain the mismatches when comparing microbial core families. These taxa were not only different at family level but also down to phylum or class.

Taken together, all comparisons indicate that the CoMA output is generally well comparable to that of Mothur and QIIME, particularly when looking at the microbial key players. Regarding less abundant taxa, differences to Mothur were found.