Evaluating 16S metagenomics data analysis pipelines using an Ion Torrent sequencer for human gut microbiome research

Abstract

Background

High-throughput 16S rRNA gene amplicon sequencing is essential for characterizing gut dysbiosis in translational research. However, microbiome profiles vary widely across sequencing platforms, library preparation protocols, targeted hypervariable regions (HVRs), and bioinformatics workflows. For users of the Ion Torrent S5 platform, there is currently no consensus methodology for performing open-source analyses as an alternative to proprietary frameworks. This study evaluates the proprietary Ion Reporter™ software against eight open-source pipelines based on QIIME2 and the SILVA database, accounting for data compositionality. Benchmarking was performed using a commercial human fecal reference material (ZymoBIOMICS™ D6323) characterized via DNA-shotgun sequencing (the ZB-Shotgun profile), as the reference standard.

Results

Poisson regression modeling showed that using two primer sets for library preparation yielded a 30 % lower expected read count than a single primer set (IRR = 0.6986, p < 0.001). Across all configurations, the choice of denoising algorithm (DADA2 vs. Deblur) significantly affected alpha diversity calculations. At the phylum level, hierarchical clustering of Bayes factors (BFs) indicated that multi-region Ion Reporter workflows produced stable consensus predictions. Conversely, at the species level, single-HVR open-source pipelines significantly outperformed multi-region approaches.

Conclusions

Open-source pipelines that combine targeted HVR2 or HVR4 inputs with state-of-the-art denoisers outperform proprietary Ion Reporter options, providing superior taxonomic resolution and precision for species-level microbiome profiling. This study establishes a traceable, reproducible, and cost-effective framework that empowers independent researchers to maximize the utility of Ion Torrent 16S rRNA data in healthcare and gastrointestinal research without relying on restrictive software ecosystems.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1186/s12866-026-05444-3.

Introduction

Genome sequencing has provided microbiologists, ecologists, and biomedical scientists with an unprecedented view of microbial diversity across a wide range of ecosystems. In particular, the commensal flora of the human gut has been extensively studied and linked to the onset of numerous human conditions [1, 2], including metabolic disorders, diabetes, inflammatory bowel diseases, and cancer [3, 4]. Recent advances in the cost-effectiveness and accessibility of commercial next-generation sequencing (NGS) platforms have facilitated their global adoption in healthcare [5]. However, translating NGS into standard clinical workflows remains a fundamental challenge for implementing these technologies in clinical settings [6, 7].

The excitement of applying these emerging technologies to novel research questions has occasionally overshadowed crucial technical considerations that affect comparisons across investigations conducted by different laboratories using different protocols [8, 9]. Indeed, the use of mock communities in microbiome studies has revealed that variations in experimental conditions and analytical methods dramatically affect the quality of the results [8, 10]. Given the reliance on NGS technologies to characterize bacterial communities, comparative studies across diverse platforms and methods are essential to ensure research reproducibility.

Most microbiome studies rely on either shotgun metagenomic sequencing (DNA-shotgun) or 16S ribosomal RNA (rRNA) amplicon sequencing (16S-amplicon). While DNA-shotgun sequencing involves sequencing the total DNA within a sample to enable taxonomic, functional, and metabolic profiling, it is generally more expensive, time-consuming, and requires more input DNA than 16S-amplicon sequencing. The latter employs PCR amplification before sequencing a short, specific fragment of the bacterial 16S rRNA gene. This gene is a well-established taxonomic biomarker of approximately 1542 base pairs (bp) with a highly conserved backbone across prokaryotes, interspersed with nine hypervariable regions (HVR1 to HVR9) that show lower sequence conservation. Sequences from these HVRs enable taxon identification in bacterial diversity studies [11].

Currently, there is no consensus on which HVR to target, and practices vary widely across fields. For instance, HVR1 and HVR2 have shown the highest resolving power for accurately identifying respiratory bacterial taxa from sputum samples [12], whereas HVR3 detects the highest bacterial diversity in the vaginal microbiota [13, 14]. HRV4 is the most commonly used region to study the gut microbiome, followed by HVR2 and HVR3 [15, 16]. Furthermore, diverse studies have confirmed that prediction of microbial diversity in a sample differ across sequencing platforms, library preparation methods, protocols, and bioinformatics pipelines used for data analysis. Optimization to prevent bias and obtain a comprehensive view of complex communities would ideally involve sequencing the full-length 16S rRNA; however, this remains challenging with short-read next-generation sequencing technology [17], which typically limits read lengths to 200–600 base pairs.

Although the Ion Torrent sequencer is not the most prevalent platform, its use in human gut microbiome studies is supported by a notable number of recent publications [18–20] and it remains a relevant tool in translational medical research [21]. Furthermore, while the current literature demonstrates a persistent, though still incomplete, effort to standardize reproducible analytical workflows for Ion Torrent 16S data [17, 22, 23], published guidance on bioinformatics choices for gut microbiome researchers using this platform remains limited. This study aims to address this gap.

When operating the Ion Torrent S5 platform (Thermo Scientific) alongside the Ion 16S™ Metagenomics Kit, users can generate two microbiome profile predictions using the manufacturer’s proprietary bioinformatics solution. The 16S metagenomics workflow (v1.1), installed as a module in Ion Reporter™ software (v5.12; from now on, Ion Reporter software) delivers a consensus profile by integrating data from multiple amplicons sequenced simultaneously. For instance, three amplicons (targeting HVR2, HVR4, and HVR8) are amplified and sequenced when samples are prepared with a single set of primers (1S). Conversely, six amplicons (targeting HVR2, HVR4, and HVR8, alongside HVR3, HVR6-7, and HVR9) are processed when two sets of primers (2S) are used. These consensus predictions are ultimately exported as feature tables at various taxonomic levels.

However, the Ion Reporter software available to Ion 16S™ Metagenomics Kit users has several technical limitations for advanced downstream analysis. It does not allow users to incorporate study-specific metadata into analyses, and the exported processed data lacks information from prior analyses, preventing downstream analysis with open source tools [24]. Crucially, Ion Reporter relies on an outdated GreenGenes database (v13.5; [25] and an obsolete QIIME version (QIIME1; [26]. Furthermore, it fails to account for the compositional nature of microbiome data during beta diversity and other taxonomic calculations. This is relevant in a research environment that implies downstream bioinformatics analysis [9], such as differential analysis of microbiome composition between study groups. The compositional nature of the data requires the use of specific statistical methods of analysis [27]. In fact, ignoring the compositional nature of microbiome data when comparing between samples can lead to spurious results [28–30].

Consequently, Ion Torrent S5 users must choose between using the manufacturer's default software and migrating data to external, updated bioinformatics frameworks, despite the lack of a clear consensus approach for Ion Torrent 16S data. Moreover, because primer sequences are not provided to Ion 16S Metagenomics Kit users, data from different HVRs are stored in separate FASTQ files with primer sequences removed Researchers must therefore choose between a multi-region consensus result generated by an outdated proprietary workflow and exploring modern, open-source, open-architecture alternatives restricted to single-HVR analyses.

Although a previous mock-community study suggested that amplicons targeting HVR4 and HVR2 yield the most consistent results on this platform [17], comprehensive validation of single-region open-source alternatives against shotgun metagenomic standards remains unavailable.

To bridge this gap, this work aims to guide gut microbiome researchers in selecting the pipeline most suited to their goals, when the starting data originate from Ion Torrent, proposing open-source and freely available alternatives and updated bioinformatics approaches that consider a single HVR.

Recent comparative methodological studies have suggested the use of Quantitative Insights Into Microbial software version 2 (QIIME2) [31] over other available bioinformatics tools such as QIIME1, Mothur and MAPseq [26, 32, 33] for 16S-amplicon sequencing data. QIIME2 is a comprehensive toolset that can process, classify, and perform downstream analyses for microbiome data. For taxonomic classification, this tool compares a set of query sequences against a defined reference database, such as SILVA [34], and assigns the most likely taxonomic lineages. Ultimately, the success of these analyses depends not only on database curation but also on the accuracy of the classification and denoising algorithms used. By default, QIIME2 uses a naive Bayes classifier [35], to assign biological sequences to a reference database [8] while state-of-the-art denoisers like DADA2 [36] or Deblur [37] are used to infer true biological sequences from raw reads. Comparative studies have demonstrated that SILVA provides significantly more specific and accurate classifications at the genus level compared to other database options, particularly within complex gut-associated microbiome [38]. SILVA database has long stood out as the most consistently updated and comprehensive database for ribosomal RNA gene sequences [39].

To date, no studies have examined alternative open-source analysis pipelines for microbiome data from the Ion Torrent S5 sequencing platform (ThermoScientific) that account for data compositionality. This study evaluates Ion Reporter consensus results obtained with one (1S) or two primer sets (2S) and six available 16S pipeline options based on QIIME2 and the SILVA database. We assess the performance of HVR2 and HVR4 independently, and in combination with widely used denoising algorithms such as DADA2 and Deblur. To ensure a rigorous evaluation, all comparisons were performed using a commercial human fecal mock sample (ZymoBIOMICS™ D6323) characterized via a proprietary DNA-shotgun metagenomic pipeline (the ZB-Shotgun) as the reference standard. Our findings highlight that primer set selection, HVR choice, and denoising algorithms significantly affect taxonomic resolution, alpha diversity calculations, and overall composition, providing guidelines to ensure data quality and cross-study reproducibility in gastrointestinal research.

Methods

Mock microbial community standard

The mock sample used in this study is a reference microbiome material, ZymoBIOMICS™ Fecal Reference with TruMatrix™ technology (D6323, Zymo Research), with a DNA concentration of 6 ng µl–1, and containing the following bacterial phyla: Firmicutes (71.33%), Bacteroidetes (23.27%), Actinobacteria (1.81%) and Verrucomicrobia (1.08%).

Technical description of the MOCK sample

The mock sample used as the reference in this study is a commercial microbiome quality control made from real human fecal material (ZymoBIOMICS^TM Fecal Reference w/TruMatrix^TM Technology, Research, cat: D6323). This reference mock is derived from healthy adult donors and has been homogenized into a single large batch/suspension using the TruMatrix™ Technology. The product was processed using the following Zymo method: 100 μl of the product was extracted using the ZymoBIOMICS™ DNA Miniprep (D4300). The shotgun library was prepared using the Illumina® DNA Prep kit. DNA sequencing was performed on the Illumina® NovaSeq®. Finally, bioinformatics analysis was performed using the private ZymoBIOMICS™ Metagenomics Pipeline, which uses Centrifuge (v1.0.4) [40] for taxonomic identification (ZB instruction manual Ver 1.0.3. 2021). The bacterial profile was determined with the ZymoBIOMICS™ Metagenomics Pipeline, and the DNA shotgun sequencing raw data from the mock fecal reference (ZB) were available from the provider.

DNA extraction, library preparation and sequencing

DNA was extracted from the Mock sample (ZB) using the MagMAX^TM Microbiome Ultra Nucleic Acid Isolation Kit (Applied Biosystems^TM, A42358), according to the manufacturer's instructions. The quality of the extract was verified on a 0.8% agarose gel stained with GelRed (Biotium). Its concentration was measured on the Qubit 3.0 Fluorometer using the Qubit^TM 1X dsDNA HS Assay kit (Invitrogen, Q33230), and its purity was assessed using a spectrophotometer (Cytation 5, BioTek). Three nanograms of DNA were used to amplify hypervariable regions of the 16S rRNA gene using one set of primers (1S) or two sets of primers (2S) from the Ion 16S^TM Metagenomics Kit (Thermo Fisher Scientific, A26216), according to the manufacturer's instructions (Life Technologies 2015). Then, 100 ng of this PCR product was used to construct a DNA library with the Ion Plus^TM Fragment Library Kit (Thermo Fisher Scientific, 4471252) and to index samples using the Ion Xpress^TM Barcode Adapters 1–16 kit (Thermo Fisher Scientific, 4471250), according to the manufacturer's instructions. The concentration of amplicon libraries was quantified using the Ion Universal Library Quantitation Kit (Thermo Fisher Scientific, A26217) on the QuantStudio 3 thermocycler (Applied Biosystems, Waltham, MA, USA). Quantified libraries were diluted to 20 pM and pooled in equimolar ratios for normalization. From the pooled libraries, 25 μl was transferred to library sample tubes in the Chef^TM reagent cartridge. Emulsion PCR, bead enrichment, and chip loading were automatically performed on an Ion Chef^TM instrument (Thermo Fisher Scientific) using the Ion 510^TM& Ion 520^TM& Ion 530^TM kit (Thermo Fisher Scientific, A34019). The Ion 510^TM chip (Thermo Fisher Scientific, A34288) was used. Base calling and run demultiplexing were performed with the Torrent Suite^TM Software v.5.12.2 (Thermo Fisher Scientific). DNA from a sample containing 100% E.coli was used as internal positive control for the sequencing process. Finally, sequencing was carried out on the Ion GeneStudio S5 Plus Sequencer (Thermo Fisher Scientific) in our laboratory, using the Ion 510^TM& Ion 520^TM& Ion 530^TM kit with read length settings at 400 and number of flows at 850.

Sequencing of mock samples

Six sample replicates from the ZB mock fecal reference and a positive control sample containing 100% E.coli were sequenced in 4 independent sequencing runs and named as “Mock1SNGS4”, “Mock1SNGS5”, “Mock1SP3” (prepared using one primer set in three independent runs), “Mock2SNGS4”, “Mock2SNGS5”, and “Mock2ST1” (prepared with primer sets in three independent runs). Data for HVR2 and HVR4 from samples sequenced with one primer set (n=3), together with Ion Torrent consensus results obtained with one or two primer sets were selected for comparison in 16S analysis (Additional file 1: Table S1).

Microbiome analysis pipelines

Because primer sequences are not made available to Ion 16S^TM Metagenomics Kit users, the raw data must be analyzed using pipeline options outside Ion Reporter software. The ThermoFisher bioinformatics team separated the FASTQ files by primer set, resulting in 3 or 6 separate FASTQ files per sample. Each FASTQ file contains data from each sequenced amplicon (HVR2, HVR3, HVR4, HVR6-7, HVR8, and HVR9), with primer sequences removed.

Ion Reporter software consensus results obtained with one (1S) or two primer sets (2S) were evaluated, together with four bioinformatics options based on QIIME2, using a custom in-house pipeline developed at the Laboratory of Translational Medicine (LMT). The pipeline used the SILVA database (v138) as the taxonomic reference and “qiime2_feature_classifier” (sklearn) as the taxonomic classification method [35]. These four pipeline options combined HVR2 and HVR4 with denoising algorithms such as DADA2 and Deblur.

The in-house pipelines consist of modules based on QIIME2 core functionalities: FASTQ input files are imported into QIIME2 via “qiime tools import” in SingleEnd FastqManifestPhred33V2 format [31]. QIIME2 v 2022.11 was used to perform quality filtering (PHRED scores < 19 were discarded) [41, 42], denoising (Deblur or DADA2), taxonomic classification using a pre-trained Naive-Bayes classifier fitted to the SILVA database (v138), and calculation of alpha and beta diversity metrics. Alpha diversity analysis generates rarefaction curves for “Shannon” and “chao1” metrics. Values representing the internal positive control sample (100% E. coli) were excluded from alpha diversity plots to improve box plot visualizations. Beta diversity analysis uses the Deicode [43]. QIIME2 library to calculate the Aitchison distance between samples using a 2-dimensional robust PCA approach. Pipeline results can be exported to BIOM and TSV file formats for further analysis. The “Dokdo” Python library 1.4 is used to generate result visualization (https://github.com/sbslee/dokdo, [44].

As a crossed-reference, raw data from HVR2 and HVR4 in samples “Mock1SNGS4”, “Mock1SNGS5”, and “Mock1SP3” were independently analyzed (n=3) using “Nephele”, a freely available platform for microbial bioinformatics analysis [45]. Its pipeline is based on QIIME2 and the SILVA database (v138). The analysis was run using default parameters and Deblur as the denoiser. One microbiome profile per HVR was obtained.

Microbiome profile comparisons

Microbiome profiles from the Ion Reporter software (IR), the Nephele web platform (Nephele), and in-house pipelines (LMT) were compared with the ZB-Shotgun profile, used as the reference standard. Feature tables from each platform were merged with the ZB-Shotgun reference profile into consolidated data frames at the phylum, genus, and species taxonomic levels. Discrepancies in taxonomic classifications arising from reference database use were manually curated for further analysis.

Alpha diversity metrics were calculated using QIIME2, and boxplot visualizations were generated using Python tools (pandas, seaborn). Beta diversity was evaluated by calculating Aitchison distances between the profiles of each 16S pipeline option and the DNA-shotgun reference profile (ZB) at the phylum, genus, and species levels. Visualizations were obtained using a 2-dimensional robust compositional PCA approach via DEICODE [43] and Dokdo [44]. The code generated and utilized during this study is publicly available in the GitHub repository: https://github.com/n-bravo/16S-data-analysis

Statistical analysis

To evaluate statistical differences in alpha diversity among 16S bioinformatics pipelines at the genus and species levels, Kruskal–Wallis tests with post-hoc pairwise comparisons were performed. The False Discovery Rate (FDR) was controlled using the two-stage linear step-up procedure of Benjamini, Krieger, and Yekutieli, as implemented in GraphPad Prism (version 10).

To quantify the relative evidence for similarity or divergence between the microbiome profiles predicted by each method and the reference standard, Bayes factors (BFs) were computed for each taxonomic classification across all levels (phylum, genus, and species). The calculations were based on observed differences in Aitchison distances between the DNA-shotgun profile and the predictions from each 16S pipeline. Becauses we expect the methods to match the standard we assume that the mean difference per taxon is zero under the null hypothesis (H₀: = 0), whereas the alternative hypothesis posits that this mean differs from zero (Hₐ: ≠ 0). Given this constraint and following established literature [46, 47], we used the Savage–Dickey density ratio to estimate Bayes factor per taxon. This approach allows a direct comparison of the posterior and prior distributions at the null value, providing an efficient means of evaluating evidence in favor of the null or alternative hypothesis (for details, see [48].

The choice of a Bayes factor framework is motivated by the nature of the question addressed here. Because the null hypothesis is a specific point hypothesis representing concordance with the reference (H₀: μt = 0), the goal is to quantify evidence in favor of similarity, not only to reject a difference. Standard null hypothesis significance testing is designed to reject the null and cannot provide evidence in its favor. Approximating this within a frequentist framework would require combining equivalence testing, which relies on predefined equivalence margins, with conventional difference testing, applied across the hundreds of taxonomic classifications evaluated here, with the associated inflation of the type I error rate. The Bayes factor resolves both directions of evidence within a single interpretable quantity per taxon [49].

This framework was implemented using a Bayesian hierarchical model in JAGS. In this model, the parameters of interest, per-taxon mean difference (μt) and standard deviation (σt), were estimated individually for each taxonomic unit t from the observed Aitchison distance differences. The hyperparameters govern the prior distributions placed on these per-taxon parameters and are shared across all taxa, allowing information sharing across taxonomic units and reducing the variance of individual estimates. Specifically, the prior distribution for μt was specified as a normal distribution centered at zero with a precision of 0.2 (i.e., μt ~ Normal (0, 0.2)), reflecting the expectation that differences between methods should be small. The prior distribution for σt was specified as a uniform distribution between 0 and 30 (i.e., σt ~ Uniform (0, 30)).

The upper bound of 30 for σt was determined through a preliminary simulation study designed to ensure correct behavior in the presence of structural zeros. In compositional microbiome data, structural zeros arise when a taxon is genuinely absent from a sample, producing zero abundance in both the 16S pipeline profile and the shotgun reference profile. After Aitchison distance calculations (which respect the compositional nature of the data throughout) these structural zeros result in an observed mean difference of zero for the corresponding taxon. The upper bound of σt was calibrated so that, under this scenario of true concordance (mean difference = 0), the model reliably favored H₀, yielding BF₀₁ > 3. This calibration ensures that structural zeros do not artifactually penalize pipeline performance.

The Savage-Dickey Ratio method was applied iteratively across taxonomic units. Posterior densities were obtained from samples generated by the hierarchical model in JAGS using custom R scripts (R version 4.3.1 and rjags version 4.14). JAGS was used to sample from the posterior distributions of the parameters given the data, enabling estimation of the posterior densities required to calculate Bayes Factors.

A threshold of BF₀₁ > 3 was established as substantial evidence supporting the null hypothesis, indicating that the evaluated method was highly similar to the reference. Conversely, BF₀₁ < 0.33 was considered evidence in favor of the alternative hypothesis, indicating significant divergence between the DNA-shotgun method and the 16S pipeline. Values between 0.33 and 3.0 fall within the “gray area” of the Bayesian method used reflecting classification uncertainty. These thresholds follow the conventional Bayes factor interpretation scale proposed by Jeffreys (1961) and widely adopted in the literature [47]. Clustered heatmaps of BFs were generated to provide a global view of the taxonomic performance, with darker colors representing higher BF values.

Alternative shotgun sequencing bioinformatics pipeline

The microbiome composition of the mock sample used in this study was previously characterized using DNA shotgun sequencing and the bioinformatics solutions of Zymo Research (see materials and methods). To investigate potential variations in Aitchison distances at the species level attributable to the choice of the shotgun data-analysis bioinformatics pipeline, the raw DNA shotgun sequencing data provided by Zymo Research were re-analyzed with an alternative bioinformatics pipeline implemented in-house at our installations and the resulting profile (SM-shotgun) was compared to the ZB-Shotgun reference profile, to quantify the sensitivity of the reference to shotgun pipeline choice. This alternative analysis began with a thorough examination of raw sequence quality using fastqc v0.11.5 [50]. Subsequently, fastp v0.23.1 [51] was used for adapter removal and trimming. Retaining shotgun reads with a quality score exceeding Q35, we performed a de novo assembly using SPAdes v3.15.5, integrated into the SqueezeMeta v1.6.2 pipeline [52]. The assembly was executed with the following parameters: "-m sequential -t 30 -a spades -assembly_options ---meta --only-assembler y -t 30 -m 60 -binners maxbin, metabat2 --doublepass." To validate the taxa predicted by both the ZymoBIOMICS and SqueezeMeta workflows, we curated a comprehensive database encompassing 533 genomes predicted by ZymoBIOMICS, alongside the SqueezeMeta assembly outputs. This database served as a reference for assessing the breadth and depth of coverage for each genome. The assessment utilized the trimmed reads from ZymoBIOMICS, mapping and filtering them through CoverM v0.6.1, default parameters [53]. The resulting species-level profile predictions were further evaluated against the primary benchmark.

Results

This study evaluates the proprietary Ion Reporter™ software against eight open-source pipelines based on QIIME2 and the SILVA database, accounting for data compositionality. Benchmarking was performed using a commercial human fecal reference material (ZymoBIOMICS™ D6323) characterized by DNA-shotgun sequencing (the ZB-Shotgun profile) as the reference standard. (Figures 1 and 2).

Fig. 1.

Schematic representation of the study design and main observations. This study evaluates eight 16S bioinformatics pipeline options, including Ion Reporter software on the Ion Torrent platform and open-source options for microbiome analysis based on QIIME2 and the SILVA database, with denoising algorithms such as DADA2 and Deblur. Benchmarking was performed using a commercial human fecal reference material (ZymoBIOMICS™ D6323) characterized via DNA shotgun sequencing (the ZB-Shotgun profile) as the reference standard

Fig. 2.

Main differences between the 16S bioinformatics pipeline options evaluated in this study. Ion Reporter consensus results obtained from multiple hyper-variable regions and based on QIIME1 (colored in blue) were evaluated, together with six bioinformatics options based on QIIME2 and SILVA database (colored in purple and green). We assess the performance of HVR2 and HVR4 independently, using widely used denoising algorithms such as DADA2 and Deblur. As a cross-reference, independent analyses of HVR2 and HVR4 were run through “Nephele”, a freely available platform for microbial bioinformatics (colored in green)

Sequencing runs and data generated from mock samples

The number of sequenced reads per hypervariable region (HVR) across the mock samples was not homogeneously distributed (Additional File 2: Figure S1). On average, read counts for HVR2, HVR4, and HVR8 using 1 primer set (1S) were 41,946; 76,437 and 66,437 respectively, In contrast, the read counts for the same HVRs utilizing two primer sets (2S) were substantially lower, yielding 17,853; 29,226 and 24,973 respectively. HVR 9 showed the lowest average read count (1,769) when using 2primer sets, whereas HVR 4 showed the highest average throughput when samples were prepared with one primer set. To evaluate the effect of primer set configurations on read count data while adjusting for HVR, a Poisson regression model was executed, with results expressed as incidence rate ratios (IRRs). Notably, library preparation using two primer sets resulted in a 30% lower expected read count compared to the single primer set protocol (IRR = 0.6986, 95% CI: 0.6958, 0.7015, p < 0.001), suggesting a significant overall effect of the number of primer sets used for library preparation. This effect was estimated while controlling additively for HVR, ensuring that differences in primer sets were evaluated independently of variations across HVR conditions. Furthermore, IRRs indicated significant differences in event counts across targeted regions relative to HVR2; HVR 4 and HVR 8 showed higher expected counts (IRRs of 1.767 and 1.5286, respectively), whereas HVR 3, HVR 6–7, and HVR 9 displayed substantially lower IRRs (0.4642, 0.4608, and 0.0296, respectively), indicating a marked reduction relative to HVR 2 (Additional file 3: Table S2).

When comparing microbiome consensus profiles generated by the Ion Reporter™ software (v5.12), Aitchison distances among triplicates prepared with one primer set (1S) were consistently smaller than those among triplicates prepared with two primer sets (2S). The tighter clustering and reduced dispersion of the 1S triplicates suggest that a single primer set (1S) yields higher analytical consistency (Additional file 4: Figure S2).

Consequently, to maximize read depth per sample while maintaining data quality, sequences from HVR2 and HVR4 of the 1S protocol (n = 3) and the consolidated Ion Reporter consensus results (IR-1S and IR-2S), were selected for downstream 16S pipeline benchmarking (Additional file 1: Table S1).

Alpha diversity metrics comparisons between pipelines

To assess whether pipeline choice affects alpha diversity calculations, Shannon and Chao1 scores for each pipeline, at the genus and species levels were compared (Figure 3). Values for the internal positive control sample (Escherichia coli) were excluded from the visualizations. Statistically significant differences (adjusted q-value < 0.05) were observed (Additional file 5: Table S3) at the genus level for the Shannon index (Figure 3A) when comparing the IR software pipelines (IR-1S and IR-2S) with the in-house pipelines utilizing the HVR2 region (LMT-DADA-HVR2 and LMT-Nephele-HVR2, respectively). Similarly, significant differences were found for the Chao1 index (Figure 3B) between the in-house pipeline LMT-DADA-HVR4 and both IR software pipelines (IR-1S and IR-2S). Notably, at the genus level, the LMT-DADA-HVR2 in-house pipeline achieved a Shannon index (on average, 4.22) closest to the ZB-Shotgun reference (4.29). Similarly, the LMT-DADA-HVR4 pipeline provided the closest Chao1 index value (on average, 64) to the ZB-Shotgun reference (Chao1 index 90).

Fig. 3.

Comparison of alpha diversity metrics between different bioinformatics approach. Alpha diversity metrics were predicted for each evaluated pipeline option, including the value of the ZB-Shotgun reference profile. Shannon and Chao1 scores were calculated at the genus level (A and C respectively) and the species level (B and D respectively). Asterisks indicate significant differences. Kruskal–Wallis tests with post-hoc pairwise comparisons were performed using the two-stage linear step-up procedure of Benjamini, Krieger, and Yekutieli to control the false discovery rate (q-value < 0.05). ZB-Shotgun was excluded from the pairwise comparisons)

At the species level, alpha diversity scores for the IR software pipelines (IR-1S and IR-2S) were consistently higher than those for the in-house pipeline options (Figures 3C, 3D and Additional file 6: Table S4). Statistically significant differences (adjusted q-value < 0.05) were confirmed for both the Shannon (Figure 3C) and Chao1 (Figure 3D) indices when comparing the IR pipelines with the HVR4-based in-house pipelines (LMT-DADA-HVR4, LMT-Deblur-HVR4, and Nephele-HVR4).

Comparisons of alpha diversity across pipelines at the genus and species levels indicate that as taxonomic resolution narrows to the species level, the IR software pipelines (IR-1S and IR-2S) yield consistently higher alpha diversity scores than all in-house alternatives. Furthermore, DADA2 consistently inferred higher diversity than Deblur, which is consistent with previous literature [54].

Unlike at the genus level, differences at the species level are not HVR-specific. Statistically significant inflation (adjusted q-value < 0.05) is observed across both Shannon and Chao1 indices, particularly when comparing IR with all HVR4-based in-house pipelines (LMT-DADA-HVR4, LMT-Deblur-HVR4, and Nephele-HVR4).

Beta diversity metrics comparison and Aitchison distances between pipeline profiles

Aitchison distances were calculated between profiles from each 16S pipeline option and compared with the ZB-Shotgun reference profile at phylum (Additional file 7: Figure S3), genus (Figure 4), and species level (Figure 5). At the phylum level, predictions integrating data from multiple hypervariable regions (IR-1S and IR-2S) exhibited the shortest Aitchison distances to the ZB-Shotgun reference. At the genus level, 16S pipeline options cluster together but are not close to the ZB-Shotgun reference in the plot. Results considering HVR4, such as LMT-DADA-HVR4, and Nephele-HVR4 are closest to the ZB-Shotgun reference, followed by results considering HVR2. Interestingly, at the species level, results obtained using single HVRs cluster separately from those obtained using multiple HVRs such as IR-1S and IR-2S, with the latter showing the greatest dispersion between triplicates. Despite the inherent limitations of species-level resolution in 16S rRNA sequencing, our bioinformatic pipeline consistently and accurately resolves all positive controls to the exact same expected position, far removed from the complex experimental samples

Fig. 4.

Evaluation of microbiome profile similarities using beta diversity metrics at the genus level. Beta diversity was calculated at the genus level using a 2-dimensional robust PCA to represent the Aitchison distances between the shotgun profile (ZB, blue circle) and the profiles predicted for each 16S pipeline option evaluated (represented by triplets of circles of different colors on the plot). Internal sequencing positive controls (with 100% E. coli) for the analyzed triplicates completely overlap and are represented by the black star

Fig. 5.

Evaluation of microbiome profile similarities using beta diversity metrics at the species level. Beta diversity was calculated at the species level using a 2-dimensional robust PCA to represent the Aitchison distances between the shotgun profile (ZB, blue circle) and the profiles predicted for each 16S pipeline option evaluated (represented by triplets of circles of different colors on the plot). Internal sequencing positive controls (with 100% E. coli) for the LMT and Nephele pipelines overlap each other (distance 0) and are represented by a black star in the plot, while the IR-1S and IR-2S controls (green and red stars, respectively) mostly overlap each other

To assess the pipeline dependence of the reference standard itself, the raw DNA shotgun sequencing data provided by Zymo Research and used as a reference in this work were re-analyzed using the alternative SqueezeMeta-based pipeline (SM-shotgun) (Additional file 8: Figure S4). Substantial variations in Aitchison distances at the species level confirm that the ZB-Shotgun profile is not pipeline-invariant [55, 56]. While single-HVR results clustered together, major structural distances emerged between profiles derived from identical raw shotgun sequencing data analyzed through different bioinformatics frameworks.

Methodological evaluation using a Bayes Factor approach

To validate the previous results, Bayes factors (BFs) were calculated for each taxonomic classification at each taxonomic level (phylum, genus, and species). The calculations were based on the differences in Aitchison distances between the DNA-shotgun profile and the microbiome profile predictions from each 16S pipeline option.

Clustered heatmaps of BFs were generated to provide a global view of the taxonomic calculations, with darker colors representing higher BF values.

At the phylum level, the clustered heatmap of BFs shows a lack of consistency between the microbiome profiles predicted by the evaluated 16S pipeline methods and the ZB-Shotgun reference profile (Figure 6) which separates them into two primary clusters. The first cluster comprises the Ion Reporter pipelines (IR-1S and IR-2S), data from multiple HVRs to predict a consensus microbiome profile. The IR pipelines show BF values over 3.0 (3.4) for abundance predictions of the phylum Desulfobacterota, indicating strong evidence that these IR pipelines are similar to the reference profile, at least for that taxonomic classification at the phylum level. On the other hand, IR-1S and IR-2S show values of 0.4 and 0.3, respectively, for predicting Verrucomicrobia abundances; the latter BF value indicates strong differences between the DNA-shotgun profile and IR-2S predictions for that taxonomic classification. The second cluster consists of the single-HVR open-source pipelines. Within this cluster, pipelines using HVR4 as input group together (LMT-DADA-HVR4, LMT-Deblur-HVR4, and Nephl-HVR4), while the pipelines using HVR2 as input show a different pattern. Only pipelines using Deblur as a denoiser group together (LMT-Deblur-HVR2 and Nephl-HVR2), and the pipeline using DADA2 as a denoiser (LMT-DADA2-HVR2) is outside the group. Except for LMT-DADA2-HVR2, all single-HVR pipelines show strong evidence of similarity to the ZB-Shotgun reference regarding unassigned ("unknown") sequences at the phylum level.

Fig. 6.

Clustered heatmap of Bayes Factors at phylum level. Global view of the consistency between the 16S bioinformatics pipelines evaluated and the ZB-Shotgun reference profile at the phylum level. Darker colors represent higher BF values and indicate strong evidence of consistency among pipelines. A BF > 3 indicates strong evidence in favor of the evaluated method being similar to the reference (marked with ∆). Conversely, BF < 0.33 is considered evidence in favor of the alternative hypothesis, indicating strong differences between the DNA-shotgun method and the 16S pipeline methods evaluated (marked with ¥). BF values showing strong evidence for the null and alternative hypotheses are further investigated (see also Table 1)

16S pipelines targeting HVR4 achieved BF values over 3.0 for predictions of the phylum Verrucomicrobia (3.4 on average) and Firmicutes (3.1 on average), but at the same time, abundance predictions for bacteria belonging to the phylum Tenericutes are very different from the ZB-Shotgun reference values (BF values of 0.1 on average).

Taken together, predictions at the phylum level show that IR pipelines deliver comparable results and have the highest average BF values compared to the other 16S pipelines (Table 1). This consistency is mainly due to the multiple hypervariable regions that IR pipelines consider when calculating a single consensus microbiome profile. In other words, at the phylum level, pipelines using only one HVR to predict microbiome profiles, are strongly influenced by the HVR chosen as input and can also be influenced by the denoiser used within the pipeline.

Table 1.

Descriptive statistics for BF values calculated between the DNA shotgun method and the evaluated 16S pipeline methods

Taxonomic level	Measures	16S Pipeline options evaluated
		Pipeline options using HVR4			Pipeline options using HVR2			Multiple HVRs
		LMT-DADA HVR4	LMT-Deblur- HVR4	Nephele- HVR4	LMT-DADA HVR2	LMT-Deblur- HVR2	Nephele- HVR2	IR-1S	IR-2S
Phylum (11)	Median	0.91	0.95	0.94	0.98	0.95	0.92	1.35	1.85
	Average	1.47	1.47	1.47	1.10	1.16	1.15	1.80	1.85
	% Classification with BF > 3	3.33	3.33	3.33	0	1.11	1.11	2.22	2.22
	% Classification with BF < 0.33	1.11	1.11	1.11	2.22	2.22	1.11	0	1.11
Genus (90)	Median	0.74	0.72	0.73	0.78	0.71	0.73	0.73	0.79
	Average	1.28	1.16	1.20	1.24	1.20	1.20	1.16	1.17
	% Classification with BF > 3	12.22	6.67	8.89	10.00	11.11	12.22	11.11	7.78
	% Classification with BF < 0.33	21.11	25.56	24.44	23.33	27.78	27.78	22.22	25.56
Species (157)	Median	1.17	1.29	1.17	1.33	1.39	1.28	1.15	1.31
	Average	1.61	1.62	1.59	1.63	1.62	1.64	1.54	1.50
	% Classification with BF > 3	23.57	22.93	22.29	22.93	22.29	21.66	17.20	15.92
	% Classification with BF < 0.33	16.56	15.92	15.92	15.29	16.56	15.29	15.92	15.29

Differences between microbiome profiles predicted by each method were assessed using Bayes factors (BFs) calculated for each taxonomic classification within each taxonomic level, including phylum, genus, and species. The median, average, and percentage of taxonomic classifications with BF values showing strong evidence for the null and alternative hypotheses are tabulated. BF > 3 indicates strong evidence in favor of the evaluated method being similar to the reference profile. Conversely, BF < 0.33 is considered evidence in favor of the alternative hypothesis, indicating strong differences between the DNA-shotgun method and the 16S pipeline methods. The highest median and average BF values are in bold

At the genus level, two clustered heatmaps of BFs were generated to improve visualization of the consistency between the DNA-shotgun and each evaluated 16S pipelinemethod, 61 taxonomic genera from the phylum Firmicutes are shown in figure 7, while 29 genera from the phyla Actinobacteria, Bacteroidetes, Proteobacteria, Tenericutes, and Verrucomicrobia are shown in figure 8. Hierarchical clustering of BF values separated the workflows into three distinct groups based on input structure: the first comprised HVR4-based pipelines, the second encompassed HVR2-based pipelines, and the third isolated the multi-region IR configurations. Overall, 11 of 90 taxonomic classifications at the genus level (12.2%) showed strong evidence of similarity to the ZB-Shotgun reference profile using pipelines based on a single HVR (LMT-DADA-HVR4 and Nephele-HVR2), while 10 genera were identified with BF values > 3 (11.1%) using integrated multiple HVRs (IR-1S). The average BF values observed across all 16S pipeline options evaluated at the genus level showed small variations, ranging from 1.16 (IR-1S and LMT-Deblur-HVR4) to 1.28 (LMT-DADA-HVR4), indicating a tendency toward low BF values for 16S profile predictions at this taxonomic level (Table 1).

Fig. 7.

Clustered heatmap of Bayes Factors at genus level, Firmicutes. Global view of the consistency between the 16S bioinformatics pipelines evaluated and the ZB-Shotgun reference profile at the genus level. Only Firmicutes are represented in this figure. Darker colors represent higher BF values and strong evidence of consistency across pipelines. BF values showing strong evidence for the null and alternative hypotheses are further investigated (see also Table 1). A threshold of BF > 3 indicates strong evidence in favor of the evaluated method being similar to the reference. Conversely, BF < 0.33 indicates strong differences between the DNA-shotgun method and the 16S pipeline methods evaluated

Fig. 8.

Clustered heatmap of Bayes Factors at the genus level. Actinobacterias, Bacteroidetes, Protobacteria, Tenericutes and Verrucomicrobiota. Global view of the consistency between the 16S bioinformatics pipelines evaluated and the ZB-Shotgun reference profile at the genus level. Actinobacterias, Bacteroidetes, Protobacterias, Tenericutes, and Verrucomicrobiota are represented in this figure. Darker colors indicate higher BF values and strong evidence of consistency across pipelines. BF values showing strong evidence for the null and alternative hypotheses are further investigated (see also Table. 1). A threshold of BF > 3 indicates strong evidence in favor of the evaluated method being similar to the reference. Conversely, BF < 0.33 indicates strong differences between the DNA-shotgun method and the 16S pipeline methods evaluated

Notably, five genera—Dorea, Dysosmobacter, Mediterraneibacter (phylum Firmicutes), Actinomyces (phylum Actinobacteria), and Butyricimonas (phylum Bacteroidetes)—achieved BF values greater than 3 across all tested 16S pipelines. Conversely, 14 of 90 taxonomic classifications showed BF values below 0.33 across all evaluated 16S pipelines; these included genera such as Lachnoclostridium, Blautia, Anaerotruncus (phylum Firmicutes), Asaccharobacter (phylum Actinobacteria), Acinetobacter, Escherichia, and Oxalobacter (phylum Proteobacteria), among others, indicating substantial significant divergence from the DNA-shotgun data.

Bacterial abundances of specific genera, such as Solobacterium, Clostridium (phylum Firmicutes), and Akkermansia (phylum Verrucomicrobia), were predicted with greater consistency relative to the ZB-Shotgun reference when pipelines utilizing HVR2 as input were considered (Figures 7 and 8).

At the species level, 37 of 157 taxonomic classifications (23.57%) showed strong evidence of similarity to the ZB-Shotgun reference profile when using pipelines based on a single HVR (LMT-DADA-HVR4) followed by LMT-Deblur-HVR4 and LMT-DADA-HVR2 (36 classifications; 22.93%). Additionally, 25 (15.92%) and 27 (17.2 %) species were identified with BF values > 3 using integrated multiple HVRs (IR-2S and IR-1S respectively). Across 16S pipeline options using single HVR, the average BF values are narrow, ranging from 1.59 (Nephele-HVR4) to 1.64 (Nephele-HVR2). This range is also higher than the BF values observed across 16S pipeline options using multiple HVRs (IR-1S and IR-2S) at this taxonomic level (Table 1). These results indicate that targeting HVR2 or HVR4 with either DADA2 or Deblur yields comparable species-level results that significantly outperform the Ion Reporter pipeline options. While multi-region consensus models offer architectural stability at higher taxonomic ranks, single-HVR open-source pipelines using HVR2 or HVR4 provide substantially higher accuracy and resolution for species-level microbiome profiling.

Discussion

Although PCR-based high-throughput amplicon sequencing introduces inherent biases regardless of the sequencing platform utilized [10], 16S amplicon sequencing remains a cornerstone for microbiome researchers due to its cost-effectiveness, low input DNA requirements, curated reference databases, and robust open-source analysis frameworks [24]. While recent advancements have accelerated the global adoption of next-generation sequencing (NGS) in clinical diagnostics [5], complex data analysis and pipeline standardization remain primary bottlenecks to routine integration into clinical workflows [6, 7]. This study addresses this critical gap by benchmarking and optimizing open-source bioinformatics workflows for the Ion Torrent S5 platform, which remains a highly relevant tool in translational medical research [21]. Our evaluation specifically targeted the genus and species levels; from a translational standpoint, these narrower taxonomic tiers are clinically determinants for identifying specific bacterial pathogens or characterizing dysbiotic signatures associated with human diseases.

Two consensus microbiome profile predictions can be generated using the 16S bioinformatics solution installed on Ion Reporter™ software (v5.12) within the Ion Torrent platform (ThermoScientific). It delivers a consensus profile predicted by integrating data from multiple amplicons sequenced simultaneously. For instance, three amplicons (targeting HVR2, HVR4, and HVR8) are amplified and sequenced when samples are prepared with a single set of primers (1S). Conversely, six amplicons (targeting HVR2, HVR4, and HVR8, alongside HVR3, HVR6-7, and HVR9) are processed when two sets of primers (2S) are used.

The Ion Reporter™ software has several limitations. This solution is based on QIIME1 (2010) [26] and the GreenGenes database [25]v13.5).

A major challenge for Ion Torrent users is the proprietary nature of the Ion Reporter™ software, which delivers a consensus profile predicted from integrated data across multiple amplicons sequenced simultaneously but runs on obsolete frameworks—specifically QIIME1 [26] and an outdated GreenGenes database [25]. Furthermore, users cannot incorporate study-specific metadata into analyses, and exported processed data lack information from prior analyses preventing downstream analysis with open-source tools [24]. This technical barrier necessitates the implementation of external open-architecture workflows capable of handling Ion Torrent data [24].

Although previous methodological studies have extensively evaluated the performance of state-of-the-art denoising algorithms such as DADA2 and Deblur [54], these benchmarks have relied heavily on paired-end reads typical of Illumina platforms. Consequently, there is limited validation for single-end read architectures generated by platforms such as Ion Torrent [17].

Our study revealed substantial variations among the 16S metagenomics workflows evaluated. We confirmed that raw read counts across targeted hypervariable regions (HVRs) are not uniformly distributed. Specifically, HVR8 and HVR9 were consistently underrepresented, whereas HVR2 and HVR4 achieved the highest read depths (Additional file 2: Figure S1), consistent with previous observations [17]. It is important to note that a substantial proportion of sequenced reads in the raw data files can be removed during quality control and noise filtering before taxonomic classification, and higher read depths can be advantageous.

Crucially, our findings show that the choice of denoising algorithm (DADA2 vs. Deblur) fundamentally alters alpha diversity estimates across all HVRs because of their distinct error-correction approaches. DADA2 uses a highly sensitive, sample-specific statistical error model to infer exact amplicon sequence variants (ASVs). While this approach maximizes detection of rare biological variants, it is inherently more likely to retain low-frequency sequencing artifacts, thereby inflating diversity metrics [54]. Conversely, Deblur employs a conservative, static error-correction threshold that prioritizes specificity. This effectively removes background noise but risks grouping closely related true biological sequences, artificially depressing perceived alpha diversity. Although these trends did not reach strict statistical significance in our sub-analyses, accounting for this algorithmic bias is vital, as alpha diversity indices are frequently correlated with clinical parameters and could otherwise lead to spurious medical conclusions [57]. Since both Nephele and LMT-Deblur utilize Deblur as their core denoiser under matched parameters, they were expected to yield highly consistent results. Our findings successfully confirmed this algorithmic consistency, thereby validating the robustness of our in-house pipeline against a standardized, peer-reviewed bioinformatics platform like Nephele. Unlike DADA2, which includes a specific parameter for Ion Torrent data (homopolymer_gap_penalty = -1), Deblur-based pipelines (Nephele and LMT-Deblur) lack an equivalent, built-in denoising optimization for this sequencing platform. While we aimed for maximum parameter consistency, taxonomic classification differences between Nephele and our in-house LMT pipelines can mainly be driven by the specific SILVA databases utilized for the taxonomic classifier: Nephele uses standard, unmodified reference databases provided by the platform (silva-138 -default-) while LMT-Deblur and LMT-DADA2 uses the most updated SILVA reference database available through QIIME2 at the time of this study and. (Additional file 9: Table S5).

We confirm that different sequenced 16S rRNA HVRs yield significant differences in taxonomic identification, using a Bayesian statistical approach that accounts for the compositionality of microbiome data. Our results, along with other studies using Ion Reporter software, show that different regions vary in their taxonomic utility [17, 24], likely due to primer bias, differential HVR sequence length, and HVR sequence uniqueness across bacterial taxa [24]. HVR8 and HVR9 have demonstrated poor performance in identifying species in the mock community, while HVR2 and HVR4 show the most consistent results [14, 17, 24]. This pattern strongly suggests that the Ion 16S™ Metagenomics Kit does not perform well at the end of the 16S rRNA gene [17].

Our results suggest that while multi-region approaches (IR software options) offer stability at higher taxonomic ranks, single-HVR pipelines using HVR2 or HVR4 provide superior resolution for species-level microbiome profiles when processed with updated open-source bioinformatics tools. It has been suggested that examining redundant data from multiple hypervariable regions may help elucidate the true microbial makeup of the sample [24]. This could be achieved by combining results from multiple HVRs using generalized linear models (GLMs) [24] or by combining sequences from multiple PCR amplicons to provide a single, unified set of taxonomic profiling results at the sequence level [58, 59]. The latter option is only possible if the sequence of each primer is available. To our knowledge, besides an application note reported by Thermofisher, there are no other published studies using the primers in the Ion 16S™ Metagenomics Kit.

This study has several limitations. First, the benchmark is based on a single reference material derived from healthy donor fecal samples. Consequently, these results may not to samples with distinct community structures. Second, due to the limited sample size (n = 3 biological replicates per condition), robust adjustment for confounding variables was not feasible and a conventional sample size or power calculation is not applicable to this design, since it would require an estimate of biological variability obtained from independent biological units, which are not present when replicates are technical replicates of a single reference standard. For this reason, the replicates in this study quantify the technical reproducibility of each pipeline against a common reference profile rather than sampling biological variability across a population. Accordingly, the conclusions of this benchmark should be read as statements of relative consistency among pipelines against a shared reference, and not as estimates of biological. Furthermore, the narrow confidence intervals observed in our Poisson model reflect high sequencing read counts rather than true biological precision, as the model treats each read as an independent observation. This approach likely underestimates biological uncertainty, and the strength of these associations should be interpreted cautiously.

Third, all bioinformatics pipelines were executed using default or author-recommended parameters. Pipeline performance with custom-tuned parameters may differ; however, we believe that the vast majority of users employ either default or author-recommended settings. We therefore aimed to compare pipelines under these typical conditions to match the most plausible use scenarios as recommended [54] Although this benchmark focused specifically on the QIIME2 ecosystem paired with the SILVA database and a Naïve Bayes classifier, we highly encourage researchers to employ the most recently updated databases available to ensure taxonomic accuracy. Fourth, fecal microbiota analysis captures only a portion of the intestinal microbial ecosystem and may not fully represent mucosa-associated communities. In addition, species-level assignments based on 16S rRNA sequencing should be interpreted with caution due to the limited taxonomic resolution of this approach. Therefore, some species-level classifications should be considered approximate taxonomic assignments rather than definitive identifications.

Finally, the ZymoBIOMICS™ Fecal Reference (D6323) is derived from pooled human fecal material rather than a synthetic community of defined organisms at known absolute abundances; its compositional profile is generated by a proprietary bioinformatics pipeline (Centrifuge v1.0.4, ZymoBIOMICS™ Metagenomics Pipeline ver. 1.0.3, 2021). Accordingly, all Bayes factors, Aitchison distances, and pipeline rankings reported here reflect agreement with this ZB-Shotgun reference profile rather than accuracy against an independently verified absolute standard. This aligns with the well-documented pipeline-dependence of shotgun metagenomic profiling, in which different classifiers applied to identical raw data yield substantially divergent taxonomic compositions [55, 56] empirically confirmed this by re-analyzing the same D6323 shotgun reads with SqueezeMeta [52], which produced a substantially different profile (Additional file 8: Figure S4). As the ZB-Shotgun profile represents the standard reference characterization distributed by the manufacturer for D6323, and given that no universally accepted alternative benchmark exists for this material, results should be interpreted as reflecting relative consistency among 16S pipelines against one specific shotgun reference profile, not absolute taxonomic accuracy.

Conclusion

This study establishes a reproducible, traceable, and open-source bioinformatics framework for analyzing single-end 16S rRNA amplicon sequencing data from the Ion Torrent platform, offering a robust alternative to proprietary workflows. Current pipelines using QIIME2 and the SILVA database significantly outperform the standard Ion Reporter™ software (v5.12) from Thermo Fisher Scientific for human gut microbiome profile predictions.

We found that while multi-region approaches (IR Software options) offer stability at higher taxonomic ranks, single-HVR pipelines using HVR2 or HVR4 provide superior resolution for species-level microbiome profiles when paired with updated bioinformatics tools. This work aims to guide researchers in choosing the pipeline most suited to their goals and supports other studies that suggest similar pipeline choices. Open-source and in-house bioinformatics solutions enable functional extensions, modular implementations, automation, and greater adaptability to updates in the bioinformatics tools and databases used for human gut microbiome research.

Our benchmarking also underscores that the choice of denoising algorithm (DADA2 vs. Deblur) significantly skews alpha diversity calculations, regardless of the HVR selected. Because alpha diversity metrics are often correlated with clinical endpoints, pipeline-induced artifacts risk generating spurious translational conclusions.

These variations underscore the urgent need for the microbiome bioinformatics community to establish standardized consensus methodologies that preserve data quality and enable cross-study meta-analyses. Ultimately, this work provides a cost-effective, adaptable, and platform-specific guide that empowers researchers to select optimal workflows aligned with their taxonomic goals, enabling the successful translation of Ion Torrent data into routine healthcare and clinical research environments without reliance on restrictive software ecosystems.

Authors’ contributions

RJ Designed the experiment, performed bioinformatics and statistical analyses of data and was a major contributor in writing the manuscript. TE and RL Contributed in the design of this work, performed statistical analyses of data and contributed to manuscript writing and editing. NB Contributed in the implementation of bioinformatics data analysis tools needed for this work and performed bioinformatics analysis of data. TE and RG performed DNA isolation, barcoding, library preparation and sequencing of samples. RG contributed to manuscript editing. AH Contributed in the publication process and raw data availability. FI Performed bioinformatics analysis of data and contributed to manuscript writing and editing. FV Contributed in the design of this work, results interpretation, advised RJ and substantively revised the manuscript. All authors read and approved the final manuscript.

Funding

Fundacion Arturo Lopez Perez, internal grant N° FALP-LMT-2021 and FALP-LMT-2022.

Data availability

The datasets used during the current study, corresponding to the raw data and data tables from the reference mock are publicly and freely available by the provider in the Zymo Research website, microbiomics section ([https://www.files.zymoresearch.com/protocols/d6323-zymobiomics_fecal_reference_protocol.pdf] (https://www.files.zymoresearch.com/protocols/d6323-zymobiomics_fecal_reference_protocol.pdf)). All data supporting the findings of this study are available within the paper and its supplementary files. Raw sequencing data is publicly available at Sequence Read Archive (SRA) under accession number PRJNA1427200 ([https://www.ncbi.nlm.nih.gov/bioproject/?term=1427200] (https://www.ncbi.nlm.nih.gov/bioproject/?term=1427200)). The code generated and utilized during the current study, is available in the Github repository linked to this work (https://github.com/n-bravo/16S-data-analysis).

Declarations

Ethics approval and consent to participate

Not applicable. No patients were recruited for this study, and no human participants, identifiable human data, or directly collected human specimens were involved. The analyses were performed using commercially available reference material. Therefore, ethics approval, informed consent, and the Declaration of Helsinki were not applicable.

Consent for publication

Not applicable.

Competing interests

The authors declare that they have no competing interests.