Updating urinary microbiome analyses to enhance biologic interpretation

Nazema Y. Siddiqui, Li Ma, Linda Brubaker, Jialiang Mao, Carter Hoffman, Lisa KarstensmedRxiv 2021 · DOI

Abstract

Objective: An approach for assessing the urinary microbiome is 16S rRNA gene sequencing, where a segment of the bacterial genome is amplified and sequenced. Methods used to analyze these data are rapidly evolving, although the research implications are not known. This re-analysis of an existing dataset aimed to determine the impact of updated bioinformatic and statistical techniques.

Methods: A prior Pelvic Floor Disorders Network (PFDN) study compared the urinary microbiome in 123 women with mixed urinary incontinence (MUI) and 84 controls. We used the PFDN’s unprocessed sequencing data of V1-V3 and V4-V6 16S variable regions, processed operational taxonomic unit (OTU) tables, and de-identified clinical data. We processed sequencing data with an updated bioinformatic pipeline, which used DADA2 to generate amplicon sequence variant (ASV) tables. Taxa from ASV tables were compared to OTU tables generated from the original processing; taxa from different variable regions after updated processing were also compared. After updated processing, data were analyzed with multiple filtering thresholds. Several techniques were tested to cluster samples into microbial communities. Multivariable regression was used to test for associations between microbial communities and MUI, while controlling for potentially confounding variables.

Results: Of taxa identified through updated bioinformatic processing, only 40% were identified originally, though taxa identified through both methods represented >99% of sequencing data in terms of relative abundance. When different 16S rRNA gene regions were sequenced from the same samples, there were differences noted in recovered taxa. When the original clustering methods were applied to reprocessed sequencing data, we confirmed differences in microbial communities associated with MUI. However, when samples were clustered with a different methodology, microbial communities were no longer associated with MUI.

Conclusions: Updated bioinformatic processing techniques recover many different taxa compared to prior techniques, though most of these differences exist in low abundance taxa that occupy a small proportion of the overall microbiome. Detection of high abundance taxa are not significantly impacted by bioinformatic strategy. However, there are different biases for less abundant taxa; these differences as well as downstream clustering methodology and filtering thresholds may affect interpretation of overall results.

Introduction

The study of the "urobiome"—the community of bacteria residing in the bladder—presents unique challenges compared to high-biomass environments like the gut. Because the bladder is a low-biomass environment, small variations in sequencing, filtering, and bioinformatic processing can be significantly magnified, potentially leading to different clinical interpretations.

Many foundational studies in urobiome research relied on bioinformatic tools that are now considered outdated, such as grouping sequences into Operational Taxonomic Units (OTUs) based on a 97% similarity threshold and utilizing the Greengenes reference database (last updated in 2013). This paper re-evaluates a major dataset from the HMS-ESTEEM study to determine if modern analytical standards—specifically Amplicon Sequence Variants (ASVs) and the SILVA database—alter the biological narrative regarding Mixed Urinary Incontinence (MUI).

Methods and Comparative Framework

The researchers conducted a "head-to-head" comparison between the original study's pipeline and an updated modern pipeline using data from 123 women with MUI and 84 age-matched controls.

Bioinformatic Evolution

The primary shift involved moving from OTU-based clustering to DADA2, which identifies Amplicon Sequence Variants (ASVs). Unlike OTUs, ASVs can distinguish sequences differing by as little as a single nucleotide, providing finer taxonomic resolution and superior error correction. Additionally, the team replaced the Greengenes database with SILVA, which offers more comprehensive representation of urinary bacteria.

Genomic Targets and Statistical Models

The study compared two different "windows" of the 16S rRNA gene: the V1-V3 and V4-V6 regions. Statistically, the authors compared the standard Dirichlet Multinomial Mixture (DMM) modeling against a newer, tree-based approach called Dirichlet Tree Multinomial Mixture (DTMM). The latter accounts for the evolutionary (phylogenetic) relationships between bacteria, theoretically providing a more nuanced view of community structure.

Key Findings

The Stability of Dominant Taxa vs. the Volatility of the "Rare Biosphere"

The updated pipeline identified 426 genera that were missed in the original analysis. However, these newly identified taxa accounted for less than 1% of the total sequence abundance. While the "heavy hitters" of the urobiome remain stable regardless of the tool used, the identification of rare bacteria is highly dependent on specific bioinformatic settings and filtering thresholds.

Sequencing Region Bias

The choice of genomic target introduced noticeable bias. The V4-V6 region was found to be more reliable for urobiome analysis. In contrast, the V1-V3 region was more susceptible to non-bacterial "noise" and was less effective at identifying Gardnerella, a genus frequently associated with female pelvic health.

The Fragility of Clinical Signals

Perhaps the most significant finding was the sensitivity of clinical associations to mathematical modeling. When the original clustering method (DMM) was applied to the newly processed data, the association between specific microbial communities and MUI was confirmed (specifically Clusters 2 and 6). However, when the more advanced DTMM model was applied, the association with MUI disappeared entirely. This suggests that the clinical "signal" in many microbiome studies may be fragile and dependent on the specific assumptions of the underlying statistical model.

Clinical Context and Confounding Variables

The re-analysis also highlighted the importance of deep clinical phenotyping. In the original study, age appeared to be a major factor influencing the urobiome. However, once the researchers created a more accurate composite variable for menopause and hormone therapy use, the independent effect of age vanished. This underscores that biological transitions (like menopause) are more influential than chronological age in shaping the urinary environment.

Discussion and Future Directions

While modern tools offer higher precision, the authors note the difficulty of establishing "ground truth" in low-biomass samples without culture-based gold standards or mock communities. The study confirms that filtering thresholds—the decisions of what constitutes "noise" versus "real" data—drastically impact results.

For future urobiome research, the authors recommend:

  1. Standardized Variables: Adoption of consistent clinical variables, such as vaginal pH and recent catheterization, to reduce confounding.
  2. Model Validation: Further validation of tree-based models (like DTMM) to determine if they are more accurate or if they risk "over-correcting" and obscuring true biological signals.
  3. Methodological Transparency: Researchers must be explicit about their bioinformatic choices, as these decisions can fundamentally change the clinical conclusions of a study.

Summary updated

Abstract

Abstract Objective An approach for assessing the urinary microbiome is 16S rRNA gene sequencing, where a segment of the bacterial genome is amplified and sequenced. Methods used to analyze these data are rapidly evolving, although the research implications are not known. This re-analysis of an existing dataset aimed to determine the impact of updated bioinformatic and statistical techniques. Methods A prior Pelvic Floor Disorders Network (PFDN) study compared the urinary microbiome in 123 women with mixed urinary incontinence (MUI) and 84 controls. We used the PFDN’s unprocessed sequencing data of V1-V3 and V4-V6 16S variable regions, processed operational taxonomic unit (OTU) tables, and de-identified clinical data. We processed sequencing data with an updated bioinformatic pipeline, which used DADA2 to generate amplicon sequence variant (ASV) tables. Taxa from ASV tables were compared to OTU tables generated from the original processing; taxa from different variable regions (e.g., V1-V3 versus V4-V6) after updated processing were also compared. After updated processing, data were analyzed with multiple filtering thresholds. Several techniques were tested to cluster samples into microbial communities. Multivariable regression was used to test for associations between microbial communities and MUI, while controlling for potentially confounding variables. Results Of taxa identified through updated bioinformatic processing, only 40% were identified originally, though taxa identified through both methods represented >99% of sequencing data in terms of relative abundance. When different 16S rRNA gene regions were sequenced from the same samples, there were differences noted in recovered taxa. When the original clustering methods were applied to reprocessed sequencing data, we confirmed differences in microbial communities associated with MUI. However, when samples were clustered with a different methodology, microbial communities were no longer associated with MUI. Conclusions Updated bioinformatic processing techniques recover many different taxa compared to prior techniques, though most of these differences exist in low abundance taxa that occupy a small proportion of the overall microbiome. Detection of high abundance taxa are not significantly impacted by bioinformatic strategy. However, there are different biases for less abundant taxa; these differences as well as downstream clustering methodology and filtering thresholds may affect interpretation of overall results.
BibTeX
@misc{siddiqui2021updating,
  author = {Nazema Y. Siddiqui and Li Ma and Linda Brubaker and Jialiang Mao and Carter Hoffman and Lisa Karstens},
  title = {{Updating urinary microbiome analyses to enhance biologic interpretation}},
  year = {2021},
  doi = {10.1101/2021.09.30.21264391},
  howpublished = {medRxiv},
}
Figure