Integrated analysis reveals strong reproducible signals within and across studies of the built environment

By integrating four diverse 16S rRNA datasets and applying a unified sampling ontology, this study demonstrates that the built environment microbiome exhibits strong, reproducible signals characterized by distinct soil-associated taxa on floors and human-associated taxa on hands and surfaces, which persist across different buildings, timepoints, and sequencing protocols without the need for batch correction.

The Gist

Imagine you are trying to figure out the "personality" of a house just by looking at the dust and germs living on its surfaces. You might think that every house is so different—different owners, different cleaning habits, different locations—that comparing the germs in one house to another would be like comparing apples to oranges.

This paper is like a group of scientists gathering four different "dust samples" from very different places: a hospital, a university dorm, a military dorm, and a private home. Their goal was to answer a big question: Is there a consistent pattern to the invisible world of germs inside our buildings, or is it just random chaos?

Here is the breakdown of their findings using simple analogies:

1. The "Three Zones" of a Building

The researchers realized that different studies used confusing names for their samples (like "desk," "bedrail," or "counter"). To fix this, they created a simple three-zone map for every building:

• The Hands: Where people touch directly (skin).
• The Hand-Associated Surfaces: The things hands touch often (doorknobs, desks, bedrails). Think of these as the "hand's shadow."
• The Floors: The ground.

2. The Big Discovery: The "Soil vs. Skin" Battle

When they looked at the data, they found a very strong, repeating pattern across all four different buildings, despite them being totally different places.

• The Floors are "Outdoors": The floors in every building were packed with germs that usually live outside in the soil. It's like the floor is a magnet for dirt from shoes. Even inside a hospital, the floor looked more like a garden patch than a sterile room.
• The Hands are "Humans": The germs on hands and the things hands touch were packed with skin bacteria. It's like a fingerprint made of microbes.
• The Result: You can almost always tell a floor sample from a hand sample just by looking at the bacteria. It's as distinct as telling the difference between a beach (sand/soil) and a swimming pool (water/human contact).

3. The "Time Travel" Test

The scientists looked at these buildings over time (months and years).

• The Good News: The pattern held up. The floors stayed "soil-like" and the hands stayed "human-like" over time.
• The Glitch: In one specific hospital time period, the pattern got a little fuzzy. The researchers suspect this was because the hospital staff changed their cleaning routine or the patient mix changed suddenly. It's like if someone suddenly started mopping the floor with a giant bucket of soil; the floor would look different for a while. But once they looked at all the data together, the true pattern re-emerged.

4. The "Hand vs. Hand-Shadow" Confusion

While it was easy to tell the floor from the hand, it was much harder to tell the difference between a hand and a desk (or bedrail).

• Why? Because hands touch desks constantly. They are constantly swapping germs back and forth. It's like trying to tell the difference between two people who are hugging; their clothes get mixed up. The signal was weak here, which makes sense biologically.

5. The "Technical Noise" Problem

Usually, when scientists combine data from different labs, the results get messy because everyone uses different microscopes, different soap, and different computers. This is called "batch effects."

• The Surprise: The researchers tried to use a special computer program to "clean up" these technical differences. But they found they didn't even need to! The biological signal (the difference between soil and skin) was so strong that it drowned out the technical noise. The "truth" of the building's microbiome was loud enough to be heard even through the static.

The Bottom Line

This paper tells us that buildings have a predictable microbial personality.

• Floors are the "outdoor" zone, full of dirt and soil bugs.
• Hands and touched surfaces are the "human" zone, full of skin bugs.

Even though every building is unique, this basic rule applies everywhere. This is great news for science because it means we can finally compare studies from different countries and different buildings and trust that they are talking about the same underlying reality. It also suggests that if we want to change the germs in a building (to make it healthier), we need to understand that the floor will always try to be "outdoors" unless we actively stop shoes from bringing dirt in.

Technical Summary

1. Problem Statement

Built environment microbiome research has identified numerous factors shaping indoor microbial communities. However, the reproducibility of these findings across different buildings, timepoints, and research groups remains unclear.

• Challenges: Cross-study comparisons are hindered by:
  • Technical Variation: Differences in sequencing protocols, extraction kits, and bioinformatics pipelines create "batch effects" that can overwhelm biological signals, especially in low-biomass environments.
  • Terminological Inconsistency: Studies use disparate naming conventions for similar sample types (e.g., "desk" vs. "countertop"), making meta-analysis difficult.
  • Environmental Heterogeneity: Variations in geography, occupancy, and cleaning regimes introduce noise.
• Gap: It is unknown whether a consistent biological signal exists across studies that can withstand these technical and environmental variations, or if batch effects render cross-study synthesis impossible.

2. Methodology

The authors integrated four publicly available 16S rRNA gene datasets representing distinct environments: a hospital, a university dormitory, an Air Force dormitory, and private residential houses.

• Ontology Construction: To standardize comparisons, the authors created a simple ontology grouping all samples into three categories:
  1. Hand: Direct human skin contact (including inner elbow for the Air Force study).
  2. Hand-associated surfaces: Environmental surfaces frequently touched by hands (e.g., desks, bedrails).
  3. Floor: Ground-level surfaces.
• Data Processing:
  • Denoising: Performed using QIIME 2 (DADA2) with no trimming/truncation.
  • Taxonomy: Assigned using a Naive Bayes classifier trained on the Silva 138 database.
  • Normalization: Log-normalization applied to correct for read depth.
• Statistical & Machine Learning Approaches:
  • Within-Study Analysis: Timepoint-specific comparisons using PERMANOVA, PCoA (Bray-Curtis), and Random Forest (10-fold cross-validation).
  • Cross-Study Analysis: Leave-One-Study-Out (LOSO) Random Forest models to test generalizability. Models were trained on three datasets and tested on the fourth.
  • Batch Correction: Applied DEBIAS-M, a method specifically designed for microbiome data, to assess if removing study-specific biases improved signal detection.
  • Visualization: "p-value vs. p-value" plots to visualize agreement between studies regarding taxa enrichment.

3. Key Contributions

• Standardized Ontology: Demonstrated that mapping diverse sample names to a unified "Hand/Hand-associated/Floor" ontology enables robust cross-study meta-analysis.
• Validation of Biological Signal: Provided evidence that biological niche partitioning (floor vs. human-contact surfaces) is a stronger driver of microbiome structure than technical batch effects.
• Evaluation of Batch Correction: Showed that for built environment data, the biological signal is often strong enough that complex batch correction (DEBIAS-M) does not significantly improve classification performance or effect sizes.

4. Key Results

A. Reproducibility Across Studies

• Floor vs. Hand/Hand-Associated: There was a strong, reproducible separation between floors and human-contact surfaces across all four studies.
  • PERMANOVA: Significant clustering by sample type ($p < 10^{-5}$) with study as a random effect. The effect of sample type outweighed the effect of the specific study.
  • Machine Learning: LOSO Random Forest models achieved high Area Under the Curve (AUC) scores for Hand vs. Floor comparisons (ranging from 0.865 to 0.921).
  • Taxonomic Drivers:
    • Floors: Consistently enriched with soil-associated taxa (e.g., KD4-96, 67-14, Skermanella, Sphingobacterium).
    • Hands/Surfaces: Enriched with skin-associated taxa (e.g., Lawsonella, Cutibacterium, Streptococcus).

B. Reproducibility Within Studies (Temporal)

• Results were generally consistent across timepoints within a study.
• Exception: One hospital timepoint (July 2013) showed weak separation between hands and floors, likely due to unmeasured environmental shifts (e.g., cleaning protocols). This highlights the risk of drawing conclusions from single timepoints.

C. Hand vs. Hand-Associated Surfaces

• This comparison yielded the weakest signal.
• While statistically significant in some studies (e.g., Air Force dorm), the separation was modest (AUCs 0.54–0.69).
• Interpretation: The bidirectional exchange of microbes between skin and frequently touched objects creates a "convergent" microbiome, making these two categories harder to distinguish than floors vs. hands.

D. Impact of Batch Correction

• Application of DEBIAS-M did not improve AUCs or effect sizes.
• Clustering by study remained significant even after correction, but the biological signal (sample type) remained the dominant factor. This suggests the underlying ecological structure is robust to technical noise.

5. Significance and Implications

• Ecological Stability: The built environment microbiome possesses a stable, reproducible structure defined by niche partitioning: floors act as reservoirs for outdoor/soil microbes, while hands and high-touch surfaces retain human skin signatures.
• Meta-Analysis Viability: The study proves that integrating disparate datasets is feasible and valuable. It allows researchers to distinguish true biological patterns from study-specific idiosyncrasies (e.g., the anomalous hospital timepoint).
• Future Applications:
  • Predictive Modeling: The reproducible signal supports the development of predictive tools for indoor air quality and health.
  • Forensics & Public Health: Standardized ontologies can help track pathogen spread or determine the origin of samples.
  • Intervention Monitoring: Establishing a "baseline" signal allows for better detection of deviations caused by cleaning practices, ventilation changes, or occupancy shifts.

Conclusion: Despite significant heterogeneity in building types, sampling methods, and sequencing protocols, the built environment microbiome exhibits a reproducible bacterial signal. Floors consistently harbor soil-derived taxa, while human-contact surfaces harbor skin-derived taxa, providing a reliable foundation for future cross-study research.