Bat anthropogenic roosting ecology influences taxonomic and geographic predictions of viral hosts

Integrating anthropogenic roosting ecology into machine learning models reveals that this trait significantly improves the prediction of bat virus hosting ability and identifies new undetected host species, particularly in Asia, highlighting its critical role in assessing spillover risks at human-wildlife interfaces.

The Gist

Imagine the world of viruses as a massive, chaotic library. Some books (viruses) are dangerous to humans, while others are harmless. Scientists are trying to figure out which animals are the "librarians" holding these dangerous books, so they can keep an eye on them and prevent the books from accidentally falling into human hands.

For a long time, scientists have been using a checklist of animal traits to guess who these librarians might be. They look at things like: How big is the animal? What does it eat? Where does it live geographically?

This new study asks a very specific, modern question: Does living in a human-made building change the odds?

The Core Idea: The "Human-House" vs. The "Cave"

Think of bats as the librarians. Some bats live in the "wild library" (caves, trees, and cliffs). Others have moved into the "human library" (attics, bridges, barns, and abandoned buildings).

The researchers wanted to know: Does moving into a human building make a bat more likely to be carrying a virus that could jump to us?

They treated "living in a human building" (anthropogenic roosting) as a new piece of data to add to their computer models. It's like adding a new ingredient to a soup recipe to see if it changes the flavor.

What They Did (The Recipe Test)

1. The Ingredients: They gathered data on nearly 1,300 bat species. They looked at their viruses, their diets, their sizes, and crucially, whether they sleep in caves or in our buildings.
2. The Simulation: They built computer models (like a super-smart guessing game) to predict which bats carry viruses.
   • Model A: Used all the old data (size, diet, location) but ignored whether the bat lived in a building.
   • Model B: Used all the old data plus the new "building-living" data.
3. The Taste Test: They compared the two models to see if Model B was better at guessing the truth.

The Surprising Results

Here is where the story gets interesting, because the answer isn't a simple "Yes, buildings are dangerous."

1. The "Flavor" Didn't Change Much
Adding the "living in a building" ingredient didn't make the computer model much more accurate overall. The model was already pretty good at guessing virus carriers just by looking at size and diet. It's like adding a pinch of salt to a soup that's already perfectly seasoned; it doesn't ruin it, but it doesn't completely transform it either.

2. But... It Found New "Librarians"
Even though the model's overall accuracy didn't skyrocket, the list of suspects changed slightly.

• Model A missed a few bats that live in buildings.
• Model B said, "Wait a minute! These specific building-dwelling bats are also likely to have viruses."

It's like a detective who already knows the main suspects but, by checking the security cameras of the local coffee shop (the buildings), realizes they missed a few people who hang out there.

3. The Geography Shift: Asia is the Hotspot
This is the most important part. When the researchers looked at where these new suspects were, they found a pattern.

• In the Americas and Africa, the "new suspects" were scattered.
• In Asia, the new suspects were concentrated in huge numbers.

It turns out that in Asia, many bats that live in human buildings were being overlooked by the old models. If you ignore the fact that they live in our houses, you miss a whole group of potential virus carriers in that region.

The Big Takeaway

Think of this study as updating a map.

• The Old Map: Showed us where to look for virus-carrying bats based on their size and diet. It was a good map.
• The New Map: Includes a layer showing "Bats living in human buildings."

The new map doesn't make the old one useless, but it highlights blind spots, especially in Asia. It tells us that if we want to find the next potential virus before it jumps to humans, we need to look not just at the deep forests, but also at the bats sleeping in our attics and under our bridges.

Why does this matter?
Because when bats live in our buildings, they are closer to us. They might share the same air, or their droppings might get on our clothes. By realizing that "living in a building" is a clue, we can focus our surveillance efforts on the right places and the right animals, potentially stopping a pandemic before it starts.

In short: Living in a human building isn't the only reason a bat might carry a virus, but it's a significant clue that helps us find the ones we were previously ignoring, especially in Asia.

Technical Summary

1. Problem Statement

Emerging infectious diseases (EIDs) are increasing, driven largely by human-wildlife interfaces (HWIs). While predictive models using host traits (e.g., body mass, geography, diet) are standard for identifying potential viral reservoirs, they often rely on broad proxies for human-wildlife contact, such as general urban presence or human population density. These proxies fail to capture specific physical interactions, such as wildlife occupying human-made structures (anthropogenic roosting).

The authors address a critical data gap: the lack of integration of specific anthropogenic roosting ecology (the ability of a species to roost in buildings, bridges, mines, etc.) into predictive models of bat viral outcomes. The study investigates whether incorporating this specific trait improves the prediction of viral hosting ability, viral richness, and the identification of undetected host species, particularly in the context of spillover and spillback risks.

2. Methodology

The study employed a machine learning approach to analyze bat-virus associations across a global scale.

• Data Sources:

  • Virus Data: Extracted from the VIRION database (Global Virome in One Network), restricted to NCBI-verified species and associations confirmed by PCR or isolation (excluding serological data to avoid cross-reactivity bias).
  • Host Data: Mapped to 1,279 extant bat species (86% of Chiroptera) using the latest phylogeny.
  • Anthropogenic Roosting Data: A novel binary dataset (Betke et al., 2024) classifying species as either "anthropogenic roosting" (using human-made structures) or "exclusively natural roosting." This covered 1,029 species.
  • Covariates: Included 63 additional traits from the COMBINE, PanTHERIA, and EltonTraits databases (e.g., life history, foraging, biogeography, body size) and geographic variables (human population density, climate).
  • Sampling Bias Control: Citation counts (total and virus-related) for each species were calculated via PubMed to control for sampling effort.

• Response Variables:

  1. Virus Hosting Status: Binary (host of at least one virus vs. non-host).
  2. Zoonotic Hosting Status: Binary (host of at least one zoonotic virus vs. non-host).
  3. Viral Family Richness: Count of viral families.
  4. Zoonotic Viral Family Richness: Count of zoonotic viral families.

• Modeling Approach:

  • Algorithm: Boosted Regression Trees (BRT) using the gbm package in R.
  • Validation: 50 random train/test splits (80/20) with 10-fold cross-validation to prevent overfitting.
  • Comparison: Models were run with and without the anthropogenic roosting variable to assess changes in:
    • Variable importance rankings.
    • Model performance metrics (AUC, sensitivity, specificity, pseudo-$R^2$).
    • Predictions of undetected hosts (species currently untested or unconfirmed).
  • Spatial Analysis: Predicted host distributions were mapped and scaled by total bat species richness to identify geographic hotspots.

3. Key Contributions

• Novel Trait Integration: This is one of the first studies to explicitly integrate fine-scale anthropogenic roosting data (rather than broad urbanization proxies) into macroecological models of viral host prediction.
• Disentangling Sampling Bias: The study rigorously controls for citation bias, demonstrating that the observed patterns are not merely artifacts of increased sampling of urban-adapted species.
• Geographic and Taxonomic Refinement: It reveals that while overall model performance metrics do not drastically change, the identity and location of predicted undetected hosts shift significantly when roosting ecology is included.

4. Key Results

• Variable Importance:

  • Anthropogenic roosting was ranked as a moderately important predictor.
  • It was most important for predicting virus hosting status (median rank 23.5/63) and less important for viral richness metrics.
  • It was less important than human population density but more important than most family, diet, and foraging traits.
  • Partial dependence plots showed a weak but observable positive effect: anthropogenic roosting bats have a slightly higher probability of hosting at least one virus compared to natural roosting bats, but no significant difference in viral richness.

• Model Performance:

  • Including the roosting variable did not significantly improve overall model performance metrics (AUC, sensitivity, specificity, or pseudo-$R^2$) compared to models excluding it. The models were already performing well with existing traits.

• Predictions of Undetected Hosts:

  • Despite similar performance metrics, the inclusion of roosting ecology expanded the list of predicted undetected hosts.
  • Models with roosting data predicted 7 additional virus hosts (all anthropogenic roosting) that were missed by models without this trait.
  • For zoonotic hosts, the difference was smaller (2 additional species), suggesting roosting ecology is more critical for general viral hosting than specifically for zoonotic potential.
  • Taxonomic Bias: The newly predicted hosts were predominantly in the families Vespertilionidae and Phyllostomidae.
  • Conservation Status: A significant portion of the newly predicted anthropogenic hosts (10 of 14) are threatened or near-threatened, highlighting a risk of unnecessary sampling on vulnerable species if these predictions are ignored.

• Geographic Hotspots:

  • While the Neotropics and Afrotropics had the highest absolute numbers of predicted hosts, these aligned with general bat diversity.
  • Crucial Finding: When scaled by bat diversity, Asia (specifically the Arabian Peninsula, North American deserts, and the perimeter of the Tibetan Plateau) emerged as a distinct hotspot for anthropogenic roosting hosts.
  • Excluding roosting data led to a significant underestimation of host species in Asia.

5. Significance and Implications

• Refining Surveillance Strategies: The study demonstrates that relying solely on broad urbanization proxies or existing trait models may miss specific high-risk species in Asia and other regions where bats co-habit human structures.
• Conservation and Ethics: By identifying that many predicted undetected hosts are threatened species, the authors argue for targeted surveillance that minimizes invasive sampling of vulnerable populations.
• Human-Wildlife Interface: The results confirm that anthropogenic roosting is a meaningful interface for cross-species transmission. The distinct spatial patterns of these hosts suggest that viral discovery efforts should be geographically prioritized in areas where anthropogenic roosting bats are dense relative to total bat diversity.
• Future Modeling: The paper advocates for moving beyond simple "urban vs. rural" dichotomies in disease ecology, urging the inclusion of specific behavioral traits (like roosting substrate) to better understand spillover dynamics.

In conclusion, while anthropogenic roosting is not the single most dominant predictor of viral richness, its inclusion is essential for accurately identifying which species are likely hosts and where they are located, particularly for refining risk assessments in Asia and protecting threatened bat populations.