New Statistical Model Maps How Air Pollution Reshapes the Human Microbiome

Fine particulate matter, the invisible haze of particles smaller than 2.5 microns that drifts from traffic, industry, and wildfire smoke, has long been linked to respiratory and cardiovascular disease. What scientists increasingly suspect is that part of that harm travels through an unexpected route: the trillions of microbes that inhabit our bodies. A new statistical framework called SpaMixed, developed by a team of biostatisticians and clinicians at New York University Grossman School of Medicine and collaborators, aims to sharpen the lens through which researchers can detect exactly which microbial species respond to environmental pollution. Published in Genome Medicine, the method addresses a stubborn analytical problem that has quietly undermined microbiome studies for years.

The challenge is fundamentally spatial. People who live near one another tend to share not only the same air but also similar diets, housing, healthcare access, and lifestyles. When researchers collect microbiome samples across a city or a region, the samples from neighboring locations are not statistically independent. Standard regression models assume that each observation stands alone, so this spatial dependence can trick conventional analyses into flagging microbial taxa as pollution-associated when they are really just markers of shared geography. Conversely, genuine signals can be masked by the noise introduced when large clusters of similar samples dominate a dataset.

A second, equally thorny problem comes from the ecology of microbes themselves. The hundreds or thousands of taxa measured in a single microbiome study do not fluctuate independently. Related organisms share evolutionary histories and metabolic functions, so their abundances rise and fall together. An analysis that treats each species as a separate statistical question ignores this web of correlations, inflating error and making it harder to distinguish a real exposure effect from background ecological coordination. SpaMixed confronts both problems simultaneously by building them into the model’s structure rather than hoping they average out.

At its core, SpaMixed is a Bayesian spatial mixed model designed for microbiome count data, the raw output of sequencing surveys that tally how many sequence reads map to each taxon. Because such counts are notoriously zero-heavy and overdispersed, the framework builds on a zero-inflated Poisson formulation, a standard device for data with more zeros than a simple Poisson distribution can accommodate. The innovation lies in the priors: the model employs conditional autoregressive, or CAR, priors, the same spatial smoothing machinery familiar from disease mapping and spatial epidemiology, applied along two dimensions at once. One CAR prior captures dependence across geographic regions, while a second captures ecological correlation across taxa, allowing information from neighboring places and related microbes to inform one another’s estimates.

The practical payoff of this architecture is a screening tool with strong feature selection behavior. In extensive simulation studies, the researchers generated synthetic microbiome datasets under a range of scenarios varying the strength of exposure effects, the number of taxa, and the degree of spatial and ecological dependence. Across these tests, SpaMixed achieved high true positive rates, meaning it reliably recovered the taxa genuinely affected by the simulated exposure, while keeping false positives low. The posterior estimates of exposure effects also carried reduced bias and mean square error compared with existing methods, including ANCOM-BC, a widely used bias-corrected compositional approach, and MaAsLin, a popular multivariable association model.

The comparison with established tools matters because microbiome statistics is a crowded field. Compositional methods address the fact that sequencing data are relative, constrained to sum to a fixed total, which distorts naive abundance comparisons. Penalized and linear-model approaches manage covariates and multiple testing through local false discovery rates or Benjamini-Hochberg corrections. Each of these tools, however, was built without explicit spatial awareness. SpaMixed does not discard their insights but layers spatial and taxonomic dependence on top, estimating effects through a Bayesian computation pipeline that leans on the integrated nested Laplace approximation, a fast alternative to the slower Markov chain Monte Carlo sampling that has historically made large spatial Bayesian models impractical.

To demonstrate real-world utility, the team applied SpaMixed to two studies examining fine particulate matter exposure. The first drew on the Food and Microbiome Longitudinal Investigation, or FAMiLI, a New York City cohort with residential location information that could be aggregated by postal code and linked to PM2.5 exposure estimates. The second was a lung microbiome study at NYU Langone Health, approved by the institutional review board with written informed consent from all participants, in which bronchoscopic samples revealed the microbial communities living deep in the airways of individuals exposed to varying pollution levels.

In both datasets, exploratory principal component analyses of centered log-ratio transformed abundances showed spatial structure, with microbial community gradients varying across geographic areas, which is precisely the pattern that confounds naive models. When SpaMixed was applied, it identified a set of PM2.5-associated taxa that the authors describe as biologically plausible, and supplementary analyses of p-value rank distributions from the competing methods suggest that several of the taxa SpaMixed flagged sat in regions of the distribution that conventional tools did not prioritize. The agreement between the spatial structure visible in the diversity maps and the model’s findings supports the argument that accounting for geography is not a statistical nicety but a substantive requirement in environmental microbiome research.

The implications extend well beyond air pollution. Environmental exposures of many kinds, from water contaminants to dietary chemicals to climate-driven shifts in local ecology, leave traces in microbial communities, and nearly all of these exposures are spatially patterned. A framework that separates genuine exposure effects from geographic clustering could improve the reliability of studies linking the microbiome to cancer risk, lung disease, and other outcomes. The work was supported by multiple grants from the National Institutes of Health, reflecting sustained investment in methods that can translate the microbiome from a descriptive curiosity into a measurable mediator of environmental health effects.

For the field, SpaMixed represents a maturing moment: microbiome science is moving past the era where simply cataloging species was enough, toward rigorously attributing community changes to specific causes. The method’s open-access publication, its simulation benchmarks, and its demonstrated performance on two distinct exposure studies give other researchers a concrete tool to adopt. As sequencing costs fall and cohorts grow, the statistical machinery connecting pollution maps to microbial readouts will determine whether the field’s boldest hypotheses, that the microbes within us translate the environment into biology, survive scrutiny. Models like this one are how that scrutiny gets done.

Subject of Research: Spatial Bayesian modeling of microbiome count data to identify taxa associated with environmental exposures such as fine particulate matter

Article Title: Spatial mixed models for assessing environmental exposure effects on the microbiome

Article References: Kim, S., Wang, C., Kwak, S., Darawshy, F., Bain, A., Segal, L. N., Ahn, J., & Li, H. (2026). Spatial mixed models for assessing environmental exposure effects on the microbiome. Genome Medicine. https://doi.org/10.1186/s13073-026-01770-3

Image Credits: AI Generated

DOI: 10.1186/s13073-026-01770-3

Keywords: microbiome, spatial mixed model, PM2.5, air pollution, Bayesian inference, conditional autoregressive prior, zero-inflated Poisson, environmental exposure, ANCOM-BC, feature selection, Genome Medicine, biostatistics