Gene-level metagenomic architectures across diseases yield high-resolution microbiome diagnostic indicators

Abstract

We propose microbiome disease “architectures”: linking >1 million microbial features (species, pathways, and genes) to 7 host phenotypes from 13 cohorts using a pipeline designed to identify associations that are robust to analytical model choice. Here, we quantify conservation and heterogeneity in microbiome-disease associations, using gene-level analysis to identify strain-specific, cross-disease, positive and negative associations. We find coronary artery disease, inflammatory bowel diseases, and liver cirrhosis to share gene-level signatures ascribed to the Streptococcus genus. Type 2 diabetes, by comparison, has a distinct metagenomic signature not linked to any one specific species or genus. We additionally find that at the species-level, the prior-reported connection between Solobacterium moorei and colorectal cancer is not consistently identified across models—however, our gene-level analysis unveils a group of robust, strain-specific gene associations. Finally, we validate our findings regarding colorectal cancer and inflammatory bowel diseases in independent cohorts and identify that features inversely associated with disease tend to be less reproducible than features enriched in disease. Overall, our work is not only a step towards gene-based, cross-disease microbiome diagnostic indicators, but it also illuminates the nuances of the genetic architecture of the human microbiome, including tension between gene- and species-level associations.

Introduction

The ecology of the human microbiome is known to be associated with both phenotype and environment^1,2. Here, we introduce “microbiome architectures”, which, analogous to human genetic architectures³, are the characteristics of the microbiome, which, as a group, correlate with human phenotype. More specifically, we compute architecture by identifying the complete set of associations between the microbiome and a given host disease. We hypothesize that these could potentially be jointly diagnostic for different aspects of host health^4,5,6,7. Moreover, identifying common—and distinct—architectures across diseases can shine light on the degree to which diseases share common etiologies. Achieving these ends, however, requires identifying how architecture changes across an array of human diseases in a manner that can easily be tested with in vivo or in vitro experiments.

Others have considered microbial community ecology across human individuals. Outside of single-disease metagenome-association studies, investigators have introduced the concept of the “enterotypes,” 3 hypothetically phylogenetically (at the phylum- and genus level) and functionally distinct groups of microbiome compositions. Enterotypes were initially identified across individuals from different backgrounds and countries^8,9. While interindividual microbiome variation and the presence of enterotypes is debated, their contribution to the field of comparative metagenomics remains fundamental^10,11,12. However, by comparing microbiome ecology across a range of host phenotypes, the concept and construction of architectures sidesteps the challenge of building grand views of a “normal” microbiome. Architectures instead enable the identification of specific (but still holistic) microbial factors associated with specific host phenotypes across sources of metagenomic variation.

At the heart of a metagenomic architecture rests a set of statistical associations between individual microbial features (e.g., species, pathways, or genes) and a given human phenotype. These associations are subject to the same challenges of any observational study, such as lower sample size (lack of power to detect associations) or confounding (e.g., not accounting for particular batch effects, geography, and/or diet). Lack of power and bias in observational studies (such as confounding) can lead to type 1 (false-positive) and type 2 (false-negative) errors.

Many studies use “meta-analyses” to aggregate and compare results across cohorts. There are a few approaches for carrying out a meta-analysis (e.g., random vs. fixed-effect meta-analyses¹³), and they provide a way to estimate an “overall” association size. Historically, they have been deployed for both randomized and observational research¹⁴, such as to aggregate effects across clinical trials¹⁵. Meta-analyses are emerging in the microbiome and have been used to discover new microbiome-disease associations in, for example, colorectal cancer^6,16,17,18.

However, meta-analyses are still potentially subject to confounding effects due to variable model specification strategies that occur in individual studies. In most situations, investigators choose a set of measured and potential confounding variables to adjust for in a model based on a prior hypothesis between the nature of the association between the independent and dependent variables. However, when the exact mechanism of potential confounding is unknown, the choice of potential measured confounders to adjust for in a model is arbitrary. The degree to which variation in model specification (e.g., adjusting for certain confounders and not others) changes the relationship between dependent and independent variables has been described as “Vibration of Effects” (VoE)^19,20,21. Modeling VoE enables researchers to identify not just that a result is irreproducible among certain model specifications, but potentially why that is the case²⁰. Briefly, we hypothesize that the larger the variation of associations that occur due to measured confounding and model choice, the less robust an association is. In other words, a robust association should persist across all or most configurations of study designs and model choices.

To be clear, we do not claim to identify the “best” method for computing architectures. Rather, we aim to propose architectures as a concept and demonstrate one method for their identification that controls for inconsistency in model output due to model specification. There are many options for computing the association between a disease and microbiome feature, ensuring these associations are robust, and meta-analyzing across datasets. Each of these steps rests upon volumes of biostatistics and microbiome literature. For example, a range of methods are used in the microbiome, from nonparametric tests to complex machine learning, like random forests.

Here, we propose one avenue for the identification of robust, multidata-type, microbial architectures in human disease by applying an analytic framework that considers a vast array of model specifications. We quantified the shared and distinct microbiome-disease architectures for seven prevalent diseases. We used the results of our meta-analysis and VoE pipeline to build high-resolution, robust multidisease architectures for seven diseases (adenoma, colorectal cancer (CRC), liver cirrhosis (CIRR), inflammatory bowel diseases (IBD), type 2 diabetes (T2D), otitis, and atherosclerotic cardiovascular disease (ACVD)), with a novel emphasis on gene-level, cross-disease associations. We specifically chose to examine otitis as a form of negative biological control, as, to our knowledge, it has limited reported association with the gut microbiome, and we expected it to have a negligible metagenomic architecture….