Interactions between species introduce spurious associations in microbiome studies

Microbiota contribute to many dimensions of host phenotype, including disease. To link specific microbes to specific phenotypes, microbiome-wide association studies compare microbial abundances between two groups of samples. Abundance differences, however, reflect not only direct associations with the phenotype, but also indirect effects due to microbial interactions. We found that microbial interactions could easily generate a large number of spurious associations that provide no mechanistic insight. Using techniques from statistical physics, we developed a method to remove indirect associations and applied it to the largest dataset on pediatric inflammatory bowel disease. Our method corrected the inflation of p-values in standard association tests and showed that only a small subset of associations is directly linked to the disease. Direct associations had a much higher accuracy in separating cases from controls and pointed to immunomodulation, butyrate production, and the brain-gut axis as important factors in the inflammatory bowel disease.Author summary: Microbiomes are important for better health, sustainable agriculture, and climate management. Since experimental studies of natural microbial communities are often challenging, microbiome wide association studies (MWAS) have been the primary method to reveal the connection between specific microbes and host phenotype. MWAS have established that many diseases are associated with a complex dysbiosis driven by a large number of microbes. We show that many of these associations may not reflect a mechanistic link with the disease, but arise instead due to interspecific interactions such as cross-feeding and competition for resources. We also propose a method grounded in the maximum entropy models of statistical physics to separate direct from indirect associations. Using both synthetic and real microbiome data, we show that this method detects only direct associations, identifies most predictive features of microbiomes, and avoids p-value inflation. We demonstrate the power of this method on one of the largest microbiome data sets on inflammatory bowel disease.