Two-thirds of reported gut microbiome-disease associations are not immediately reproducible Braden T Tierney1,2,3,4, Yingxuan Tan1, Zhen Yang2,3,4, Bing Shui5, Michaela J Walker6, Benjamin M Kent7, Aleksandar D Kostic2,3,4+, Chirag J Patel1+ +co-corresponding author Lead contacts: Chirag J Patel email@example.com Aleksandar D Kostic Aleksandar.Kostic@joslin.harvard.edu 1Department of Biomedical Informatics, Harvard Medical School, Boston, MA 02115, USA 2Section on Pathophysiology and Molecular Pharmacology, Joslin Diabetes Center, Boston, MA 02215, USA 3Section on Islet Cell and Regenerative Biology, Joslin Diabetes Center, Boston, MA 02215, USA 4Department of Microbiology and Immunobiology, Harvard Medical School, Boston, MA 02115, USA 5Department of Cancer Biology, Dana Farber Cancer Institute 6Independent 7US Marine Corps Reproducibly associating human gut microbes with disease is prerequisite to understanding the role of the microbiome in human health. We aimed to reproduce 581 species-level associations for 6 prevalent and well-studied diseases using 15 public cohorts (2,343 samples). We used “vibration of effects” to find robust disease indicators, reproducing only 162 associations (27.9%) and identifying a preponderance of “janus effects” -- where associations can be positive or negative depending on slight changes to modeling strategy. Published associations in type 1 and type 2 diabetes were particularly non-robust. We reproduced prior findings implicating bias, showing that features like medication (e.g. metformin) and body mass index (BMI) confound associations. Further, we highlighted the influence of other important variables like sequencing depth, glucose levels, cholesterol, and age. Overall, our results demonstrate that simple un-adjusted correlations or single modeling strategies in isolation are not sufficient to engender reproducible or biologically informative conclusions.