Thread by @segal_eran

Our new preprint analyzes largest metagenomic microbiome dataset to date (~30000 samples from Israel, and ~4000 U.S. samples)

We find robust associations of bacteria with several traits (e.g., age, BMI, HbA1C%) that replicate across continents and provide these as a public atlas

At this scale of data, we find remarkable correspondence between relative abundances of bacteria in the Israel and U.S. datasets

We find that the number of different bacteria that people have (alpha diversity) increases with age, and HDL cholesterol, and decreases with BMI, HbA1C%, fasting glucose and triglycerides

We quantify the fraction of the variance of different traits that is explained by alpha diversity (left column) and by the microbiome as a whole (right column)

"Explained" is in a statistical sense, not to be confused with causality which these results do not imply

We develop models that predict traits using only microbiome data and show that they work on held out samples from Israel (~3000 samples) and that microbiome models from Israel can also predict traits on U.S. samples

One of our key results: We sample smaller datasets from this large dataset and show that models that we develop from hundreds of samples have huge variation in their predictive power

This shows that we need thousands of microbiome samples to get robust models and associations

We show that the agreement between the set of species that associate with traits in the Israel and U.S. increases as we have more samples

This shows that many discrepancies between associations found in the literature may be due to small datasets used

Great work led by students Sigal Leviatan & Daphna Rothschild, and great collaboration (as always) with @oweissb and with @daytwohealth

biorxiv.org/cgi/content/sh…