Monday, March 2nd, 2020-
Today, I finally checked out a paper that has been on my reading list for a while, titled The predictive power of the microbiome exceeds that of genome-wide association studies in the discrimination of complex human disease by Tierney et al., the preprint of which was published on December 31, 2019. The microbiome describes the sum total of all the genomes of the microbes that we play host to.
Given that the microbiome, which houses “tens of millions of microbial genes” compared to our 20,000-25,000 genes, is an ever-shifting landscape that reflects host interactions with the environment, whereas our own genomes are far less dynamic, it makes sense that our microbiomes could perhaps better reflect disease phenotypes than our own genomes. This might particularly be true for “multifactorial” disorders, which describe disorders that occur due to genes, lifestyle, and environmental factors. Obesity and diabetes would be good examples of multifactorial disorders.
To test this hypothesis, the authors analyzed data from a multitude of both microbial association studies (MAS) and genome-wide association studies (GWAS). Whereas MAS describes the features of the microbiome (for example, species abundance and diversity) that tightly correlate with a disease phenotype in comparison to healthy controls, GWAS looks at how well specific genetic variants called SNPs correlate with a disease phenotype.
Importantly, they compared the ability of the microbiome and genome to predict 13 common multifactorial diseases based on a metric called AUC. AUC stands for area-under-the-receiver-operator-curve, and it’s a common measure of how good a model is at discriminating between categories. In this case, it’s a measure of how good the model is at predicting risk for a disease based on microbiome or genome.
Here’s a great diagram explaining how to evaluate AUC, where an AUC of 1 indicates a perfect classifier model, and an AUC of 0.5 indicates the model is no better than random choice. It’s a common metric that one will encounter for many predictive models:
Tierney et al. found that the metagenome outperformed the genome in predicting human disease status. Look at this amazing result from their paper!
This is a very exciting finding, and raises questions about what other host phenotypes the microbiome could possibly be better at predicting than the host genome, and whether handling the microbiome data a different way (rather than looking at composition indices, looking at metabolic pathways) and rerunning analyses could result in even better predictive capability.
Importantly, microbiome-disease associations does not mean that the microbiome causes these diseases, just that the microbiome could be a good proxy for disease status. Inferring causality often relies on studies involving disease-associated microbiota transplantations into germ-free mice and assess the resulting mice phenotypes.
I feel so happy to be in this field and can’t wait until the number of metagenomic studies booms so we can perform meta-analyses with larger sample sizes. The more data, the better!
*contact me if I am not allowed to include the figure from the paper here and I will take it down*