Analysis of Human Gut Metagenomes for the Prediction of Host Traits with Tree Ensemble Machine Learning Models

Abstract

The human gut microbiota is made of a myriad of microorganisms, among which not only bacteria but also archaea. Present at lower abundances, technically more challenging to quantify, and under-represented in databases, archaea are often overseen when describing the human gut microbiome. Nonetheless, the main archaeon in terms of prevalence and abundance is Methanobrevibacter smithii, family Methanobacteriaceae. It has been associated with various host phenotypes such as slow transit or diet habits. Remarkably, contrasting evidence shows an association between M. smithii and body mass index (BMI): it is enriched in lean or obese individuals according to population studies. Reasonable hypotheses relying on the metabolism of the archaeon support these conflicting findings. For instance, its slow replication time supports its association with slow transit. M. smithii and all members of the Methanobacteriaceae family are methanogens: their metabolism relies on the reduction of simple carbon molecules to methane. In the human gut, methanogenesis starts from bacterial fermentation products. In particular, H2 and CO2 are the primary substrates of M. smithii, formate can also be used but with a lower energy yield. By uptaking fermentation products, M. smithii can boost specific fermentation pathways, consequently affecting the production of short-chain fatty acids (SCFA). These byproducts of bacterial fermentation are absorbed by the host, where they mediate host energy and inflammatory metabolisms. Accordingly, its overall effect may de- pend on the fermentation potential of the gut microbiome, itself defined by the microbiome composition. Hence, M. smithii may influence its host by consuming fermentation products. Because we know so little about the interactions between M. smithii and fermenting bacteria, gaining knowledge on their diversity and specificity and the underlying mechanisms would improve our understanding of methanogens’ role in the human gut. This work aims at providing insights into the associations between M. smithii and gut bacteria. Due to the fastidiousness of methanogens’ culture, I performed a meta-analysis of human gut metagenomes using machine learning models. To decipher the variable interactions captured by the model, I developed a tool for interpreting tree ensemble models. My new method allowed me to infer biologically relevant associations between the methanogen and components of the human gut environment. In particular, I found a clear association between M. smithii and an uncultured family of the Christensenellales order, as well as members of the Oscillospirales order predicted to have a slow replication time and be associated with slow transit. Furthermore, predictions from the model revealed a gradient in relative abundances of a core group of taxa associated with the colonization of human guts by Methanobacteriaceae. This gradient generally followed microbiome composition types, i.e., enterotypes, previously correlated with human population traits. This suggests that associations between methanogens and phenotypes known to be associated with certain enterotypes, such as BMI is correlated with the ETB enterotype, may be spurious. Then, I further explored the association between M. smithii and members of the Christensenellales order. For this, I compared co-cultures of M. smithii with Christensenella minuta, a human gut iso- late of the Christensenellaceae family, and Bacteroides thetaiotaomicron, a common H2-producer from the human gut. Results demonstrated a syntrophy via H2-transfer between Christensenellaceae and the methanogen, accompanied by a switch in SCFA production. Altogether, my findings complement the current knowledge on interactions between the human gut methanogen M. smithii and fermenting bacteria. They support the hypothesis that M. smithii preferentially interacts with specific H2-producers in the human gut, e.g., members of the Christensenellales order, as well as a core group of bacteria favoring its colonization of the gut environment. Syntrophy may underlie the identified associations, with potential effects on bacterial fermentation. In addition, my method for interpreting machine learning models applies to all sorts of problems being studied with tree ensemble models. Thus, its potential in helping understand complex systems is not limited to the microbiome field and will hopefully appear useful to other researchers in the future.