Probabilistic Ranking Of Microbiomes with Taxa Selection to discover and validate microbiome function models for multiple litter decomposition studies

The overwhelming complexity of microbiomes makes it difficult to decipher functional relationships between specific microbes and ecosystem properties. While machine learning analyses have demonstrated an impressive ability to correlate microbial community composition with macroscopic functions, mechanisms that dictate model predictions are often unknown, and predictions often lack an assigned metric of uncertainty. In this study, we apply Bayesian networks to build on prior feature selection analyses and construct easy-to-interpret probabilistic models, which accurately predict levels of dissolved organic carbon (DOC) from the relative abundance of soil bacteria (16S rRNA gene profiles). In addition to standard cross-validation, we show that a Bayesian network model trained using samples from a pine litter decomposition study accurately predicts DOC of samples from an independent oak litter decomposition study, suggesting that mechanisms driving variation in soil carbon storage may be conserved across different types of decomposing plant litter. Furthermore, the structure of the resulting Bayesian network model defines a minimal set of highly informative taxa, whose abundances directly constrain the probability of high or low DOC conditions. Significant accuracy of the Bayesian network model with independent data sets supports the validity of the identified relationships between taxa abundance and DOC. SummaryUnderstanding the interplay between microbiomes and the environments they inhabit is a daunting task. While recent advances in gene sequencing technology provide a means of profiling the relative abundance of microbial species, the resulting data are noisy, sparse, and limited to small sample sizes. Despite these challenges, machine learning approaches have demonstrated a promising ability to discover patterns linking the microbiome with macroscopic behavior. However, most machine learning models applied to microbiome data do not estimate prediction uncertainty and provide little insight regarding how predictions are made. In this study, we couple machine learning approaches for feature reduction with Bayesian networks to model the relationship between the soil microbiome and dissolved organic carbon (DOC). We show that Bayesian networks are accurate and provide a transparent link between microbial abundance and DOC. To validate Bayesian networks, we demonstrate accurate predictions for held-out testing data and with data from independent decomposition experiments.