A Comparative Evaluation of Tools to Predict Metabolite Profiles From Microbiome Sequencing Data

Abstract

Metabolomic analyses of human gut microbiome samples can unveil the metabolic potential of host tissues and the numerous microorganisms they support, concurrently. As such, metabolomic information bears immense potential to improve disease diagnosis and therapeutic drug discovery. Unfortunately, as cohort sizes increase, comprehensive metabolomic profiling becomes costly and logistically difficult to perform at a large scale. To address these difficulties, we tested the feasibility of predicting the metabolites of a microbial community based solely on microbiome sequencing data. Paired microbiome sequencing (16S rRNA gene amplicons, shotgun metagenomics, and metatranscriptomics) and metabolome (mass spectrometry and nuclear magnetic resonance spectroscopy) datasets were collected from six independent studies spanning multiple diseases. We used these datasets to evaluate two reference-based gene-to-metabolite prediction pipelines and a machine-learning (ML) based metabolic profile prediction approach. With the pre-trained model on over 900 microbiome-metabolome paired samples, the ML approach yielded the most accurate predictions (i.e., highest F1 scores) of metabolite occurrences in the human gut and outperformed reference-based pipelines in predicting differential metabolites between case and control subjects. Our findings demonstrate the possibility of predicting metabolites from microbiome sequencing data, while highlighting certain limitations in detecting differential metabolites, and provide a framework to evaluate metabolite prediction pipelines, which will ultimately facilitate future investigations on microbial metabolites and human health.