EVALUATING PREDICTIVE METABOLIC MODELLING TOOLS USING MICROBIOME SEQUENCING-METABOLOMIC DATA

Abstract

Metabolomic technologies have been utilized increasingly in tandem with sequencing technologies in microbiome studies. Combined microbiome-metabolome approaches have fueled the discovery of many convincing associations between disease-associated metabolites and the microbial taxa and genes driving their variation, and new computational tools to integrate these datasets are just emerging. A subset of these tools can mathematically predict the metabolic activity of the microbiome using the genomic content encoded in the microbiome. If these tools are accurate, they could allow researchers to circumvent the difficult and time-consuming process of performing metabolomics experiments by generating metabolite data using microbiome sequencing data alone, however few groups have validated their predictions against experimentally measured data. I sought to examine the performances of six metabolic modelling tools by calculating the correlations between predicted and experimentally measured metabolite profiles using paired sequencing and metabolomic data taken from the human gut and vaginal microbiomes. Out of all surveyed tools, MelonnPan generated predictions that were correlated best to experimentally measured data, although the observed correlations were generally poor across all tools and varied depending on sample characteristics such as donor disease status and sampling site. I did not observe any metabolites that were robustly predicted across all datasets and tools, and in addition all tools generally performed poorly in identifying differentially abundant metabolites. In this work, I have demonstrated the feasibility of predicting metabolites solely from microbiome sequencing data, while raising important limitations relating to the robustness of these predictions.