Abstract:
|
Metabolomics profiles small molecules in microbial communities, which play crucial roles in signaling and immune system modulation. Identifying disrupted metabolic processes can provide new diagnostics from the microbiome, or new therapeutic targets. However, these data can be costly and difficult to obtain at scale, while shotgun metagenomic sequencing is readily available for populations of many thousands. Here, we develop a high-dimensional predictor of microbiome metabolite profiles given metagenomic sequencing data, incorporating biological prior knowledge in the form of microbial gene families or pathway abundances. We extended the two-stage ISIS procedure to predict multivariate metabolite compounds, which screens variables by iterative conditional regressions in the first stage, followed by penalized regression on the selected 'weak learners' in the second stage. This allowed us to accurately infer abundances of 2,472 microbial metabolites from a pool of 1 million microbial genes. Our results demonstrate that with sufficient dimensionality reduction, microbial gene products can provide useful insights into community metabolomic profiles when only metagenomics are available.
|