Chemical structure informing statistical hypothesis testing in metabolomics.

Abstract

Metabolomics has been shown as an effective tool to study various biological and biomedical phenotypes, whereas interrogating the inherently noisy metabolite concentration data with limited sample size remains a major challenge. Accumulating evidence suggests that metabolites' structures are relevant to their bioactivities. We present a new strategy to boost the statistical power of hypothesis testing in metabolomics by incorporating quantitative molecular descriptors for each metabolite. The strategy selects potentially informative summary molecular descriptors and outputs chemical structure-informed false discovery rates. The effectiveness of the proposed strategy is demonstrated by both simulation studies and a real application. In a metabolomic study on Alzheimer's disease, the posterior inclusion probability for summary molecular descriptors reaches 0.97. By incorporating the structure data, our approach uniquely identifies multiple Alzheimer's disease signatures, which are consistent with existing evidence. These results evidently suggest the value of the proposed approach for metabolomic hypothesis-testing problems. A code package implementing the strategy is freely available at https://github.com/HongjieZhu/CIMA.git.