Probabilistic Approach for Evaluating Metabolite Sample Integrity

The success of metabolomics studies depends upon the "fitness" of each biological sample used for analysis: it is critical that metabolite levels reported for a biological sample represent an accurate snapshot of the studied organism's metabolite profile at time of sample collection. Numerous factors may compromise metabolite sample fitness, including chemical and biological factors which intervene during sample collection, handling, storage, and preparation for analysis. We propose a probabilistic model for the quantitative assessment of metabolite sample fitness. Collection and processing of nuclear magnetic resonance (NMR) and ultra-performance liquid chromatography (UPLC-MS) metabolomics data is discussed. Feature selection methods utilized for multivariate data analysis are briefly reviewed, including feature clustering and computation of latent vectors using spectral methods. We propose that the time-course of metabolite changes in samples stored at different temperatures may be utilized to identify changing-metabolite-to-stable-metabolite ratios as markers of sample fitness. Tolerance intervals may be computed to characterize these ratios among fresh samples. In order to discover additional structure in the data relevant to sample fitness, we propose using data labeled according to these ratios to train a Dirichlet process mixture model (DPMM) for assessing sample fitness. DPMMs are highly intuitive since they model the metabolite levels in a sample as arising from a combination of processes including, e.g., normal biological processes and degradation- or contamination-inducing processes. The outputs of a DPMM are probabilities that a sample is associated with a given process, and these probabilities may be incorporated into a final classifier for sample fitness.