Abstract

We propose a novel statistical approach to improve the reliability of (1)H NMR spectral analysis in complex metabolic studies. The Statistical HOmogeneous Cluster SpectroscopY (SHOCSY) algorithm aims to reduce the variation within biological classes by selecting subsets of homogeneous (1)H NMR spectra that contain specific spectroscopic metabolic signatures related to each biological class in a study. In SHOCSY, we used a clustering method to categorize the whole data set into a number of clusters of samples with each cluster showing a similar spectral feature and hence biochemical composition, and we then used an enrichment test to identify the associations between the clusters and the biological classes in the data set. We evaluated the performance of the SHOCSY algorithm using a simulated (1)H NMR data set to emulate renal tubule toxicity and further exemplified this method with a (1)H NMR spectroscopic study of hydrazine-induced liver toxicity study in rats. The SHOCSY algorithm improved the predictive ability of the orthogonal partial least-squares discriminatory analysis (OPLS-DA) model through the use of "truly" representative samples in each biological class (i.e., homogeneous subsets). This method ensures that the analyses are no longer confounded by idiosyncratic responders and thus improves the reliability of biomarker extraction. SHOCSY is a useful tool for removing irrelevant variation that interfere with the interpretation and predictive ability of models and has widespread applicability to other spectroscopic data, as well as other "omics" type of data.