Abstract

We propose a novel data-driven approach aiming to reliably distinguish discriminatory metabolites from nondiscriminatory metabolites for a given spectroscopic data set containing two biological phenotypic subclasses. The automatic spectroscopic data categorization by clustering analysis (ASCLAN) algorithm aims to categorize spectral variables within a data set into three clusters corresponding to noise, nondiscriminatory and discriminatory metabolites regions. This is achieved by clustering each spectral variable based on the r(2) value representing the loading weight of each spectral variable as extracted from a orthogonal partial least-squares discriminant (OPLS-DA) model of the data set. The variables are ranked according to r(2) values and a series of principal component analysis (PCA) models are then built for subsets of these spectral data corresponding to ranges of r(2) values. The Q(2)X value for each PCA model is extracted. K-means clustering is then applied to the Q(2)X values to generate two clusters based on minimum Euclidean distance criterion. The cluster consisting of lower Q(2)X values is deemed devoid of metabolic information (noise), while the cluster consists of higher Q(2)X values is then further subclustered into two groups based on the r(2) values. We considered the cluster with high Q(2)X but low r(2) values as nondiscriminatory, while the cluster with high Q(2)X and r(2) values as discriminatory variables. The boundaries between these three clusters of spectral variables, on the basis of the r(2) values were considered as the cut off values for defining the noise, nondiscriminatory and discriminatory variables. We evaluated the ASCLAN algorithm using six simulated (1)H NMR spectroscopic data sets representing small, medium and large data sets (N = 50, 500, and 1000 samples per group, respectively), each with a reduced and full resolution set of variables (0.005 and 0.0005 ppm, respectively). ASCLAN correctly identified all discriminatory metabolites and showed zero false positive (100% specificity and positive predictive value) irrespective of the spectral resolution or the sample size in all six simulated data sets. This error rate was found to be superior to existing methods for ascertaining feature significance: univariate t test by Bonferroni correction (up to 10% false positive rate), Benjamini-Hochberg correction (up to 35% false positive rate) and metabolome wide significance level (MWSL, up to 0.4% false positive rate), as well as by various OPLS-DA parameters: variable importance to projection, (up to 15% false positive rate), loading coefficients (up to 35% false positive rate), and regression coefficients (up to 39% false positive rate). The application of ASCLAN was further exemplified using a widely investigated renal toxin, mercury II chloride (HgCl2) in rat model. ASCLAN successfully identified many of the known metabolites related to renal toxicity such as increased excretion of urinary creatinine, and different amino acids. The ASCLAN algorithm provides a framework for reliably differentiating discriminatory metabolites from nondiscriminatory metabolites in a biological data set without the need to set an arbitrary cut off value as applied to some of the conventional methods. This offers significant advantages over existing methods and the possibility for automation of high-throughput screening in "omics" data.