One of the consequences in analyzing biological data from noisy sources, such as human subjects, is the sheer variability of experimentally irrelevant factors that cannot be controlled for. This holds true especially in metabolomics, the global study of small molecules in a particular system. While metabolomics can offer deep quantitative insight into the metabolome via easy-to-acquire biofluid samples such as urine and blood, the aforementioned confounding factors can easily overwhelm attempts to extract relevant information. This can mar potentially crucial applications such as biomarker discovery. As such, a new algorithm, called Selective Paired Ion Contrast (SPICA), has been developed with the intent of extracting potentially biologically relevant information from the noisiest of metabolomic data sets. The basic idea of SPICA is built upon redefining the fundamental unit of statistical analysis. Whereas the vast majority of algorithms analyze metabolomics data on a single-ion basis, SPICA relies on analyzing ion-pairs. A standard metabolomic data set is reinterpreted by exhaustively considering all possible ion-pair combinations. Statistical comparisons between sample groups are made only by analyzing the differences in these pairs, which may be crucial in situations where no single metabolite can be used for normalization. With SPICA, human urine data sets from patients undergoing total body irradiation (TBI) and from a colorectal cancer (CRC) relapse study were analyzed in a statistically rigorous manner not possible with conventional methods. In the TBI study, 3530 statistically significant ion-pairs were identified, from which numerous putative radiation specific metabolite-pair biomarkers that mapped to potentially perturbed metabolic pathways were elucidated. In the CRC study, SPICA identified 6461 statistically significant ion-pairs, several of which putatively mapped to folic acid biosynthesis, a key pathway in colorectal cancer. Utilizing support vector machines (SVMs), SPICA was also able to unequivocally outperform binary classifiers built from classical single-ion feature based SVMs.