Pattern recognition classification of the site of nephrotoxicity based on metabolic data derived from proton nuclear magnetic resonance spectra of urine

The computer-based pattern recognition procedures of nonlinear mapping and principal-component analysis have been applied to analyze 1H NMR-generated metabolic data on the biochemical effects of 15 acute nephrotoxin treatments affecting the renal cortex and/or renal medulla in rats. The 1H NMR signal intensities for 16 urinary metabolites representative of several major intermediary biochemical pathways were estimated using either a simple semiquantitative scoring system or complete peak intensity quantitation. NMR-derived data were treated as input coordinates in a multidimensional metabolic space and were analyzed by pattern recognition methods through which the dimensionality was reduced for display and categorization purposes. Different nephrotoxin treatments were initially classified using semiquantitative metabolite scores on the basis of their 1H NMR-detectable biochemical effects, and a good separation of renal cortical toxin treatments from renal medullary toxin treatments was achieved. The refinement of using exact peak heights rather than metabolic data scores utilized the available metabolic information more fully and provided a unique classification of each type of toxin according to its pattern of biochemical effects and site of toxic action. Principal-component analysis provided consistently better results than did nonlinear mapping in terms of discrimination between different sites of toxicity, and maps generated from correlation matrices gave improved discrimination, compared with those based directly on the original metabolic data. A comparison between the use of an added internal quantitation standard (3-trimethylsilyl-[2,2,3,3-2H4]-1-propionate) and independently determined glucose excretion rates for scaling to the NMR-detected urinary glucose levels demonstrated that the consistent classification of site-specific nephrotoxicity was independent of the quantitation standard used. This study has provided a rigorous assessment of data processing, relative quantitation, and pattern recognition methods, and the utility of applying these methods to the classification of NMR-derived toxicological data. The considerable potential of the NMR-pattern recognition approach in the assessment of nephrotoxicity has also been confirmed with the discovery of new combinations of molecular markers of renal cellular damage.