Application of pattern recognition methods to the analysis and classification of toxicological data derived from proton nuclear magnetic resonance spectroscopy of urine.

A computer-based pattern recognition (PR) approach has been applied to the classification and interrogation of 1H NMR-generated urinalysis data, in a variety of experimental toxicity states in the rat. 1H NMR signal intensities for each endogenous urinary metabolite were regarded as coordinates in multidimensional space and analyzed using PR methods, through which the dimensionality was reduced for display and categorization purposes. The changes in the NMR spectral patterns were characterized by 17 metabolic dimensions, which were then analyzed by employing the unsupervised learning methods of hierarchical cluster analysis, two-dimensional nonlinear map (NLM) analysis, and two- or three-dimensional principal components analysis (PCA). Different types of toxin (hepatotoxins and cortical and papillary nephrotoxins) were classified according to NMR-detectable biochemical effects. PCA provided consistently better results than NLMs in terms of discrimination of toxicity type, and maps based on correlation matrices also gave improved discrimination over those based on raw data. Various refinements in the data analysis were investigated, including taking NMR urinalysis data at three time points after exposure of the rats to six different nephrotoxins, as well as employing a dual-scoring system (time and magnitude of change). The maps generated from the time-course information produced the best discrimination between nephrotoxins from different classes. The robustness of the classification methods (in particular NLMs and PCA based on correlation matrices) and the influence of the addition of new scored biochemical data, reflecting dose-response situations, nutritional effects on toxicity, sex differences in biochemical response to toxins, the addition of a new toxin class (cadmium chloride, a testicular toxin and renal carbonic anhydrase inhibitor), and an additional metabolite descriptor (creatine), to the PR analysis were also evaluated. Initial training set maps were fundamentally stable to the addition of new data, and both NLM and PCA methods correctly "predicted" the toxicological effects from NMR data for test compounds, suggesting that the approach using PR and 1H NMR urinalysis for the generation and classification of acute toxicological data has wide applicability.