Peak-finding partial least squares for the analysis of ¹H NMR spectra

Metabonomic analysis of biofluids and extracts of biological tissues is increasingly being used to diagnose important metabolic differences induced by toxicity, disease processes or genetic differences. ¹H nuclear magnetic resonance (NMR) has been shown to be very useful for monitoring the low-molecular weight metabolite milieu typical of many systems. In this paper, a rigorous comparison of five different methods of data reduction and classification has been made. The five methods include principal components analysis (PCA) followed by linear discriminant analysis (LDA), PCA followed by logistic regression, a combined peak-picking-PCA and LDA algorithm, partial least squares (PLS), and a peak-picking PLS algorithm. To evaluate these five methods, a data set consisting of ¹H NMR spectra of the extracts of 29 malignant renal tumors and 17 normal tissues were analyzed. It was determined that peak-picking with PLS was the most efficient algorithm for correctly classifying this data set. Also, the peak-picking algorithm makes identification of the metabolites responsible for establishing class membership easier than with the other methods. A variety of different metabolites, including several amino acids and choline containing compounds were identified as markers for malignancy.