Analytic Properties of Statistical Total Correlation Spectroscopy Based Information Recovery in 1H NMR Metabolic Data Sets

详细信息    查看全文

• 作者：Alexessander Couto Alves ; Mattias Rantalainen ; Elaine Holmes ; Jeremy K. Nicholson* ; Timothy M. D. Ebbels*
• 刊名：Analytical Chemistry
• 出版年：2009
• 出版时间：March 15, 2009
• 年：2009
• 卷：81
• 期：6
• 页码：2075-2084
• 全文大小：490K
• 年卷期：v.81,no.6(March 15, 2009)
• ISSN：1520-6882

文摘

Structural assignment of resonances is an important problem in NMR spectroscopy, and statistical total correlation spectroscopy (STOCSY) is a useful tool aiding this process for small molecules in complex mixture analysis and metabolic profiling studies. STOCSY delivers intramolecular information (delineating structural connectivity) and in metabolism studies can generate information on pathway-related correlations. To understand further the behavior of STOCSY for structural assignment, we analyze the statistical distribution of structural and nonstructural correlations from 1050 ¹H NMR spectra of normal rat urine samples. We find that the distributions of structural/nonstructural correlations are significantly different (p < 10⁻¹¹²). From the area under the curve of the receiver operating characteristic (ROC AUC) we show that structural correlations exceed nonstructural correlations with probability AUC = 0.98. Through a bootstrap resampling approach, we demonstrate that sample size has a surprisingly small effect (e.g., AUC = 0.97 for a sample size of 50). We identify specific signatures in the correlation maps resulting from small matrix-derived variations in peak positions but find that their effect on discrimination of structural and nonstructural correlations is negligible for most metabolites. A correlation threshold of r > 0.89 is required to assign two peaks to the same metabolite with high probability (positive predictive value, PPV = 0.9), whereas sensitivity and specificity are equal at 93% for r = 0.22. To assess the wider applicability of our results, we analyze ¹H NMR spectra of urine from rats treated with 115 model toxins or physiological stressors. Across the data sets, we find that the thresholds required to obtain PPV = 0.9 are not significantly different and the degree of overlap between the structural and nonstructural distributions is always small (median AUC = 0.97). The STOCSY method is effective for structural characterization under diverse biological conditions and sample sizes provided the degree of correlation resulting from nonstructural associations (e.g., from nonstationary processes) is small. This study validates the use of the STOCSY approach in the routine assignment of signals in NMR metabolic profiling studies and provides practical benchmarks against which researchers can interpret the results of a STOCSY analysis.