Two Issues Of NMR Data Analysis

Currently, nuclear magnetic resonance (NMR) becomes one of the most widely used analyt-ical techniques with the large scale applications and developments. How to analyze NMR data is always the focus of attention for scientists to obtain comprehensive information from research objects. While because of their mutual promotion between the development of NMR techniques and data science, we can process the data more efficiently and specifically. Therefore, this article discusses two issues of NMR data analysis.The first issue is about using support vector machine (SVM) to detect the abnormal metabo-lites. At first, including the cross-validation, it describes the partial least squares (PLS) method and the orthogonal projections to latent structures (OPLS) (when both of PLS and OPLS are used together, they could be co-written as O-PLS). Then as the rising method without widely using in metabonomics, I present the construction of L1-norm SVM method. The comparison is applying the O-PLS and L1-norm SVM method in metabolic data which are got from mice infected with Schistosoma. The results show that L1-norm SVM not only can detect abnormal metabolites as the same with O-PLS method, but also can result in the detection of possible minor abnormal metabolic products more clearly. L1-norm SVM is also able to remove the influence brought by redundant information. Including the clustering and feature selection advantages of L1-norm SVM, its application in the field of metabonomics has great potential and prospects.The second issue is about the improvement of multi-exponential inversion method. First-ly it introduces the concepts of multi-exponential model and its inversion. Then I describe the non-negative least squares (NNLS) method so that we can make multi-exponential inversion sim-ulations. From the result analysis, we can see original NNLS method needs a low noise condition (SNR^ 50). And its simulation property shows us this one-time calculation is easy to make bias results. By improving the original NNLS with the greedy algorithm and NS (Non-stationary mea-sure) to get better inversion results. New NNLS-NS can work well with greater noise (SNR≥ 30) and output more accurate T2 distribution detection. It has been more effective than the original NNLS and shows new thinking for multi-exponential inversion.