Applications Of Support Vector Machine In Medicine And Environmental Chemistry

Support Vector Machine (SVM), developed by Vapnik, as a novel type of machine learning method, has become the hotspot of machine learning because of its rigorous theory background and remarkable generalization performance. The recent researches on SVM mainly concentrated on the property of SVM and diversity and depth of its applications. So far, SVM has been applied to pattern recognition, regression analysis, function estimation and so on. In this dissertation, the SVM was introduced into the medicine and environmental chemistry, which expanded the application of SVM and strengthened the intercrossing of subjects.Chapter 1 of the dissertation included a brief description of the quantitative structure-property relationship (QSPR) principle, realization process and research status. In this section, we also described the support vector machine and a review of its applications of SVM in biology, chemistry and environment area. These theory and study status provided a theoretical basis for us to launch our work.In chapter 2, we studied the quantitative relationship between the molecular structures and the gas chromatographic retention times of 44 alkylphenols. The descriptors were calculated by CODESSA. The heuristic method (HM) was used to pre-select the descriptors and build the linear model (LM). The support vector machine (SVM) was applied to develop the non-linear model. The correlation coefficient (R~2) for the LM and SVM was 0.98 and 0.92, and the corresponding RMS was 0.99 and 2.77, respectively. By comparing the stability and prediction ability of two models, it was found that the linear model was a better method for describing the quantitative relationship between the retention times of alkylphenols and the molecular structures. The results obtained suggested that the linear model could be used for the chromatographic analysis of alkylphenols of known molecular structural parameters.In chapter 3, we built the quantitative structure-property relationships by HM and SVM between the molecular structures and the brain-blood barrier (BBB) permeability of 55 compounds and further discussed the structural factors that influenced the BBB permeability of compounds. The predicted results indicated that the performance of the non-linear SVM model (R~2=0.89, MSE=0.06) was better than that of the linear model (R~2=0.82, MSE=0.11). Descriptor HASA2, N_O, FPSA3 and Enn(CH) were the important structural factors that influenced the logBB of compounds. The SVM model was a simple and efficient tool to predict the logBB values of the candidate molecules in drug design.In chapter 4, the linear discriminant analysis (LDA) and multi-class support vector machine (MSVM) was applied to develop the linear and non-linear three-class classification model for diagnosing the thyroid gland diseases, respectively. The classification results showed that MSVM had the better prediction ability. SVM represented an effective and reliable method for diagnosing the thyroid gland diseases and could been used to diagnose other diseases.