The Application Of Support Vector Machines On Medical And Bioinformatics

Support vector machines, as a new generation of data mining methods, have delivered a state-of-the-art performance in real-world applications. In this work, we describe two successful applications of SVMs.One is the application to breast cancer diagnosis, which has shown good generalization. We take use of non-symmetrical C-SVM to solve the problem of unbalanced training examples, In order to gain a fast searching method for parameters of the model, a margin-based bound on generalization is more effective than traditional k-fold cross-validation. After feature subset selection by a cross-entry filter, we even gained a perfect prediction accuracy.The other one is the application to prediction of pre-MicroRNAs. MicroRNAs (miRNAs) are noncoding RNAs of ～ 22 nucleotides that play versatile regulatory roles in multicelluler organisms. MiRNAs identification by experimental methods are always biased towards abundant miRNAs, while computational approaches provide useful complements and are efficient for identifying those highly constrained tissue-and time-specific expressed miRNAs. In this work, a de novo Support Vector Machine (SVM) classifier is developed for identifying pre-miRNAs in plants. To build a classification model, 12 features of pre-miRNAs representing global and sub-structure features were employed. Trained on 790 plant pre-miRNAs and 7,900 pseudo hairpins, the SVM classifier achieves 96.43% five-fold cross-validation accuracy. Tested on 53 updated arabidopsis miRNAs and 62,883 pseudo pre-miRNAs, it reports accuracy 99.85% with sensitivity 79.25% and specificity 99.87%. While tested on 3447 non-plant pre-miRNAs dataset, it only achieves 46.53% accuracy, which suggests that some features of pre-miRNAs are specific in plants. At last, a feasible genome-wide application of this classification model for miRNAs prediction is suggested, which would be useful for identifying novel miRNAs (especially for those species-specific miRNAs) in plants without relying on phylogenetical conservation.