This thesis presents a novel method, RNAMultifold, for development of a non-coding RNA (ncRNA) classification model based on features derived from folding the consensus sequence of multiple sequence alignments using different folding programs: RNAalifold, CentroidFold, and RSpredict. The method ranks these folding features according to a Class Separation Measure (CSM) that quantifies the ability of the features to differentiate between samples from positive and negative test sets. The set of top-ranked features is then used to construct classification models: Naive Bayes, Fisher Linear Discriminant, and Support Vector Machine (SVM). These models are compared to the performance of the same models with a baseline feature set and with an existing classification tool, RNAz.
The Support Vector Machine classification model with a radial basis function kernel, using the top 11 ranked features, is shown to be more sensitive than other models, including another ncRNA prediction program, RNAz, across all specificity values for the RNA families under study. In addition, the target feature set outperforms the baseline feature set of z score and structure conservation index across all classification methods, with the exception of Fisher Linear Discriminant. The RNAMultifold method is then used to search the genome of a Trypanosome species (Trypanosoma brucei) for novel ncRNAs. The results of this search are compared with known ncRNAs and with results from RNAz.