Desenvolvimento de uma metodologiapara previsão de sítios de início detradução

• Ano de defesa: 2007
• Autor(a) principal: Cristiane Neri Nobre
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de Minas Gerais; UFMG
• Palavras-chave em Português: Smote; Sítio de Início de Tradução; de Dados Desbalanceado; Inferência Transdutiva; Support Vector Machine; Con-junto; Bioinformática
• Link de acesso: http://hdl.handle.net/1843/GRFO-7P4LQ9

The correct prediction of the translation start site in mRNA sequences is an im-portanttask in genomic annotation. However, attaining a correct prediction is nottrivial. Frequently the translation starts on the first AUG, but that is not a rule.Thus, this problem can be modeled as a classification problem between positive (co-dingsequences) and negative patterns (non coding sequences). To approach thisproblem the authors of this work propose the following methodology: (1) an alterna-tiveextration of negative patterns; (2) using of shorter sequence window; (3) modi-ficationof the codification for the nucleotides; (4) utilization of Smote - method forclass balance, since the problem is highly unbalanced (1:29 fold in average) for thebases used in this work; (5) use of a transductive approach besides the traditionalinductive inference; and finally, (6) use of the Support Vector Machine (SVM) classi-fier- with simple kernel functions. To test this methodology sequences collected byPetersen and Nielsen and RefSeq (Reference Sequences) sequences from NCBI (Na-tionalCenter for Biotechnology Information) from five organisms were used: Daniorerio, Drosophila melanogaster, Homo sapiens, Mus musculus and Rattus norvegicus,under six distinct inspection levels (reviewed, provisional, predicted, validated, mo-deland inferred). As a result, accuracy, adjusted accuracy, precision, sensitivityand specificity over 95% were attained, in average, by using negative patterns out offrame during training step, 24 nucleotide windows, codification by triples, patternbalancing with Smote, SVM classifier and by considering a scanning model, in which validation is tested up to TIS.