Reconstruction and classification of unknown DNA sequences

• Autor(a) principal: Lourenço, Alexandre Emanuel Monteiro
• Data de Publicação: 2021
• Tipo de documento: Dissertação
• Idioma: eng
• Título da fonte: Repositório Científico de Acesso Aberto de Portugal (Repositórios Cientìficos)
• Texto Completo: http://hdl.handle.net/10773/33658
• topic: Sequence classification; Machine learning; Data compression; Experts ensemble; Metagenomics

The continuous advances in DNA sequencing technologies and techniques in metagenomics require reliable reconstruction and accurate classification methodologies for the diversity increase of the natural repository while contributing to the organisms' description and organization. However, after sequencing and de-novo assembly, one of the highest complex challenges comes from the DNA sequences that do not match or resemble any biological sequence from the literature. Three main reasons contribute to this exception: the organism sequence presents high divergence according to the known organisms from the literature, an irregularity has been created in the reconstruction process, or a new organism has been sequenced. The inability to efficiently classify these unknown sequences increases the sample constitution's uncertainty and becomes a wasted opportunity to discover new species since they are often discarded. In this context, the main objective of this thesis is the development and validation of a tool that provides an efficient computational solution to solve these three challenges based on an ensemble of experts, namely compression-based predictors, the distribution of sequence content, and normalized sequence lengths. The method uses both DNA and amino acid sequences and provides efficient classification beyond standard referential comparisons. Unusually, it classifies DNA sequences without resorting directly to the reference genomes but rather to features that the species biological sequences share. Specifically, it only makes use of features extracted individually from each genome without using sequence comparisons. RFSC was then created as a machine learning classification pipeline that relies on an ensemble of experts to provide efficient classification in metagenomic contexts. This pipeline was tested in synthetic and real data, both achieving precise and accurate results that, at the time of the development of this thesis, have not been reported in the state-of-the-art. Specifically, it has achieved an accuracy of approximately 97% in the domain/type classification.