Date of Award

11-30-2022

Document Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Information Science

First Advisor

Intawat Nookaew

Abstract

Advancement in metagenomic sequencing technologies have led to a tremendous improvement in public health through the identification of microbial organisms in their natural environment and the role they play. Developments in second and third generation sequencing technologies have contributed immensely to the increase in prokaryotic and eukaryotic reference genomes. Alignment of metagenomic reads to all these reference genomes in the quest to identify the microbial organism present in the sample could be computationally expensive and time consuming, thus the need for a more quick and accurate taxonomic classifiers that can aid in the fast identification of these organisms. Construction of a more representative reference genome sketch using sketching technique tools requires the selection of an appropriate k-mer length which is often chosen arbitrary. The choice of k-mer length used is important to ensuring optimal resolution of reference genomes in order to infer biologically meaningful phylogenetic relationships. Thus, in this study we developed a novel software KITSUNE for identifying the empirical optimal k-mer length for virus, bacteria and fungi genomes. Using the optimal k-mer length obtained, we created a more representative reference sketch of the reference genomes which we later used in the development of a taxonomic classifier for identifying microbial organisms present in metagenomic samples. We showed that our taxonomic classifier (CAIM) which uses two different sketching tools (Kssd and Mash screen) and the genome coverage information as a means to filter out false positives rather than the relative abundance performs equally well or even better than some existing classifiers when applied on many mock community datasets. Lastly, we also demonstrated that the use of the genome coverage cutoff to filter out false positives improved the prediction performance of our prediction models (Random Forest (RF), Support Vector Machines (SVM) and Least Absolute Shrinkage and Selection operator (LASSO)) as compared to using the relative abundance to filter out false positives in discriminating fecal shotgun metagenomes from healthy controls and colorectal cancer patients.

Download

Share

COinS