Whole-genome based bacterial phenotype predictions with machine learning.

Signe Tang Karlsen

Research output: Book/ReportPh.D. thesis

26 Downloads (Pure)

Abstract

Bacteria have been used for thousands of years to ferment food and today they are used in a wide range of applications for instance in production of food, medicin, cemicals, and enzymes. Bacteria have a high diversity and functional variation, and whereas some can be applied in the production of fermented foods, others can be used as sustainable alternatives to chemical pesticides and
fertilizers in agriculture, or as sustainable alternatives to growth-promoting use of antibiotics for livestock. Many bacterial traits relevant for products cannot be determined from taxonomy alone; and to select the best suited bacterial strains, bacteria can be characterized by measuring different phenotypes (observable traits). In this PhD thesis, the application of computational methods for predicting bacterial phenotypes from genome sequences were investigated. The majority of the thesis is focused on the application machine learning models to predict phenotypes and on analysis of the resulting models - with the purpose of learning more about the genomic elements involved in the phenotypes. The thesis includes a review manuscript on the challenges involved in designing, training, and analysing such models. The thesis further includes two manuscripts, in which machine learning was used to model industrially relevant phenotypes of Lactococcus lactis and Bacillus respectively - bacteria with high importance to the production of cheese and products for improvement of plant and animal health. In a fourth manuscript, a hypothesis-driven approach for the identification of pathogenic bacteria is described. For this purpose machine learning was not used due to biased data structures.
Original languageEnglish
PublisherTechnical University of Denmark
Number of pages166
Publication statusPublished - 2021

Fingerprint

Dive into the research topics of 'Whole-genome based bacterial phenotype predictions with machine learning.'. Together they form a unique fingerprint.

Cite this