| الوصف: |
Antimicrobial resistance is a global threat to human and animal health, and costs thousands of lives every year. Inappropriate usage of antimicrobials is one of the main reasons for the rapid spread of antimicrobial resistance. Therefore, prior to prescribing an antimicrobial, detecting the resistance phenotypes of the pathogen could help to prevent usage of antimicrobials for which the pathogen is already resistant. Antimicrobial resistance phenotypes can be detected using in vitro or in silico methods. The in silico methods, alignment-based or data-driven, offer automated solutions by searching for resistance-conferring patterns in the bacterial genomes, and the main difference between these approaches is whether prior knowledge on resistance determinants is required. Alignment-based methods predict antimicrobial resistance based on known antimicrobial resistance determinants, therefore, these methods only succeed if the targeted organism and antimicrobial combination is well characterized. On the other hand, data-driven approaches explore data to find meaningful antimicrobial resistance patterns by finding correlations between genotypes and phenotypes. In this PhD study, data-driven approaches with machine learning methods were explored to find an alternative for antimicrobial resistance detection which is less dependent on prior knowledge and allows the discovery of novel antimicrobial resistance determinants. This PhD thesis includes four manuscripts on the prediction of antimicrobial resistance and resistance related features using machine learning. In Manuscript-I, a novel artificial neural network was developed for predicting antimicrobial resistance phenotypes for multiple species and antimicrobials in a rapid and efficient way. This model succeeded in predicting antimicrobial resistance phenotypes from genomic variations with at least 90% accuracy and detected novel antimicrobial resistance genes. This study proved that machine learning is able to detect complex resistance mechanisms from whole genome sequences. In Manuscript-II, we focused on predicting antimicrobial resistance phenotypes from partial genome sequences since complete genomes can typically not be reconstructed from metagenomic sequences. Our random forest models reached the accuracy of 70% to predict resistance phenotypes from approximately 1% of chromosomes of multiple species. Moreover, chromosomal regions with strong and weak antimicrobial resistance signals were detected using this model trained with partial sequences. Therefore, this novel study is fundamental for further studies focusing on predicting antimicrobial resistance from metagenomic data. In Manuscript-III, we developed a random forest model, implemented in SourceFinder that could identify whether complete or partial genomes originated from chromosomes, plasmids, or bacteriophages. This classifier identified the highly fragmented assemblies with at least 70% accuracy regardless of the host taxonomy. This method could be used in projects that study horizontal gene transfers in communities by characterization of assemblies. Complementing SourceFinder, in Manuscript-IV we developed another pipeline around a random forest model, called PlasmidHostFinder, for the identification of plasmid host ranges from the plasmid assemblies by avoiding in vitro analyses. This classifier succeeded in identifying the plasmid host ranges with high accuracy despite sequence plasticity. Thus, this study is critical for closing the gaps in our understanding of plasmid disseminations. Enhanced knowledge of antimicrobial resistance is necessary to take effective action against the rapid spread of antimicrobial resistance. Therefore, throughout this PhD, we focused on developing user-friendly solutions with machine learning models that are important for the understanding of antimicrobial resistance mechanisms and dissemination of genetic resistance elements. |
