What is a successful antibiotic resistance gene? A conceptual model and machine learning predictions

Sammanfattning

Antibiotic resistance is a global public health threat and it causes bacterial infectionsto become more difficult to treat. The spread of antibiotic resistance genes (ARGs)is predominantly driven by horizontal gene transfer (HGT) that enables bacteria toshare genetic information directly between cells. The ability of an ARG to spread isinfluenced by a range of factors, and has become a popular field of research, aimingto find characteristics that enable rapid antibiotic resistance dissemination. Thisfacilitates the identification of ARGs that possess the ability to disseminate rapidly,and for proactive measures against the dissemination to be implemented.Bioinformatics tools were used to study the prevalence of 4775 known ARGs in 867318 bacterial genomes. A conceptual model describing the success of an ARG wasdeveloped containing four different measures of dissemination, over taxonomic barriers,in different GC-environments, geographical dissemination, and disseminationto pathogenic bacteria. By using a top-down approach studying the success of agene, the thesis complements research studying factors that characterizes successfuland rapid HGT. The conceptual model resulted in a success-score for each ARGthat reflected the overall performance in the four components. Among the ARGsfound to be highly successful the most common class was multidrug resistance, followedby aminoglycoside, β-lactam, and MLS antibiotic resistance. Furthermore,the success-score together with information about the genes, were used to investigatethe possibility to predict the success of an ARG with the use of machine learningin a binary classification Random forest algorithm. The model was built to evaluatethe predictive performance using decreasing amounts of observations of each gene.As expected, the predictive performance of the model improved as the number ofobservation increased. Based on only one observation, it was possible to predict theclass of each gene with an average sensitivity of ~70% at 90% specificity, and with250 observations a sensitivity of 98% could be attained. Sequence related featuressuch as gene length and codon usage were important when only a few observationsof a gene were used, but as the number of observations grew, non-sequence relatedfeatures such as number of countries and pathogens a gene was found in, becamemore relevant. A meta-analysis also aims to explore the managerial and policy implicationsof antibiotics resistance, and findings include that policies facilitating formachine learning are important to implement. This study can be used as a startingpoint in the modelling of antibiotic resistance gene success, aiming to help identifyemerging ARGs that have the possibility to become future threats.