Drug resistant variants detection using an evolutionary algorithm applied on whole genome sequencing data

The excessive usage of antibiotics is driving the rise of antibiotics resistance. The drug susceptibility testing (DST) based on culture for drug resistance detection is the gold standard for managing the Mycobacterium tuberculosis infection. However, these tests are manual and require time, sophisticated laboratory infrastructure and qualified staff able to use the instrument. To provide an alternative to manual DST, in recent years, many software have been created, including Mykrobe, an open source software, that performs anti microbial resistance predictions (AMR) on microbial deep sequencing data. This prediction is performed in a few minutes and doesn’t require qualified staff to run it. The prediction is possible thanks to a Mykrobe internal database of variants and the relative drug resistances, called panels. The ability of Mykrobe to predict resistance, as well as the panels itself, is being evaluated by the major health organizations in the world, and it was seen that it can produce good results, compared to programs based on manual phenotype testing, to guide a treatment program for tuberculosis. Despite the high quality of its results, Mykrobe isn’t able to predict drug resistances on data that are related to other bacteria except for those already present in the tool, but it allows the user to build its own panels with custom variants. Exploiting this feature, the goal of this thesis is to use an evolutionary algorithm to optimize a Mykrobe panel starting from scratch, with the following steps: I) build a Mykrobe panel from a set of variants; II) perform a prediction, using the panel created; III) based on the output of the AMR, decide which variants to exclude and which to send forward to the next iteration; IV) apply mutation and crossover operator to the selected variants; V) repeat the process over a number of generations. The aim is to build a tool that is able to find the variants present in the default panels. Then the model can be used on samples of little known organisms to detect novel relevant mutations that can be explored to design accurate diagnostic tools.