Feature selection and aggregation for antibiotic resistance GWAS in Mycobacterium tuberculosis: a comparative study

Drug resistance (DR) of pathogens remains a global healthcare concern. In contrast to other bacteria, acquiring mutations in the core genome is the main mechanism of drug resistance for Mycobacterium tuberculosis (MTB). For some antibiotics, the resistance of a particular isolate can be reliably predicted by identifying specific mutations, while for other antibiotics the knowledge of resistance mechanisms is limited. Statistical machine learning (ML) methods are used to infer new genes implicated in drug resistance leveraging large collections of isolates with known whole-genome sequences and phenotypic states for different drugs. However, high correlations between the phenotypic states for commonly used drugs complicate the inference of true associations of mutations with drug phenotypes by ML approaches.

Recently, several new methods have been developed to select a small subset of reliable predictors of the dependent variable, which may help reduce the number of spurious associations identified. In this study, we evaluated several such methods, namely, logistic regression with different regularization penalty functions, a recently introduced algorithm for solving the best-subset selection problem (ABESS) and "Hungry, Hungry SNPos" (HHS) a heuristic algorithm specifically developed to identify resistance-associated genetic variants in the presence of resistance co-occurrence. We assessed their ability to select known causal mutations for resistance to a specific drug while avoiding the selection of mutations in genes associated with resistance to other drugs, thus we compared selected ML models for their applicability for MTB genome wide association studies.

In our analysis, ABESS significantly outperformed the other methods, selecting more relevant sets of mutations. Additionally, we demonstrated that aggregating rare mutations within protein-coding genes into markers indicative of changes in PFAM domains improved prediction quality, and these markers were predominantly selected by ABESS, suggesting their high informativeness. However, ABESS yielded lower prediction accuracy compared to logistic regression methods with regularization.