A machine learning-based strategy to elucidate the identification of antibiotic resistance in bacteria.

Abstract

Microorganisms, crucial for environmental equilibrium, could be destructive, resulting in detrimental pathophysiology to the human host. Moreover, with the emergence of antibiotic resistance (ABR), the microbial communities pose the century's largest public health challenges in terms of effective treatment strategies. Furthermore, given the large diversity and number of known bacterial strains, describing treatment choices for infected patients using experimental methodologies is time-consuming. An alternative technique, gaining popularity as sequencing prices fall and technology advances, is to use bacterial genotype rather than phenotype to determine ABR. Complementing machine learning into clinical practice provides a data-driven platform for categorization and interpretation of bacterial datasets. In the present study, k-mers were generated from nucleotide sequences of pathogenic bacteria resistant to antibiotics. Subsequently, they were clustered into groups of bacteria sharing similar genomic features using the Affinity propagation algorithm with a Silhouette coefficient of 0.82. Thereafter, a prediction model based on Random Forest algorithm was developed to explore the prediction capability of the k-mers. It yielded an overall specificity of 0.99 and a sensitivity of 0.98. Additionally, the genes and ABR drivers related to the k-mers were identified to explore their biological relevance. Furthermore, a multilayer perceptron model with a hamming loss of 0.05 was built to classify the bacterial strains into resistant and non-resistant strains against various antibiotics. Segregating pathogenic bacteria based on genomic similarities could be a valuable approach for assessing the severity of diseases caused by new bacterial strains. Utilization of this strategy could aid in enhancing our understanding of ABR patterns, paving the way for more informed and effective treatment options.