MACI: A machine learning-based approach to identify drug classes of antibiotic resistance genes from metagenomic data

Novel methodologies are now essential for identification of antibiotic resistant pathogens in order to resist them. Here, we are presenting a model, MACI (Machine learning-based Antibiotic resistance gene-specific drug Class Identification) that can take metagenomic fragments as input and predict the drug class of antibiotic resistant genes. We trained the model to learn underlying patterns of genes using the Comprehensive Antibiotic Resistance Database. It comprises of 116 drug classes with a total of 2960 representative sequences. Among these 116 drug classes, we found 22 categories (contributing approximately 85% of the overall sequence-data) surpassed other 94 drug classes based on the number of fragments. The model showed an average precision of 0.83 and a recall of 0.81 for these 22 drug classes. Moreover, the model predicted multidrug resistant classes with higher performance score (precision and recall: 0.9 and 0.88 respectively) compared to single drug resistant categories (0.77 and 0.75). Post to this, we analysed these 22 drug classes to find out class-specific overlapping patterns of nucleotides that led to accurate classification. This way, we found five drug classes viz. "carbapenem;cephalosporin;penam", "cephalosporin", "cephamycin", "cephalosporin;monobactam;penam;penem", and "fluoroquinolone". Additionally, the positions of these significant patterns corroborated with the functional domains of majority of antibiotic resistance genes in that drug class, indicating their biological importance. These class-specific patterns play a pivotal role in rapid identification of some drug classes comprising antibiotic resistance genes. Further analysis showed that bacterial species, containing these five-drug classes, were found to have well-known multidrug resistance property.