DRAMMA: a Multifaceted Machine Learning Approach for Novel Antimicrobial Resistance Gene Detection in Metagenomic Data

Background: Antibiotics are essential for medical procedures, food security, and public health. However, ill-advised usage leads to increased pathogen resistance to antimicrobial substances, posing a threat of fatal infections and limiting the benefits of antibiotics. Therefore, early detection of antimicrobial resistance genes (ARGs), especially in pathogens, is crucial for human health. Most computational methods for ARG detection rely on homology to a predefined gene database and therefore are limited in their ability to discover novel genes.

Results: We introduce DRAMMA, a machine learning method for predicting new ARGs with no sequence similarity to known ARGs or any annotated gene. DRAMMA utilizes various features, including protein properties, genomic context, and evolutionary patterns. The model demonstrated robust predictive performance both in cross-validation and an external validation set annotated by an empirical ARG database. Analyses of the high-ranking model-generated candidates revealed a significant enrichment of candidates within the Bacteroidetes/Chlorobi and Betaproteobacteria taxonomic groups.

Conclusions: DRAMMA enables rapid ARG identification for global-scale genomic and metagenomic samples, thus holding promise for the discovery of novel ARGs that lack sequence similarity to any known resistance genes. Further, our model has the potential to facilitate early detection of specific ARGs, potentially influencing the selection of antibiotics administered to patients. Video Abstract.