From sequence to signature: Machine learning uncovers multiscale feature landscapes that predict AMR across ESKAPE pathogens.

Since the clinical introduction of antibiotics in the 1940s, antimicrobial resistance (AMR) has become an increasingly dire threat to global public health. Pathogens acquire AMR much faster than we discover new drugs (antibiotics), warranting innovative methods to better understand its molecular underpinnings. Traditional approaches for detecting AMR in novel bacterial strains are time-consuming and labor-intensive. However, advances in sequencing technology offer a plethora of bacterial genome data, and computational approaches like machine learning (ML) provide an optimistic scope for in silico AMR prediction. Here, we introduce a comprehensive multiscale ML approach to predict AMR phenotypes and identify AMR molecular features associated with a single drug or drug family, stratified by time and geographical locations. As a case study, we focus on a subset of the World Health Organization's Bacterial Priority Pathogens, the frequently drug-resistant and nosocomial ESKAPE pathogens: Enterococcus faecium, Staphylococcus aureus, Klebsiella pneumoniae, Acinetobacter baumannii, Pseudomonas aeruginosa, and Enterobacter species. We started with sequenced genomes with lab-derived AMR phenotypes, constructed pangenomes, clustered gene and protein sequences, and extracted protein domains to generate pangenomic features across molecular scales. To uncover the molecular mechanisms behind drug-/drug class-specific resistance, we trained logistic regression ML models on our datasets. These yielded ranked lists of AMR-associated genes, proteins, and domains. In addition to recapitulating known AMR features, our models identified novel candidates for experimental validation. The models were performant across molecular scales, data types, and drugs while achieving a median normalized Matthews correlation coefficient of 0.89. Prediction performance showed resilience even when evaluated on geographical and temporal holdouts. We also evaluated model generalizability and cross-resistance across the drug-/drug class-specific models cross-tested on other available drug-/drug class genomes. Finally, we uncovered multiple drug class resistance features using multiclass and multilabel models. Our holistic approach promises reliable prediction of existing and developing resistance in newly sequenced pathogen genomes, while pinpointing the mechanistic molecular contributors of AMR. All our models and results are available at our interactive web app, https://jravilab.org/amr.