Amira: gene-space de Bruijn graphs to improve the detection of AMR genes from bacterial long reads

Accurate detection of antimicrobial resistance (AMR) genes is essential for the surveillance, epidemiology and genotypic prediction of AMR. This is typically done by generating an assembly from the sequencing reads of a bacterial isolate and running AMR gene detection tools on the assembly. However, despite advances in long-read sequencing that have greatly improved the quality and completeness of bacterial genome assemblies, assembly tools remain prone to large-scale errors caused by repeats in the genome, leading to inaccurate detection of AMR gene content and consequent impact on resistance prediction. In this work we present Amira, a tool to detect AMR genes directly from unassembled long-read sequencing data. Amira leverages the fact that multiple consecutive genes lie within a single read to construct gene-space de Bruijn graphs where the k-mer alphabet is the set of genes in the pan-genome of the species under study. Through this approach, the reads corresponding to different copies of AMR genes can be effectively separated based on the genomic context of the AMR genes, and used to infer the nucleotide sequence of each copy. Amira achieves significant improvements in genomic copy number recall and nucleotide accuracy, demonstrated through objective simulations and comparison with alternative read and assembly-based methods on samples with manually curated truth assemblies. Applied to a dataset of 32 Escherichia coli samples with diverse AMR gene content, Amira achieves a mean genomic-copy-number recall of 98.4% with precision 97.9% and nucleotide accuracy 99.9%. Finally, we show that Amira consistently detects more true AMR genes across all E. coli, K. pneumoniae and E. faecium nanopore datasets from the ENA (n=8580, 2448 and 415 respectively) than an assembly-based approach.