The challenge that Amira addresses begins with how resistance genes are organised in bacterial genomes. Genes conferring resistance to antibiotics are frequently multicopy, meaning that a single bacterial genome may carry two, three or many more copies of the same gene, and these copies are often enriched at repeat loci, regions where long stretches of near-identical sequence recur. These repetitive neighbourhoods are precisely where genome assemblers tend to fail, collapsing distinct copies into a single contig or producing fragmented, unreliable reconstructions. Because the exact sequence and copy number of a resistance gene can determine how resistant a pathogen actually is, errors at these loci translate directly into errors in predicting a bacterium’s drug susceptibility.

Existing computational strategies split into two broad camps, and both struggle with the multicopy problem. The first relies on obtaining a correct genome assembly before searching it for resistance genes, which as noted is unreliable when the genes of interest live in repeats. The second avoids assembly altogether and instead works with k-mers, short sequences of a fixed length extracted directly from the reads. K-mer methods can flag the presence of a resistance gene, but they have a fundamental blind spot: when a gene exists in several slightly different copies, k-mers shared between those copies cannot be attributed to any particular one, so the tools cannot reliably distinguish the different copies from one another or determine how many copies are present.

Amira takes a different path, building what the authors call gene-space de Bruijn graphs. A de Bruijn graph is a classic data structure in bioinformatics in which sequences are broken into short overlapping words and connected according to how those words overlap, producing a network that encodes sequence similarity and variation. Rather than building such a graph over an entire genome, Amira restricts its graphs to the space of known antimicrobial resistance genes, effectively zooming in on the regions that matter for surveillance. Reads from long-read sequencing technologies, which produce fragments long enough to span repetitive regions, are then threaded through these gene-space graphs, allowing the algorithm to assign each read to a specific copy of a resistance gene.

This read-assignment step is the heart of the method. Once reads have been associated with distinct gene copies, the per-copy depth of coverage becomes visible, which enables accurate copy-number estimation, and the reads assigned to each copy can be assembled to recover the precise nucleotide sequence of that copy. In effect, Amira converts a global assembly problem, which is hardest exactly where resistance genes live, into a set of much smaller, localised assembly problems that are tractable because the graph structure has already disentangled the copies. The approach preserves the ability of long reads to resolve repeats while sidestepping the weaknesses of whole-genome assembly and the copy-blindness of k-mer counting.

The performance of the method was evaluated across a range of settings. In simulations, in curated samples where the true composition of the genome was known, and in public datasets, the authors report that Amira improved both copy-number recall and nucleotide accuracy compared with existing methods. Copy-number recall refers to the tool’s ability to correctly identify how many copies of a gene are present, while nucleotide accuracy measures how faithfully the reconstructed sequence matches the true gene sequence, a critical property because small sequence differences within a resistance gene family can correspond to meaningful differences in resistance phenotype.

The evaluation drew on a substantial collection of bacterial isolates. Supplementary materials accompanying the paper describe work with reference assemblies spanning dozens of bacterial species, an Escherichia coli pan-genome reference used both to construct the method’s gene space and to evaluate it, and a truth evaluation involving 32 Escherichia coli isolates with paired Nanopore and Illumina sequencing data. The team also worked with Klebsiella pneumoniae isolates, another major pathogen in which resistance genes are often found in complex repeat structures, and examined plasmid structural variants, reflecting the fact that resistance genes frequently travel on mobile plasmids that shuffle between bacterial strains.

The practical implications reach into public health surveillance. Programmes such as PulseNet, whose Asia-Pacific team at the Centre for Pathogen Genomics contributed data to the study, depend on rapid, accurate genomic characterisation of pathogens circulating in food supplies and hospitals. A tool that can reliably genotype resistance genes directly from long reads, without waiting for or trusting a full assembly, could shorten the path from sequencer to surveillance report and reduce the risk of silent errors in multicopy genes. Because long-read sequencers are increasingly portable and can be deployed in field settings, methods that extract maximum information directly from raw reads are particularly valuable in outbreak and low-resource contexts.

The software’s computational footprint was benchmarked against competing tools across the Escherichia coli evaluation samples, information provided in the paper’s supplementary tables, giving practitioners a sense of the resources required to run the method at scale. The work was led by Daniel Anderson, supported by the European Molecular Biology Laboratory international PhD programme, with senior authorship from Zamin Iqbal, a long-standing figure in pathogen genomics and genetic epidemiology, and contributions from Trieu Le, Leandro Lima, Louise Judd and Ryan Wick, whose groups at Melbourne have deep expertise in bacterial genomics and long-read sequencing.

For a field in which the difference between one copy and three copies of a resistance gene can alter a clinical prediction, the arrival of a method designed explicitly around the structure of resistance loci marks a meaningful step. By combining gene-space de Bruijn graphs with direct read assignment from long reads, Amira offers surveillance teams and researchers a way to see resistance genes as they actually are: duplicated, repetitive and embedded in the hardest parts of the genome. The article is open access, published on 17 October 2026, and its method and benchmarks are available to the community as genome surveillance worldwide confronts the growing burden of antimicrobial resistance.

Subject of Research: Computational detection of single- and multi-copy antimicrobial resistance genes directly from long-read sequencing data

Article Title: Amira: detection of single and multi-copy antimicrobial resistance genes directly from long reads using gene-space de Bruijn graphs

Article References: Anderson, D., Lima, L., Le, T., Judd, L., Wick, R., & Iqbal, Z. (2026). Amira: detection of single and multi-copy antimicrobial resistance genes directly from long reads using gene-space de Bruijn graphs. Genome Biology. https://doi.org/10.1186/s13059-026-04301-w

Image Credits: AI Generated

DOI: 10.1186/s13059-026-04301-w

Keywords: antimicrobial resistance, long-read sequencing, de Bruijn graphs, genome assembly, bacteria, genomics, pathogen surveillance, copy-number detection, bioinformatics, Escherichia coli, Klebsiella pneumoniae, genotyping