Whole-genome sequencing lets you screen a bacterial isolate for antimicrobial-resistance (AMR) determinants in silico — quickly, comprehensively, and in a form that can be shared for surveillance. But moving from “a resistance gene is present in the genome” to “this isolate is resistant to this drug” is a genuine interpretive leap, and getting it right requires understanding both the tools and their limits. This guide covers AMR detection from WGS honestly: how it works, and where genotype and phenotype can diverge.
Key takeaways
- Resistance arises from acquired genes and from chromosomal point mutations — detection must cover both.
- Detection depends on curated databases and on identity/coverage thresholds that shape the results.
- Genotype predicts phenotype with high but imperfect concordance — a detected gene is not proof of resistance.
- Phenotypic susceptibility testing (AST) remains the reference standard for clinical decisions.
Two mechanisms of resistance
Bacteria become resistant in two broad ways, and a complete analysis addresses both. Acquired resistance genes are typically gained horizontally — on plasmids, transposons, or integrons — and encode functions like drug-inactivating enzymes (for example beta-lactamases). Chromosomal point mutations alter existing genes so a drug no longer binds its target, as with mutations in gyrA conferring fluoroquinolone resistance or in rpoB for rifampicin. Additional mechanisms — efflux pumps, reduced permeability, and target modification — complicate the picture further, and not all are captured by gene-presence screening. Recognising this is the first step toward honest interpretation.
Step 1 — Reads or assembly
AMR detection can run on either an assembled genome or the raw reads. Assembly-based detection (screening contigs) is the common route and integrates neatly with the rest of a bacterial WGS workflow; read-based detection maps reads directly to reference gene databases and can be more sensitive for genes that fragment across contig breaks. Either way, the quality of the input matters: a poor assembly can split or drop genes, so resistome screening is only as reliable as the genome it runs on.
Step 2 — Databases and tools
Detection compares your sequence against a curated reference database of known resistance determinants. Three are widely used and actively maintained. NCBI’s AMRFinderPlus, built on the Bacterial Antimicrobial Resistance Reference Gene Catalog, detects both acquired genes and point mutations from protein or nucleotide sequence. The Comprehensive Antibiotic Resistance Database (CARD), queried with its Resistance Gene Identifier (RGI), includes a broad set of determinants including efflux systems. ResFinder, from the Center for Genomic Epidemiology, focuses on acquired genes and pairs with PointFinder for chromosomal mutations. Because these databases have different scopes and curation philosophies, the tool you choose shapes what you find — a point we return to below.
Step 3 — Identity and coverage thresholds
A match is called when a database gene aligns to your sequence above set thresholds for percent identity (how similar the sequences are) and coverage (how much of the reference gene is present). ResFinder, for instance, has commonly been used with a threshold around 98% identity and 60% length coverage. These thresholds are consequential: set them too stringently and you miss divergent variants of a known gene; set them too loosely and you call spurious or partial hits. Reporting the thresholds used is part of making a resistome analysis reproducible, and interpreting borderline hits — a gene present at low coverage, or a distant homolog — requires judgement rather than a blanket rule.
Step 4 — Point mutations
Acquired-gene screening alone misses resistance caused by chromosomal mutations, which are central for several important drug classes. Dedicated methods — PointFinder, and the point-mutation module of AMRFinderPlus — check specific positions in known target genes against catalogues of resistance-associated changes. This matters because, for drugs like fluoroquinolones or rifampicin, the difference between susceptible and resistant can come down to a single amino-acid substitution that a gene-presence screen would never flag.
Step 5 — From genotype to phenotype
This is the crux, and where honesty is essential. Studies comparing WGS-predicted resistance to phenotypic testing report high concordance — often around 98% for well-characterised gene–drug relationships in species like Salmonella, E. coli, and Campylobacter. That is genuinely useful. But concordance is not perfect, and the gaps are instructive. A resistance gene may be present but not expressed (a “silent” gene), so the genotype over-predicts resistance. Resistance may arise from a novel or uncatalogued mechanism the database does not contain, so the genotype under-predicts. And mechanisms such as efflux, permeability changes, and regulatory effects are only partially captured by gene screening. The correct framing is therefore that a detected determinant indicates likely resistance and flags it for confirmation — not that it proves a clinical phenotype. For clinical decisions, phenotypic antimicrobial susceptibility testing (AST), interpreted against CLSI or EUCAST breakpoints, remains the reference standard; genomic prediction complements it rather than replacing it.
Why databases disagree
Run the same isolate through different tools and you will often get different gene counts. Comparative studies show that AMRFinderPlus, CARD, and ResFinder can report different numbers of determinants for the same genomes — CARD, with its broad inclusion of efflux and related systems, typically reports more — and even detection of a specific gene can vary between databases. None of this means one tool is “wrong”; it reflects different scopes, thresholds, and curation. The practical implication is to choose tools deliberately, document versions (databases update frequently), and, for important results, cross-check across databases rather than trusting a single run. This kind of methodological rigour underpins our infectious-disease and surveillance work.
Common pitfalls
The recurring AMR-detection mistakes are interpretive. Equating gene presence with resistance ignores expression, mechanism, and context. Screening only for acquired genes misses point-mutation resistance. Not reporting thresholds or database versions makes results irreproducible. Running on a poor assembly splits or drops genes. And trusting a single database can miss or over-call determinants. Each is avoidable with a documented, versioned pipeline and honest reporting — treating genomic AMR prediction as strong evidence to be interpreted, not as a verdict.
Confirmation and surveillance context
Because genomic prediction is strong evidence rather than proof, important resistance findings are confirmed phenotypically. Antimicrobial susceptibility testing (AST) — commonly broth microdilution — measures the minimum inhibitory concentration (MIC), the lowest drug concentration that inhibits growth, which is then interpreted as susceptible or resistant against EUCAST or CLSI breakpoints. When genotype and phenotype disagree, the discrepancy is informative: it may point to a silent gene, a novel mechanism, or a determinant the database misclassified, and resolving it improves both the isolate’s interpretation and the reference databases themselves.
At population scale, WGS-based AMR screening feeds surveillance. Consistent, documented methods let laboratories compare resistomes across time and geography, track the spread of specific determinants, and support One Health monitoring across human, animal, and environmental isolates. This only works if analyses are reproducible — same databases, versions, and thresholds — which is why version-controlled pipelines and clear reporting matter as much for AMR as the detection step itself. Interpreting these datasets soundly is where careful biostatistics and honest reporting earn their place.
AMR beyond single isolates
The same principles extend from cultured isolates to whole communities. Metagenomic resistome analysis screens shotgun metagenomic data — from gut, wastewater, soil, or clinical samples — for the full complement of resistance genes present, without needing to culture the organisms carrying them. This is powerful for surveillance because it captures resistance in unculturable and rare community members, but it comes with important caveats: metagenomic detection generally cannot tell you which organism carries a gene or whether it sits on a mobile element, and low-abundance determinants may be missed at typical sequencing depths.
This community view is central to One Health surveillance, which tracks resistance across human, animal, and environmental reservoirs on the understanding that they are connected. Whether working from isolates or metagenomes, the discipline is the same: use curated databases, document versions and thresholds, and frame gene detection as evidence of resistance potential in the sample rather than confirmed phenotype. Read-based tools that map directly to reference databases are often preferred for metagenomic data, since assembly of complex communities is itself error-prone. Interpreting these large, noisy datasets soundly is where domain expertise matters most.
Conclusion
AMR detection from WGS is fast, comprehensive, and increasingly central to surveillance — but its value depends entirely on honest interpretation. Cover both acquired genes and point mutations, choose and document your databases and thresholds, and present results as well-supported predictions that phenotypic testing confirms rather than as clinical conclusions. Handled that way, genomic resistome screening is a powerful tool. If you’d like your isolates screened and reported with that rigour, our team supports the full workflow — tell us what you’re working with.