AMR Prediction from Genotype

Written By Kanna Nandakumar
AMR Prediction from Genotype: Why a Resistance Gene in an Assembly Is Not a Resistant Phenotype
Zetobit · Bioinformatics Insight Series Genomic AST
BIOINFORMATICS INSIGHT SERIES AMR Prediction from Genotype Why a Resistance Gene in an Assembly Is Not a Resistant Phenotype DB MATCH EXPRESSION INTEGRITY DOSAGE HOST BACKGROUND MIC Kanna Nandakumar, PhD ZETOBIT
Bioinformatics Insight Series

AMR Prediction from Genotype

Why a Resistance Gene in an Assembly Is Not a Resistant Phenotype

Kanna Nandakumar, PhD · Zetobit, LLC

An isolate is sequenced, assembled, and run through a resistance gene finder. The report comes back with blaOXA-23. Somewhere between that line and the clinical conversation, the finding becomes “carbapenem-resistant,” and the treatment discussion starts from there.

What the pipeline established is narrower than that. A region of the assembly matched a database entry above a similarity threshold. Whether the gene is transcribed, whether it encodes an intact protein, how many copies the cell carries, and what the rest of the genome does to the drug’s access to its target are all unobserved. Each of those steps can fail on its own, and each failure is silent, because a gene finder has no output category for present and irrelevant.

This is not an argument against genomic susceptibility prediction. It is an argument about what a detection is evidence of. The EUCAST subcommittee that reviewed the field in 2017 concluded that for most bacterial species the evidence did not support using sequence-inferred susceptibility to guide clinical decisions1. The subcommittee’s 2025 update records substantial progress since — growing evidence that sequence can separate wild-type from non-wild-type isolates, particularly for Mycobacterium tuberculosis and where clinical breakpoints align with the epidemiological cut-off — while noting that important challenges remain before the approach becomes a clinical decision-making tool2. Both of those statements are about the same gap. This piece is about what sits inside it.

“Detected” is a threshold decision, not an observation

Before any biology enters the argument, there is a parameter. ResFinder 4.0 runs by default at 80% identity over 60% of the length of the reference gene3. The identity threshold is intuitive enough. The coverage threshold is more interesting, and the reason for it is stated plainly in the benchmarking literature: 60% is set so that genes lying at the edge of a contig, or split across two contigs, are not missed because of imperfect assembly4.

That is a sound engineering decision, and it has a consequence that follows from it directly. The tolerance that rescues a real, intact gene fragmented by an assembly artifact is the same tolerance that admits a gene which is genuinely truncated. A hit covering 65% of the reference is consistent with both, and the presence or absence of the remaining 35% in the assembly does not distinguish an incomplete assembly from an incomplete gene. AMRFinderPlus handles this better than most by explicitly annotating partial matches and internal stop codons — which is the correct response, and which also means the distinction lives in a column that summary tables routinely drop on the way to a report.

Tool choice moves the answer too. In a validation across 6,242 NARMS isolates, ResFinder did not report 1,147 loci that AMRFinder identified5. At catalogue scale the divergence is larger still: a 2026 preprint comparing ten detection pipelines across more than 270 million prokaryotic genes reported up to a 45-fold difference in the number of genes called, with a mean Jaccard index of 16% between pipelines6. That analysis is metagenomic rather than clinical-isolate work, and it has not yet been peer reviewed, so the specific figures should be read as an upper bound on divergence rather than a routine expectation. The direction, though, is not in dispute. A resistome is partly a property of the software that produced it.

Expression: the promoter is usually not inside the gene

The clearest demonstration in clinical microbiology comes from Acinetobacter. The carbapenem-hydrolysing oxacillinase gene blaOXA-23 is frequently silent in its native configuration and becomes consequential when an insertion sequence lands upstream of it. Sequencing the region around the gene showed the transcription start site sitting inside ISAba1, which supplies the promoter that drives expression7. In a survey of clinical isolates, those carrying blaOXA-23 without ISAba1 upstream were carbapenem-susceptible8; in a separate collection where the element was present, expression of the gene was elevated between roughly two- and eightfold9.

Same gene. Same database match. Opposite clinical meaning. The feature that decides between them is a mobile element outside the coding sequence, and whether a pipeline sees it depends entirely on whether the assembly happens to place it on the same contig.

The converse case is equally common and even less visible: resistance driven by regulation alone, with no acquired gene to find. Efflux pump overexpression and derepression of a chromosomal cephalosporinase leave the coding sequences intact and change only how much protein is made. When 414 clinical Pseudomonas aeruginosa isolates were profiled with gene presence, sequence variants, and expression data together, transcriptomic information significantly improved resistance prediction over genomic information alone10. The improvement is the size of the layer the DNA cannot report.

Integrity: a gene can be present, detected, and broken

Oxacillin-susceptible, mecA-positive Staphylococcus aureus is the textbook case of a resistance determinant that a pipeline finds and a laboratory cannot confirm. In one collection these isolates made up 24% of community-acquired MRSA against 2% of hospital-acquired MRSA11. Whole-genome sequencing of seven such isolates traced the susceptibility to mutations in nucleotide-repeat regions within mecA itself12.

It would be tempting to file this as a genotypic false positive and move on. The same two studies make that reading untenable. Most of the isolates were heteroresistant, and exposure to sub-inhibitory concentrations of a β-lactam selected secondary mutations that restored high-level resistance1112. The genotype is not wrong about the capability. The phenotype is not wrong about the current state. They are answers to different questions, and treating the organism as though only the second one existed is the mechanism by which the first one becomes true.

The genotype describes a capability. The phenotype describes a population of cells on the day it was tested. Discordance between them is not automatically an error in either.

Dosage: how many copies, in how many cells

OXA-48 hydrolyses carbapenems weakly. Whether an isolate carrying it crosses a carbapenem breakpoint turns out to depend on quantity. Combining sequencing with mass spectrometry on clinical Klebsiella pneumoniae, one 2025 study found that high carbapenem MICs were driven by increased copy number of the blaOXA-48-carrying plasmid together with loss of the OmpK36 porin, without any contribution from efflux; in mutants selected by meropenem exposure, the same pattern appeared, with the porin lost through IS1 insertion or partial deletion of ompK3613.

Copy number, then, is a scientific parameter rather than an assembly detail. It is also the parameter a consensus assembly is least equipped to report, because tandem copies of a sequence collapse into a single representation by construction.

Heteroresistance takes the same problem one level deeper. Across 41 clinical isolates of Escherichia coli, Salmonella enterica, K. pneumoniae and Acinetobacter baumannii tested against 28 antibiotics, 27.4% of the 766 organism–drug combinations were heteroresistant, and the majority of those cases were unstable, arising from spontaneous tandem amplifications that typically included known resistance genes14. The resistant subpopulation in this phenotype typically sits at a frequency of about 10−7 to 10−415.

Consider what that frequency means for a pipeline. A subpopulation present in one cell in a million contributes nothing to a consensus assembly, contributes essentially nothing to read depth, and is invisible to any variant caller running at a sensible allele-fraction threshold. It is also unstable, which means it can be absent from the culture that was sequenced and present in the population being treated. Both the standard phenotypic test and the standard genomic pipeline fail here, and they fail in the same direction: toward susceptible.

WHAT THE PIPELINE SEES Sequence match in the assembly A contig region passing an identity and coverage threshold — e.g. ≥80% identity over ≥60% of the reference gene. WHAT IT DOES NOT SEE Expression Is the gene transcribed? The promoter often sits outside the coding sequence — an upstream insertion element can be the difference between a silent gene and a resistant isolate. FLANKING CONTEXT Integrity Does it encode a working protein? A partial hit, an internal stop, or a repeat-region mutation yields a detected gene that does not confer the phenotype attributed to it. ALLELE-LEVEL DETAIL Dosage How many copies, in how many cells? Tandem amplifications and plasmid copy number set the enzyme level — and a consensus assembly collapses repeats to a single copy. COPY NUMBER Host background What does the rest of the cell do? Porin loss, efflux capacity and target abundance decide how much drug reaches the enzyme — the same gene gives different MICs in different strains. WHOLE-CELL STATE MIC a measured concentration → Breakpoint or ECOFF a threshold someone chose → S / I / R the line in the report
Figure 1. Schematic. A gene finder reports the top box; a susceptibility category describes the bottom row. The four layers between them fail independently of one another, and three of the four are determined by information a fragmented short-read assembly represents poorly or not at all.

Host background: the gene needs a cell

The OmpK36 result above is not incidental to the carbapenemase — it is half the mechanism. Porins are how carbapenems enter the cell, and an enzyme that hydrolyses slowly can only keep up if influx is also reduced. The dependency runs in the other direction too, and this is the case that most clearly defeats gene-centric prediction. In a six-year collection of carbapenem-resistant Enterobacterales from a single hospital, 44 of 334 isolates (13.2%) produced no carbapenemase at all: the phenotype came from an ESBL or AmpC enzyme combined with porin loss, and none of the carbapenemase genes searched for were present16. A pipeline looking for acquired carbapenemases returns a clean report on those isolates.

Background effects are also quantitative rather than binary. When the same carbapenemase plasmid, pOXA-48, was introduced into a panel of diverse enterobacterial strains, both its fitness cost and the level of resistance it conferred varied substantially across recipients — evidence of genetic interactions between the plasmid’s genes and the host chromosome that are not predictable from either sequence alone17. A resistance gene does not carry an MIC. It carries a contribution to one.

The assembly is the wrong shape for the question

There is a structural irony here worth naming. The features that decide whether a detected gene matters — the insertion sequence upstream of it, its copy number, whether it sits on a plasmid or the chromosome — are repeats or are defined by repeats. Repeats are exactly what fragment a short-read assembly. In work developing complete assemblies for multidrug-resistant Bacteroides fragilis, the authors put it directly: fragmented short-read assemblies often prevent identification of the insertion sequences upstream of AMR genes, which is necessary for predicting resistance from whole-genome sequencing18.

So the coverage threshold that tolerates fragmentation, the gene finder that reports presence, and the assembly that lost the flanking context are all responding to the same underlying limitation, and none of them surfaces it. Long-read or hybrid assembly is not a quality nicety in this setting. It changes which of the four layers in Figure 1 are knowable at all.

What machine learning learns instead

Predicting the phenotype directly from the genome, without a mechanistic model, is an appealing shortcut, and it has a specific failure mode. Analysing more than 24,000 genomes across five priority pathogens and 27 antibiotics, and fitting 6,740 models, one 2025 study showed that when resistance is confounded with phylogeny — which is the normal condition of clinical isolate collections — models perform poorly on held-out clades, and increasing the training sample size fails to rescue them. The models conflate markers of lineage with genuine determinants of resistance19.

The consequence is precise rather than general. A model that has learned “this sequence type is resistant” will report accurately in surveillance, where lineage frequencies are stable, and will fail on the susceptible member of a resistant lineage or the resistant member of a susceptible one. Those are the isolates a clinician needs an answer about. Readers of the earlier piece on population stratification in GWAS will recognise the structure: the confounder and the signal are correlated in the training data, and a model with no mechanism to distinguish them will take whichever is easier to fit.

Where prediction works, and what its accuracy looks like

M. tuberculosis is the strongest case in the field, and it is instructive precisely because performance is well characterised. In a large multi-country study, genotypic prediction of resistance to isoniazid, rifampicin, ethambutol and pyrazinamide reached sensitivities of 97.1%, 97.5%, 94.6% and 91.3%20. In the pooled analysis behind the 2021 WHO mutation catalogue, sensitivity exceeded 80% for rifampicin, isoniazid, ethambutol, the fluoroquinolones and streptomycin, and specificity exceeded 95% for every drug except ethionamide (91.4%), moxifloxacin (91.6%) and ethambutol (93.3%)21. Applied prospectively to an unselected regional cohort of 785 isolates, the same catalogue produced considerably lower sensitivities while retaining very high specificity22.

Genotypic prediction of first-line TB resistance, two settings
Drug Sensitivity, large curated multi-country study20 Sensitivity, unselected regional cohort22 Specificity, curated20 Specificity, regional22
Isoniazid97.1%85.4%99.0%>99.6%
Rifampicin97.5%73.3%98.8%>99.6%
Ethambutol94.6%50.0%93.6%>99.6%
Pyrazinamide91.3%57.1%96.8%>99.6%

The two studies used different catalogues, populations and phenotypic methods, and the regional cohort contained few resistant isolates per drug, so its confidence intervals are wide. The comparison is not a head-to-head benchmark; it illustrates that a prediction rule’s sensitivity is a property of the population it is applied to, while its specificity — governed by how conservatively the catalogue was curated — travels better.

That asymmetry is the important pattern, and it generalises well beyond tuberculosis. A catalogue of known mechanisms is precise about what it names and silent about what it has not seen. Sensitivity therefore tracks how completely the mechanisms circulating in a given population are represented in the database, which is why it falls when the same catalogue moves to a new setting. The operational consequence is worth stating flatly: absence of a known determinant is not evidence of susceptibility. It is absence of evidence, and a pipeline that renders it as “S” has converted one into the other silently.

The comparator is also an estimate

It is worth resisting the assumption that phenotypic testing is ground truth and the genome is the thing being graded. In the regional TB cohort above, four isolates classified as phenotypically susceptible carried rpoB mutations, including Ile491Phe, that are known to produce borderline rifampicin resistance which standard phenotypic tests miss22. There the genotype was right and the phenotype was wrong.

More often, apparent discordance is definitional. EUCAST’s recommendation has been that genotype–phenotype concordance be assessed against epidemiological cut-off values rather than clinical breakpoints, because the question sequence can answer is whether an isolate is wild-type or non-wild-type, not whether an infection is treatable1. Those are different lines drawn on the same MIC distribution. In the ResFinder 4.0 validation, concordance was at or above 95% for 46 of 51 antimicrobial–species combinations in Gram-negative bacteria and 25 of 32 in Gram-positive bacteria; where it fell below, the authors attributed the discrepancies mainly to criteria for interpreting the phenotypic tests and to sequence quality rather than to the tool3. Two measurements disagreeing near a threshold that neither of them defines is not, by itself, a finding about biology.

What to do about it

  1. Report the detection, not the phenotype. “blaOXA-23 detected, 99% identity over 100% of length, contig 14, ISAba1 present 25 bp upstream” is a finding. “Carbapenem-resistant” is an inference, and it should be labelled as one wherever the pipeline is not validated to make it.
  2. Carry identity, coverage and partial-match flags through to the output. A hit at the coverage threshold and a full-length hit are different observations. If internal stops and partial matches are annotated upstream and dropped in the summary, the pipeline has discarded its own best evidence about integrity.
  3. Resolve context when context is the mechanism. Insertion sequences, plasmid location and copy number require long-read or hybrid assembly. If the deliverable includes any claim about expression or genomic context, short reads alone cannot support it.
  4. Estimate copy number from read depth even without long reads. The depth of a resistance gene relative to single-copy chromosomal genes is a cheap, imperfect proxy for the dosage axis, and it is better than the implicit assumption of one copy.
  5. Record the tool, the database version and the thresholds. Two tools on one assembly disagree, and databases change. A resistome without those three values is not reproducible, in the same sense a corrected expression matrix is not reproducible without the method that corrected it.
  6. Treat unassessed mechanism classes as unmeasured. A pipeline with no module for efflux regulation, porin loss or promoter variants should say those were not assessed. Reporting “no resistance determinants found” when a whole mechanism class was never in scope makes an absence of coverage look like a negative result.
  7. State the comparator. ECOFF or clinical breakpoint, and which version. Concordance figures are not comparable across studies that made different choices here.
  8. Validate machine-learning models on phylogeny-aware splits, and report performance on held-out clades rather than random splits. A random split of a structured collection measures the model’s memory of the collection.
  9. Keep phenotypic testing where the decision is therapeutic. For most species this is still what the evidence supports2, and the honest framing of genomic AST for a client is complementary information with a faster turnaround, not a replacement.

The shape of the error

Readers who have followed this series will recognise the pattern by now. A clonal haematopoiesis variant is genuinely somatic, correctly called, and from the wrong tissue. A low-VAF C>T in an FFPE specimen is real chemistry that happened in a cassette rather than in a patient. An ambient transcript is a real molecule counted against the wrong cell. In each case nothing malfunctioned; the measurement was accurate and the inference attached to it was not.

Resistance gene detection belongs to the same family. The sequence is real. The match is real. The database entry describing what that gene does — in some organism, in some genetic context, at some level of expression — is real too. What the assembly does not contain is the cell: the promoter that was or was not there, the copy number that collapsed into a single contig, the porin that was or was not made, and the subpopulation at one in a million that will still be present on day three of therapy.

A resistance gene is a claim about capability. A susceptibility category is a claim about a population of cells under a drug, on a particular day, against a threshold someone chose. Pipelines report the first. Reports are read as the second. Most of the work of doing this well is refusing to let one become the other without saying so.

References

  1. Ellington MJ, Ekelund O, Aarestrup FM, et al. The role of whole genome sequencing in antimicrobial susceptibility testing of bacteria: report from the EUCAST Subcommittee. Clinical Microbiology and Infection 2017;23(1):2–22. pubmed.ncbi.nlm.nih.gov/27890457
  2. Samuelsen Ø, López-Causapé C, Aarestrup FM, et al. The role of whole genome sequencing in antimicrobial susceptibility prediction of bacteria: 2025 update from the EUCAST Subcommittee. Clinical Microbiology and Infection (in press). eucast.org — WGS subcommittee
  3. Bortolaia V, Kaas RS, Ruppé E, et al. ResFinder 4.0 for predictions of phenotypes from genotypes. Journal of Antimicrobial Chemotherapy 2020;75(12):3491–3500. academic.oup.com/jac/article/75/12/3491
  4. Benchmarking of methods for identification of antimicrobial resistance genes in bacterial whole genome data. Journal of Antimicrobial Chemotherapy 2016;71(9):2484–2493. academic.oup.com/jac/article/71/9/2484
  5. Validating the AMRFinder tool and resistance gene database by using antimicrobial resistance genotype–phenotype correlations in a collection of isolates. Antimicrobial Agents and Chemotherapy 2019;63(11):e00483-19. journals.asm.org/doi/10.1128/aac.00483-19
  6. The elusive resistome: a global comparison reveals large discrepancies among detection pipelines. bioRxiv 2026 (preprint; not peer reviewed). biorxiv.org — 2026.05.11.724158
  7. Poirel L, Nordmann P. Genetics and expression of the carbapenem-hydrolyzing oxacillinase gene blaOXA-23 in Acinetobacter baumannii. Antimicrobial Agents and Chemotherapy 2007;51(4):1530–1533. journals.asm.org/doi/10.1128/aac.01132-06
  8. First report of carbapenem-resistant Acinetobacter nosocomialis isolates harboring ISAba1-blaOXA-23 genes in Latin America. Journal of Clinical Microbiology 2013;51(7):2423–2425. pmc.ncbi.nlm.nih.gov/articles/PMC3719618
  9. Over-expression of ISAba1-linked intrinsic and exogenously acquired OXA-type carbapenem-hydrolyzing class D β-lactamase-encoding genes is key mechanism underlying carbapenem resistance in Acinetobacter baumannii. Frontiers in Microbiology 2019;10:2809. frontiersin.org — fmicb.2019.02809
  10. Predicting antimicrobial resistance in Pseudomonas aeruginosa with machine learning-enabled molecular diagnostics. EMBO Molecular Medicine 2020;12(3):e10264. ncbi.nlm.nih.gov/pmc/articles/PMC7059009
  11. Phenotypic and genomic characteristics of oxacillin-susceptible mecA-positive Staphylococcus aureus, rapid selection of high-level resistance to beta-lactams. 2023. pubmed.ncbi.nlm.nih.gov/37515660
  12. Emergence of oxacillin resistance in stealth methicillin-resistant Staphylococcus aureus due to mecA sequence instability. Antimicrobial Agents and Chemotherapy 2019;63(9):e00558-19. journals.asm.org/doi/10.1128/aac.00558-19
  13. Proteogenomic analysis demonstrates increased blaOXA-48 copy numbers and OmpK36 loss as contributors to carbapenem resistance in Klebsiella pneumoniae. Antimicrobial Agents and Chemotherapy 2025. journals.asm.org/doi/10.1128/aac.00107-25
  14. Nicoloff H, Hjort K, Levin BR, Andersson DI. The high prevalence of antibiotic heteroresistance in pathogenic bacteria is mainly caused by gene amplification. Nature Microbiology 2019;4(3):504–514. nature.com/articles/s41564-018-0342-0
  15. Heidarian S, Hjort K, Nicoloff H, Andersson DI. Deletions of recombination genes impair tandem amplification and reshape heteroresistance mechanisms in Escherichia coli. mBio 2026. ncbi.nlm.nih.gov/pmc/articles/PMC12892994
  16. Role of association of OmpK35 and OmpK36 alteration and blaESBL and/or blaAmpC genes in conferring carbapenem resistance among non-carbapenemase-producing Klebsiella pneumoniae. International Journal of Antimicrobial Agents 2018. sciencedirect.com — S0924857918300980
  17. Global epistasis in plasmid-mediated antimicrobial resistance. Molecular Systems Biology 2024;20:311–320. link.springer.com — s44320-024-00012-1
  18. Complete hybrid genome assembly of clinical multidrug-resistant Bacteroides fragilis isolates enables comprehensive identification of antimicrobial-resistance genes and plasmids. Microbial Genomics 2019;5(11). ncbi.nlm.nih.gov/pmc/articles/PMC6927303
  19. Yu MK, Wheeler NE, Barquist L, et al. Biased sampling driven by bacterial population structure confounds machine learning prediction of antimicrobial resistance. PLOS Biology 2025;23(12):e3003539. journals.plos.org — pbio.3003539
  20. The CRyPTIC Consortium and the 100,000 Genomes Project. Prediction of susceptibility to first-line tuberculosis drugs by DNA sequencing. New England Journal of Medicine 2018;379:1403–1415. nejm.org/doi/full/10.1056/NEJMoa1800474
  21. Walker TM, Miotto P, Köser CU, et al. The 2021 WHO catalogue of Mycobacterium tuberculosis complex mutations associated with drug resistance: a genotypic analysis. The Lancet Microbe 2022;3(4):e265–e273. thelancet.com — S2666-5247(21)00301-3
  22. Role of the first WHO mutation catalogue in the diagnosis of antibiotic resistance in Mycobacterium tuberculosis in the Valencia Region, Spain: a retrospective genomic analysis. The Lancet Microbe 2024;5(1):e43–e51. thelancet.com — S2666-5247(23)00252-5
Zetobit, LLC · Bioinformatics consulting · Lexington, KY · zetobit.com
Kanna Nandakumar https://www.zetobit.com
Previous

cfDNA Fragmentomics
Next

Long-Read Methylation Calling: Why the Methylome You Get Depends on Which Model Read the Signal