Genomic resistance predictions need biological context, review finds

Whole-genome sequencing can identify antimicrobial-resistance genes with remarkable speed, but a 2026 mini-review argues that simply recording whether a determinant is present or absent often produces the wrong answer. The central problem is representational: resistance depends on the functional state of genes, their regulation and dosage, their genomic surroundings, the strain carrying them and the distribution of resistant cells within a bacterial population. By reducing these multidimensional systems to binary resistance profiles, genomic pipelines can label susceptible bacteria as resistant and resistant bacteria as susceptible. The authors conclude that more reliable predictions require models that selectively represent the biological context relevant to each resistance mechanism, supported by quantitative phenotypic data and validation beyond the populations used for model development.

The distinction matters because phenotypic antimicrobial susceptibility testing remains the reference standard for measuring bacterial response to antimicrobial agents. Disk-diffusion and minimum inhibitory concentration assays require viable cultures and can take 24 to 72 hours, while their standardized laboratory conditions may not reproduce biofilm growth, host stress, antimicrobial induction or other states encountered outside the dish. Sequencing avoids the need to culture every organism and supports rapid surveillance of acquired resistance genes, resistance-associated mutations and mobile genetic elements. It is therefore well suited to detection and transmission analysis, but sequence detection and resistance prediction are not interchangeable. A resistance gene can be intact yet silent, disrupted, inducible or active only at a particular copy number, while resistance can also arise through changes in permeability or efflux that no simple gene call captures.

Sequence-based prediction is most reliable when a well-characterized determinant has a large and reproducible effect within a validated organism–antimicrobial pairing. The review cites extended-spectrum β-lactamases, carbapenemases and established quinolone-resistance mutations in gyrA and parC as examples for which genomic tools can achieve high genotype–phenotype concordance. Long-read sequencing may improve these predictions by resolving complete plasmids, integrons and chromosomal neighborhoods that remain fragmented in short-read assemblies. The benefit is not universal, however: long reads are especially valuable when plasmid structure, gene order or mobile-element context is consequential, rather than being obligatory for every prediction task. Binary detection becomes less dependable when a phenotype depends on sequence integrity, expression, gene dosage, host background, population heterogeneity or environmental induction.

Several biological states can be hidden by a simple presence–absence matrix. A structurally intact gene may be transcriptionally or translationally attenuated by a weak promoter, repression, stable messenger RNA structures or unfavorable codon usage. Its activity may vary with antimicrobial exposure, growth phase, nutrients, stress responses and regulatory background, so absence of expression in one laboratory assay does not prove permanent inactivity. Other determinants contain frameshifts, premature stop codons, deletions, disruptive insertions or damaging substitutions that can render them nonfunctional while leaving enough sequence similarity for homology-based annotation. Such pseudogenes differ from silent genes because exposure alone cannot restore their function; genetic reversion, recombination, compensatory mutation or acquisition of an intact allele would be required. Pipelines that do not check open-reading-frame integrity can consequently report a disrupted allele as functional resistance.

Resistance magnitude can also depend on interactions between the acquired determinant and the rest of the bacterial genome. In Gram-negative organisms, efflux-pump overexpression following regulatory changes can lower intracellular drug concentrations, whereas loss or modification of porins such as OmpK35, OmpK36 and OprD can restrict influx. The same acquired β-lactamase may therefore be associated with different minimum inhibitory concentrations in different strains. In Klebsiella pneumoniae, loss or modification of OmpK35 and OmpK36 can amplify the phenotypic effect of KPC production, while OprD inactivation is a major contributor to imipenem resistance in Pseudomonas aeruginosa. Gram-positive examples likewise involve background-dependent regulation: NorA, NorB, NorC and MepA contribute to fluoroquinolone resistance only when upregulated, while cell-wall remodeling and mutations affecting daptomycin susceptibility involve several interacting genes. Whole-genome sequencing can identify some coding and regulatory variants, but conventional presence–absence analysis does not directly measure porin or efflux expression.

Expression timing and gene dosage create another layer of uncertainty. Chromosomal ampC in Enterobacter cloacae, Citrobacter freundii, Serratia marcescens and related taxa may remain weakly expressed until β-lactam exposure or become constitutively derepressed through changes involving ampD, ampR or associated regulatory pathways. Erythromycin-resistant but clindamycin-susceptible staphylococci can carry inducible erm-mediated macrolide–lincosamide–streptogramin B resistance that is revealed by a D-test, while attenuator deletions or duplications can create constitutive expression. Promoter variation within Tn4401 can alter blaKPC transcription, and gene or plasmid copy number can increase minimum inhibitory concentrations depending on permeability and genomic background. A binary profile cannot distinguish basal, inducible, derepressed, constitutive and high-copy states, nor can it show why a mecA-positive Staphylococcus aureus isolate may remain susceptible until expression and auxiliary changes alter the phenotype.

Minor resistant subpopulations add a population-level problem. Heteroresistance describes a resistant fraction within an otherwise susceptible isolate and may result from low-frequency variants, plasmid copy-number changes or transient amplification. Conventional susceptibility testing can miss such minorities, while consensus sequencing can mask the same variants or amplification. The discrepancy is partly methodological: sequencing and susceptibility testing may examine different population states, colonies selected for testing may not include the resistant minority, and transient amplification can disappear during culture or passage. Population-analysis profiles can confirm heteroresistance, while deep sequencing, minor-variant analysis and normalized read depth can provide complementary evidence, but their absence in routine whole-genome data does not exclude a resistant fraction below a workflow’s detection limit. A multicenter study of Mycobacterium tuberculosis mixtures illustrates this limitation: an agar-proportion method detected a 1% resistant subpopulation, whereas routine sequencing workflows reported rifampicin-resistance variants at 10% but not 1%. No universal sequencing threshold applies because sensitivity depends on depth, library preparation, variant calling, quality filters and reporting rules.

The proposed alternative is a tiered, mechanism-aware system rather than an indiscriminate demand for every possible omics technology. Routine whole-genome sequencing and validated allele calls remain appropriate for high-effect determinants. Full-length alleles and open-reading-frame assessment can address disrupted genes, while promoter sequences, standardized expression data, read-depth estimates, gene neighborhoods and plasmid or chromosome localization add context where the relevant mechanism requires them. Regularized linear or tree-based models may work for sparse tabular features, whereas graph-based or attention-based approaches are candidates for genomic topology or regulatory sequences, and late-fusion ensembles may combine several data types. Missing information must remain explicitly unknown rather than being coded as absence. More complex models also have to demonstrate added value, calibration, external validity, scalability and interpretable errors rather than merely improving performance within a development dataset.

Models must also be tested for transportability across species, lineages, regions and the human, animal, food and environmental compartments encompassed by the One Health framework. The review reports that ResFinder 4.0 achieved 97% overall concordance in a Danish animal-and-meat Escherichia coli dataset, 98% in a UK animal-surveillance dataset and 95.3% in a blind dataset of human- and animal-origin isolates, although concordance for ertapenem fell to 54.6%. Differences cannot be attributed to ecological compartment alone because testing methods, resistance prevalence, lineages and represented mechanisms also varied. A model trained on 980 pathogenic E. coli isolates from seven livestock and companion-animal hosts achieved a mean F1 score of 98.0%, yet its predictors for the same antimicrobial differed among hosts. These comparisons indicate why random train–test partitions are insufficient: closely related isolates can appear in both sets, producing optimistic estimates. Lineage-aware, geographically separated, temporally separated and cross-compartment evaluations, along with subgroup-specific calibration and uncertainty estimates, provide more informative tests.

Metagenomic resistance-gene detection presents a related but distinct challenge because it can miss the physical and functional links needed to infer phenotype. A gene and bacterial taxon may occur in the same sample without the gene being carried by that organism. Short-read assemblies may separate resistance genes from taxonomic markers, plasmid backbones or neighboring regulatory regions, while abundance data reveal neither expression nor mobility. Long-read, hybrid, targeted-enrichment and proximity-ligation approaches can improve host attribution and structural resolution, but they remain limited by sequencing depth, community complexity, DNA quality and analytical thresholds. Metatranscriptomics may show that a detected gene is expressed, yet transcripts can originate from multiple taxa, vary with environmental conditions and fail to represent minority resistant populations. The review therefore treats metagenomic findings as resistance potential: even confirmed host linkage and expression do not define the susceptibility phenotype of a particular host population. A pragmatic workflow is to combine validated genomic prediction for high-concordance pairings with targeted escalation when phenotypes are unexplained or consequences are high, while retaining phenotypic testing as a complement rather than allowing unresolved genomic uncertainty to be presented as a definitive clinical result.

Subject of Research: Context-aware genomic prediction of antimicrobial resistance across One Health settings

Article Title: From genome to phenotype: advances, challenges and One Health perspectives in the genomic prediction of antimicrobial resistance

Article References: Gomes, E., Figueiredo, R. J., Mesquita, J. R., Almeida, C., Araújo, D., Azevedo, N. F., & Castro, J. (2026). From genome to phenotype: advances, challenges and One Health perspectives in the genomic prediction of antimicrobial resistance. Frontiers in Cellular and Infection Microbiology, 16, Article 1924354. https://doi.org/10.3389/fcimb.2026.1924354

Image Credits: AI Generated

DOI: 10.3389/fcimb.2026.1924354

Keywords: antimicrobial resistance, whole-genome sequencing, genotype-phenotype prediction, One Health, resistome, phenotypic susceptibility testing, machine learning, metagenomics, heteroresistance, long-read sequencing, genome, phenotype