From genotype to phenotype: EUCAST on predicting antimicrobial susceptibility from whole-genome sequencing
Table of Contents
Publication details
Article: The role of whole genome sequencing in antimicrobial susceptibility prediction of bacteria: 2025 update from the European Committee on Antimicrobial Susceptibility Testing Subcommittee
Authors: Ørjan Samuelsen, Carla López-Causapé, Frank Møller Aarestrup et al.
Journal: Clinical Microbiology and Infection, 2026, Vol. 32, Issue 7S, pp. S1–S34
DOI: 10.1016/j.cmi.2026.05.012
Article type: EUCAST subcommittee guidance
Download
Clinical Microbiology and Infection has published an updated EUCAST subcommittee opinion on the role of whole-genome sequencing (WGS) in predicting bacterial antimicrobial susceptibility.
The previous EUCAST report, published in 2017, took a cautious position: the available evidence was insufficient to consider WGS a complete alternative to phenotypic antimicrobial susceptibility testing (AST).
The field has advanced substantially since then. Databases of resistance determinants have expanded, long-read sequencing has developed, species-specific bioinformatics tools and machine-learning methods have emerged, and far more paired genomic and phenotypic observations are now available.
Nevertheless, the central conclusion remains measured: for most organism–antimicrobial combinations, genomic prediction cannot yet fully replace phenotypic AST.
The distinction between detecting a genetic resistance marker and predicting a clinically meaningful phenotype is therefore one of the report’s main themes.
Detecting a resistance gene is not enough
At first glance, the task may appear straightforward: if a genome contains a known resistance gene or mutation, the organism might be expected to be resistant to the corresponding drug.
In practice, the genotype–phenotype relationship is much more complex. The presence of a genetic determinant does not always produce clinical resistance, while the absence of a known marker does not guarantee susceptibility.
Phenotype may be influenced by:
- the level of resistance-gene expression;
- gene copy number;
- porin alterations;
- efflux-pump activity;
- regulatory mutations;
- the specific gene or protein variant;
- the organism’s lineage;
- interactions among multiple resistance mechanisms.
Some mechanisms produce only a small increase in minimum inhibitory concentration (MIC), with clinical resistance emerging only when several mechanisms are combined.
Detecting a resistance gene and assigning an S, I, or R susceptibility category are therefore fundamentally different tasks.
Genotype identifies biological change more readily than a clinical category
EUCAST emphasizes the distinction between epidemiological and clinical interpretive criteria.
Genomic analysis may effectively identify isolates that differ from the wild-type population by detecting acquired resistance mechanisms or relevant mutations. This comparison is made against the epidemiological cut-off value (ECOFF), which separates wild-type from non-wild-type organisms.
However, the ECOFF and the clinical breakpoint do not always coincide. An organism may carry a resistance mechanism and be classified as non-wild type while remaining clinically susceptible to a drug.
The path from
genetic marker → resistance mechanism → MIC change → S/I/R category
cannot therefore always be automated. This is particularly important when genomic data are used to support clinical decisions.
How far genomic prediction has progressed
The updated EUCAST report reviews the evidence organism by organism. The following table summarizes the assessment presented in the article.
| Organism | Evidence base for WGS prediction | Main achievements and limitations |
|---|---|---|
| Enterobacterales | High | Species-specific tools and large paired genotype–phenotype datasets are available. Remaining challenges include combined mechanisms, gene copy number, differences between species and lineages, and newer agents. |
| Staphylococcus aureus | High | Genotype–phenotype relationships are well described for several major resistance mechanisms. Machine learning and gene-expression analysis may add further value. |
| Mycobacterium tuberculosis | High | The most mature application. The WHO mutation catalogue is available, and WGS is already used in routine diagnostics in several countries. |
| Pseudomonas aeruginosa | Moderate | Dedicated tools exist, but prediction is complicated by mutational mechanisms, porins, efflux systems, and interactions among mechanisms. |
| Acinetobacter baumannii | Moderate | Machine-learning models are developing rapidly. Standardized approaches and well-characterized isolate collections remain necessary. |
| Neisseria gonorrhoeae | Moderate | Large genomic and phenotypic datasets are available. Predictive accuracy and resistance prediction for newer agents require further work. |
| Streptococcus pneumoniae | Moderate | Promising results have been achieved for predicting beta-lactam MICs, although methodological limitations remain. |
| Enterococcus spp. | Moderate | Genotype–phenotype associations are established for several standard drugs. Resistance to daptomycin, tigecycline, and some other agents remains more difficult to predict. |
| Haemophilus influenzae | Moderate | Prediction of beta-lactam resistance, including PBP3 alterations, is advancing but requires further external validation. |
| Clostridioides difficile | Low | The volume of paired genomic and high-quality phenotypic data remains limited. |
| Bacteroides fragilis | Low | Evidence exists for individual antimicrobials, but the overall evidence base remains limited. |
The labels high, moderate, and low refer to the amount and maturity of the evidence, not to universal predictive accuracy across all drugs. Even within one species, performance may vary substantially by antimicrobial agent and resistance mechanism.
Why Mycobacterium tuberculosis is an exception
EUCAST considers prediction of drug resistance in Mycobacterium tuberculosis to be the most mature application.
One reason is the WHO mutation catalogue, which systematizes associations between genetic changes and phenotypic resistance. In several countries, genomic testing has already reduced the amount of conventional phenotypic testing required for first-line drugs.
Even here, EUCAST stresses the continuing need for high-quality phenotypic data. Such data are essential for identifying new mechanisms, updating mutation catalogues, and assessing resistance to new agents.
What data are needed to link genotype and phenotype
One of the report’s most important conclusions concerns the quality of data used to develop and validate predictive models.
EUCAST recommends linked datasets combining:
genomic data + quantitative phenotype + metadata
For the phenotypic component, storing the original quantitative measurement—especially the MIC—is preferable to retaining only the final S/I/R category.
Clinical breakpoints can change. If a database contains only a categorical interpretation, retrospective reinterpretation becomes difficult. Retaining the original MIC allows results to be interpreted again when criteria are revised.
Information about the isolate and the context in which it was obtained is equally important. Robust genotype–phenotype studies require representative collections containing both wild-type organisms and isolates with diverse resistance mechanisms, including MIC values close to clinical breakpoints.
These data requirements are directly relevant to the development of machine-learning methods. A predictive model can reproduce only the genotype–phenotype patterns represented in its linked training dataset, making phenotypic testing quality, metadata completeness, and isolate representativeness as important as the choice of algorithm. We discuss the available approaches and their limitations in Can machine learning predict antimicrobial resistance?.
Joint genotype–phenotype analysis is becoming a distinct analytical task
Genomic and phenotypic methods have often been treated as competing approaches to resistance detection. The updated EUCAST position supports a more productive model: genotype and phenotype should be treated as complementary sources of information.
Linking the two levels can help answer more complex questions:
- Does a detected genetic mechanism correspond to the observed phenotype?
- Which genetic markers are accompanied by an MIC increase?
- Which markers occur in phenotypically susceptible isolates?
- Which resistant phenotypes remain unexplained by known determinants?
- How does genotype–phenotype concordance vary between species?
- Which combinations of mechanisms produce the greatest changes in susceptibility?
Discordance should not automatically be interpreted as an error in one method. It may reflect incomplete knowledge of resistance mechanisms, differences in gene expression, interactions between mechanisms, or the distinction between biological and clinical resistance.
Genotype and phenotype in ABioGram
This approach is especially relevant to the development of microbiological surveillance systems.
ABioGram enables genetic antimicrobial resistance markers to be analyzed together with phenotypic susceptibility results. Rather than treating these as two independent datasets, the platform can link them as characteristics of the same isolate.
The resulting analytical chain is:
organism → genetic marker → resistance mechanism → antimicrobial agent → phenotypic result
This makes it possible to identify both expected genotype–phenotype combinations and discordant cases, including situations in which:
- a genetic marker is present but phenotypic resistance is absent;
- phenotypic resistance is present but no known genetic mechanism is detected;
- several resistance mechanisms occur simultaneously;
- the same marker is associated with different phenotypes in different isolates.
As linked datasets grow, this type of analysis may provide a stronger basis for investigating the local epidemiology of resistance mechanisms and their actual phenotypic expression.
Implications for practical microbiology
The updated EUCAST report reflects a transition from the relatively simple task of
“finding a resistance gene”
to the more demanding question:
“What phenotypic and clinical significance does this genetic feature have in a particular organism?”
Sequencing alone cannot answer that question. The required infrastructure must connect:
phenotypic testing → genomic analysis → resistance-mechanism interpretation → metadata → accumulation of linked results
Phenotypic methods therefore do not lose their importance as genomics advances. On the contrary, high-quality phenotypic data remain the foundation for validating and improving genomic predictions.
Main conclusion
Eight years after the previous EUCAST report, the potential of genomic antimicrobial susceptibility prediction has expanded substantially. For several organisms, the evidence base is now extensive, dedicated algorithms are available, and large paired genotype–phenotype datasets have been assembled.
WGS is nevertheless not a universal substitute for phenotypic AST across most clinically important organism–drug combinations.
The most promising direction is therefore not to set genotype and phenotype against each other, but to analyze them together. Combining phenotypic results, genetic resistance markers, and contextual data enables a transition from simply identifying resistance mechanisms to understanding how those mechanisms are expressed in real clinical isolates.
Source
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 European Committee on Antimicrobial Susceptibility Testing Subcommittee.
Clinical Microbiology and Infection. 2026;32(7S):S1–S34.
DOI: 10.1016/j.cmi.2026.05.012