Can genomic AI reduce missed antibiotic resistance?

The direct answer: fewer dangerous misses in retrospective benchmarks, not yet better care

PredKlebAMR made fewer false-susceptible predictions than AMRFinderPlus and ResFinder in three external genomic benchmarks, including cohorts from Africa and Thailand. That is an important technical result because a false-susceptible call can make a resistant Klebsiella pneumoniae isolate look treatable. But the study did not place the model in a live laboratory, compare prescribing decisions or follow patients. It therefore supports further prospective evaluation, not a claim that the system already improves treatment.

The model was designed to prioritise sensitivity: finding as many resistant isolates as possible. On the authors' held-out global test set it reached 91.92% accuracy, 94.67% sensitivity, 89.47% specificity and a Cohen's kappa of 0.839. Those figures describe agreement with the study's resistance labels under its retrospective protocol. They do not measure the probability that a particular medicine will work for a particular patient, and they should not replace phenotypic antimicrobial-susceptibility testing where that testing is clinically required.[1]

What the model learned from 5,942 genomes

The researchers retrieved 6,000 Klebsiella pneumoniae genomes from the Bacterial and Viral Bioinformatics Resource Center. After preprocessing, 5,942 genomes were used for model development. Rather than feeding raw sequence directly into a black box, the team used 38 predictors generated through Kleborate, a genomic surveillance tool. The features included known resistance determinants and bacterial lineage information. A random-forest classifier was trained with a stratified 80:20 split, while five-fold cross-validation was used to select hyperparameters.

Reported influential predictors included fluoroquinolone-resistance mutations, high-risk sequence types, the KPC-2 and KPC-3 carbapenemase genes, and changes in the OmpK35 and OmpK36 porins. That makes the model more biologically legible than a purely opaque sequence classifier: its decisions draw on features already associated with resistance. It also creates a boundary. A model built from known, engineered signals can underperform when new mechanisms, unfamiliar lineages or combinations absent from its training data appear.

The paper concerns resistance profiling for K. pneumoniae, a pathogen associated with hospital and community infections and with globally important antimicrobial resistance. It is not a universal bacterial-resistance model. Performance for this species, these data sources and the studied drugs cannot be transferred automatically to other organisms, laboratories or resistance phenotypes.[1]

How it compared with established genomic tools

The authors compared PredKlebAMR with AMRFinderPlus version 4.2.7 and ResFinder version 4.1 on independent cohorts described as global, African and Thai. In the global cohort, PredKlebAMR's sensitivity was 95.45%, compared with 90.91% for AMRFinderPlus and 86.36% for ResFinder. The model produced one very major error, versus two and three respectively. In the African cohort, sensitivity was 95.56%, compared with 91.11% and 82.2%, with two very major errors rather than four or eight.

The African sensitivity comparison was the one reported as statistically significant, using McNemar's test with p=0.041. In the Thai cohort, sensitivities were 95.62% for PredKlebAMR, 92.5% for AMRFinderPlus and 63.12% for ResFinder; the reported very major error counts were seven, 12 and 59. A very major error here means a resistant isolate was predicted to be susceptible. It is a high-consequence analytical error category, but it is not the same as a documented adverse event in a patient.

The accessible source record does not provide the exact number of isolates or drug-isolate comparisons behind each external percentage. That missing denominator limits interpretation: one percentage point can represent very different evidence in a small cohort and a large one. It also prevents a reader from independently reconstructing the uncertainty around every comparison from the public summary. The full evaluation should be read with cohort-specific sensitivity, specificity, confidence intervals and denominators side by side, not sensitivity alone.[1]

Why the African and Thai tests matter—and what they cannot prove

External evaluation matters because bacterial lineages, resistance mechanisms, sequencing practices and clinical breakpoints vary across place and time. A model can look strong when tested on data resembling its training set and then fail where the mix of strains changes. Testing separate global, African and Thai cohorts is therefore more informative than reporting only an internal split, and the African collaboration is particularly valuable in a literature that often under-represents genomic data from the continent.

Geographic labels are not substitutes for representative sampling. An 'African cohort' does not capture every country, laboratory, patient group or circulating lineage across Africa, just as one Thai cohort cannot represent South-East Asia. Public genomic repositories can over-represent well-resourced laboratories, outbreak investigations and unusual resistant isolates. They may also combine data produced under different sequencing, curation and susceptibility-testing protocols. Those composition effects can make performance look better or worse than it would in routine local practice.

The relevant transfer test is prospective and local: freeze the model, run it on consecutive eligible samples in several independent laboratories, and compare its calls with a predefined phenotypic reference under current breakpoints. Results should be reported by drug, lineage, specimen, hospital and demographic group where appropriate. New or uncertain mechanisms need an abstention or escalation path rather than a confident answer.[1]

What the result could mean for patients and laboratory teams

If prospective studies reproduce the sensitivity advantage, genomic prediction could help laboratories flag likely resistance earlier in investigations where whole-genome sequencing is already available. Fewer false-susceptible calls could reduce the chance that a stewardship team is falsely reassured by genomic evidence. The model's dashboard may also help researchers compare resistance profiles and prioritise cases for review. Those are plausible workflow benefits, not outcomes measured by this paper.

Real clinical value depends on the full pathway. A laboratory needs a suitable isolate or sample, sequencing capacity, quality control, a reproducible bioinformatics pipeline, current interpretive rules and a secure route to the clinician. Turnaround time and cost must be compared with existing culture and susceptibility testing. A result that arrives after treatment is chosen, or that cannot be explained and confirmed, may add little. A higher-sensitivity system can also generate more false-resistance calls if specificity falls, potentially steering clinicians away from useful narrower-spectrum antibiotics.

For patients, the evidence does not support changing medication on the basis of this model alone. Antibiotic decisions depend on infection site, severity, dose, allergies, kidney function, local guidance and laboratory confirmation, not just the presence of genomic features. The safe near-term role is decision support inside a governed microbiology and antimicrobial-stewardship process, with qualified staff accountable for interpretation and a route to resolve disagreement between genomic and phenotypic results.[1]

Evidence quality, interests and the next decisive study

Scientific Reports published the work as a peer-reviewed open-access article on 9 October 2026. The publisher describes the available item as accepted research that may still receive editorial changes before the final version of record. The authors declared no competing interests and said that the research received no specific funding. They acknowledged National Institutes of Health Office of Data Science Strategy support for an April 2026 codeathon while stating that it did not provide financial support for the article.

This is one research group's retrospective study. Independent teams have not yet shown that a locked version reproduces the results prospectively or that using it changes care. The next decisive evaluation should pre-register primary endpoints, publish exact cohort denominators and confidence intervals, and compare the model with current laboratory practice as well as competing software. It should measure false susceptibility, false resistance, abstentions, turnaround, staff workload, antibiotic changes and disagreement resolution.

A clinically persuasive trial would then ask whether adding the genomic model to usual care shortens time to effective treatment without increasing unnecessary broad-spectrum use. Patient outcomes could include treatment escalation, intensive-care admission, length of stay and mortality, with appropriate adjustment and safety oversight. Until that evidence exists, PredKlebAMR is a promising sensitivity-focused research tool whose strongest contribution is a better-tested hypothesis: genomic machine learning may reduce missed resistance across diverse datasets, but whether that benefit survives the bedside pathway remains unproven.[1]