According to the World Health Organization’s October 2025 assessment, the global pipeline of antibacterial agents under development shrank from 97 to just 90 over two years, with only 15 classified as novel and merely five targeting pathogens on WHO’s “critical” priority list.
The scarcity of new antibiotics makes it key to ensure they reach the right patients and are withheld from those who don’t need them. This requires quickly identifying which drugs an infection resists. According to the WHO, one in six common bacterial infections is now antibiotic-resistant, and drug-resistant infections were linked to an estimated 4.71 million deaths in 2021.
AI Models for Resistance Prediction
AI models that analyze a microbe’s genome to predict effective drugs could reduce the wait time for results from days to hours. However, the data underpinning these models is inadequate. Two major reviews in 2026, one from a EUCAST subcommittee and another from the Align Foundation and Google DeepMind, reached a similar conclusion.
While genome-based prediction is advancing, it relies on high-quality lab testing that much of today’s training data lacks. Public datasets also lack experimentally validated susceptibility results. For instance, in 2020, researchers found over 2 million contaminated entries in GenBank, the NIH’s public sequence database.
Each entry is labeled as a single organism but also contains DNA from other organisms, so an AI model trained on it may learn from sequences unrelated to the pathogen. Additionally, susceptible isolates rarely make it into public databases because a drug that worked seems unremarkable. However, an AI model needs them to learn to distinguish resistant genomes from susceptible ones.
Challenges in Building Trustworthy AI Models
Drug developers relying on these AI models to accelerate discovery should consider the following: the training data must be trustworthy, the AI model must understand the underlying biology, and the AI model must hold up on strains it has never seen.
For instance, a 2024 benchmark study compared four machine-learning (ML) models with ResFinder, a rule-based tool that flags known resistance genes, across 78 species-antibiotic datasets. The study found that when test strains were less closely related to the training strains, ResFinder came out ahead on 44% to 50% of the datasets.
The gap between familiar and unfamiliar strains is critical because the most concerning strains for clinicians are often the least represented in training sets. Without standardized test sets and data splits, accuracy comparisons across teams are unreliable. The Acinetobacter baumannii, a WHO critical-priority pathogen, exemplifies this, as its carbapenem resistance depends on the activation of a mobile genetic element.
As a result, AI models that only check for the presence of a resistance gene may misclassify many susceptible strains as resistant. The same genetic makeup can produce different outcomes, and training data rarely captures the context explaining these variations, creating blind spots for gene-focused AI models.
Addressing these challenges will require a concerted effort from clinical labs, culture collections, drug developers, and AI model builders to contribute well-characterized isolates and adopt common benchmarks. The ATCC Genome Portal, for example, links authenticated strains to reference-grade genomes and includes pathogens on WHO’s 2024 Bacterial Priority Pathogens List.
Building reliable AI resistance prediction models depends on reference data that supports the model throughout its lifecycle, including authenticated strains traceable to verified sources, with genomes and MICs generated from the same isolate using a consistent method.
Culture collections can provide these materials, and the ATCC Genome Portal connects authenticated strains to high-quality genomes.
Protecting the Next Generation of Antibiotics
New antibiotics generate little revenue initially, partly because they are reserved as last-resort treatments. Fast, accurate AI-based resistance prediction can help target them to the right patients.