DocumentRX-9040
Issued
Read5 min

Antibacterial pipeline shrinks from 97 to 90, and AI resistance tools inherit bad data

WHO reports antibacterial candidates fell from 97 to 90 in two years, with only five hitting critical pathogens. Reviews from EUCAST and Google DeepMind-backed Align Foundation say the data under AI resistance prediction is not ready.

TAG T-2239 · 964 words on the filing

New antibiotics need AI that gets resistance right
New antibiotics need AI that gets resistance rightbiancamentil / pixabay

Summary of findings

  • WHO reported in October 2025 that clinical-stage antibacterial agents fell from 97 to 90 in two years; only 15 are innovative and only five cover at least one critical-priority pathogen.
  • WHO estimates one in six common bacterial infections is resistant, and drug-resistant infections were associated with an estimated 4.71 million deaths in 2021.
  • A 2024 benchmark study found ResFinder beat four ML models on 44–50% of datasets when test strains were less related to training strains, versus 25% under random splits.
  • Researchers identified more than 2 million contaminated entries in GenBank in 2020.
  • ATCC authors call for authenticated strains with paired genomes and MICs generated under ISO 20776-1 reference methods and shared community benchmarks.

The antibacterial clinical pipeline contracted from 97 candidates to 90 in two years, the World Health Organization reported in October 2025. Only 15 of those agents qualify as innovative, and only five act against at least one pathogen on WHO's critical-priority list. The count landed as a sobering backdrop to this year's World AMR Awareness Week, and it sharpens a question every developer now faces: can the AI tools meant to shepherd scarce antibiotics to the right patients actually be trusted?

The scale of the problem is measurable. WHO estimates one in six common bacterial infections is now resistant to antibiotics, and drug-resistant infections were associated with an estimated 4.71 million deaths in 2021. AI models that read a microbe's genome and predict which drugs will work could cut the wait for a susceptibility answer from days to hours. The data underneath those models, however, is not ready. Two major 2026 reviews — one from a EUCAST subcommittee, one from the Align Foundation and Google DeepMind — reach the same conclusion: genome-based prediction is advancing, but it depends on reference-quality laboratory testing that much of today's training data lacks, and public datasets run short of experimentally validated susceptibility results.

Can the training data be trusted?

Every resistance model learns from antimicrobial susceptibility testing (AST), which measures a minimum inhibitory concentration (MIC) — the smallest dose that stops the microbe growing in the lab. Labs then compare that number to a clinical breakpoint to assign susceptible or resistant. Breakpoints move. In 2010, the Clinical and Laboratory Standards Institute (CLSI) lowered its carbapenem breakpoints for Enterobacteriaceae, so an isolate labeled susceptible before then could be classed as resistant today. Strip that context from a label and the training data quietly corrupts. The raw MIC survives a breakpoint revision; the category does not. MIC values therefore make better training data than susceptible/resistant labels.

The genomes carry their own defects. In 2020, researchers identified more than 2 million contaminated entries in GenBank, the NIH's public sequence database — each labeled as a single organism while containing DNA from others. A model trained on those entries can learn from sequences unrelated to the pathogen. Missing negatives compound the problem: susceptible isolates rarely enter public databases because a drug that worked seems unremarkable, yet models need them to distinguish resistant genomes from susceptible ones.

Does the model understand the biology?

Resistance rarely maps onto a list of genes. A gene can be present but silent; some resistance depends on regulation or cell-to-cell variation that a presence-or-absence check cannot detect. Consider Acinetobacter baumannii, a WHO critical-priority pathogen. It carries an OXA-51-like carbapenemase gene in its own chromosome, but that gene typically drives carbapenem resistance only when the mobile element ISAba1 inserts just upstream and activates it. A model that only checks for the gene will call many susceptible strains resistant.

Heteroresistance creates a parallel trap: a small resistant subpopulation hides within a mostly susceptible culture, standard testing misses it, and the strain is reported as susceptible — after which treatment selects for the resistant cells. When an isolate's MIC sits at the breakpoint, ordinary variation between runs can tip it from susceptible to resistant. These gray-zone cases, where treatment decisions are hardest, are scarce in public data.

Does the model hold up on unseen strains?

A 2024 benchmark study compared four machine-learning models with ResFinder, a rule-based tool that flags known resistance genes, across 78 species-antibiotic datasets. The split decided the outcome. Under random splits, where test strains closely resembled training strains, ResFinder beat the ML models on only 25% of datasets. When test strains were less closely related, ResFinder led on 44% to 50%. Much of the ML advantage came from strains similar to those already seen — precisely the strains clinicians worry about least. Without a common test set and fixed data splits, accuracy figures from different teams cannot be compared.

What will trustworthy prediction require?

ATCC's Jonathan L. Jacobs, senior director of bioinformatics, and Shahin Ali, senior scientist, set out a reference-data agenda in three stages:

  • Build and train. Use authenticated strains traceable to a verified source, with genome and MIC generated from the same physical isolate under one reference method such as ISO 20776-1 broth microdilution.
  • Benchmark and validate. Establish a community-governed test set containing verified susceptible strains, borderline strains, and genotype/phenotype discordant cases, with fixed published data splits; let labs append protein mass spectra or gene-expression data to anchored isolates.
  • Monitor and update. Apply version control and documented benchmark updates, drawing on both recent clinical isolates and historic strains. In ATCC's comparison of historic and recent A. baumannii isolates, the team observed strains progressively acquiring resistance genes over time — temporal collections that teach algorithms how resistance evolves across lineages.

Culture collections can supply the materials; the ATCC Genome Portal already links authenticated strains to reference-grade genomes and includes pathogens on WHO's 2024 Bacterial Priority Pathogens List.

What does this change for developers?

New antibiotics earn little at launch, partly because they are rightly held back as last-resort drugs. Fast, accurate AI-based resistance prediction can route them to the right patients and slow resistance — but only if the models rest on reference materials pairing authenticated strains, genomes and MICs from the same isolate, plus well-characterized susceptible, borderline and discordant cases. The same materials strengthen preclinical test panels for new candidates.

The ATCC authors close with a direct call: clinical labs, culture collections, drug developers and model builders should contribute well-characterized isolates, adopt common benchmarks, and hold AI resistance prediction to the same standard of evidence as the drugs it protects. New antibiotics are too scarce to squander.

via who.int (Original)

Share this article:

More from Olivia Hart

Olivia Hart

Show full bio

Senior reporter covering money, funding and market shifts around global pharmaceutical industry at PharmInsider.

13 articles

Linked filings

  1. K-6477
  2. K-7275
  3. E-4696
  4. T-7549
  5. E-9238

Next filing »