AI Science Tracker

Evidence dossier · Editorial review 16 Nov 2026

Beyond the petri dish?

How far have AI-discovered antibiotics progressed beyond laboratory assays?

What these four studies establish

Several candidates reached mouse infection experiments; that is a step beyond a laboratory growth assay. The selected source reports do not establish efficacy in patients. They describe different compounds, organisms and experiments, so they cannot be treated as successive clinical stages of one drug.

AI contributed candidate selection or molecular design. Researchers synthesized or obtained compounds and measured activity experimentally. A large virtual library is a search space, not a count of effective medicines.

Scope: four catalog studies, 2020–2025, selected to contrast screening, generative design, infection experiments and toxicity-only animal testing. This is an illustrative synthesis, not an exhaustive or systematic review. The original institutional reports linked below were checked on 16 Nov 2026; full papers, trial registries and all later literature were not re-reviewed. Absence from this dossier is not absence from the research literature.

Compare the actual experiments

Counting unit: one existing catalog study per card. The candidates and conditions differ; no pooled success rate is calculated.

Halicin · 2020

Result: antibacterial activity and clearance in reported mouse infection experiments.

AI contribution: learned structure–activity screening selected an existing compound.

Test and comparison: bacterial assays, mouse infection models, and a laboratory resistance comparison with ciprofloxacin.

Limit: the source planned further development for humans; mouse efficacy does not establish patient benefit.

Original MIT report · 20 Feb 2020

Abaucin · 2023

Result: targeted activity against A. baumannii and treatment of mouse wound infections.

AI contribution: ranked compounds for experimental selection.

Test and comparison: activity against resistant isolates; little activity against other bacterial species in the reported assays.

Limit: narrow-spectrum activity is a property in these experiments, not demonstrated clinical safety or efficacy.

Original MIT report · 25 May 2023

SyntheMol · 2024

Result: six of 58 synthesized compounds killed a resistant A. baumannii strain in laboratory tests.

AI contribution: generated molecules and synthesis recipes.

Test and comparison: antibacterial assays; two compounds underwent mouse toxicity testing. Four others were not water-soluble.

Limit: toxicity testing is not an infected-animal efficacy experiment. The source describes that as a next step; mechanism remained unresolved.

Original Stanford report · 28 Mar 2024

Generative NG1 / DN1 · 2025

Result: NG1 activity in a mouse gonorrhea model and DN1 clearance of MRSA skin infection in mice.

AI contribution: fragment-based and unconstrained molecular generation, followed by computational screening.

Test and comparison: synthesis and bacterial assays followed by separate pathogen-specific infection models.

Limit: only two of 80 selected fragment candidates could be synthesized. Further preclinical optimization was underway; these reports do not demonstrate human benefit.

Original MIT report · 14 Aug 2025

What connects these results?

MIT explicitly places abaucin after the halicin screening program and describes the 2025 generative work as building on both. These are program extensions with new compounds. They are not independent replications or clinical follow-ups of the earlier candidates. SyntheMol is included for its contrasting synthesis-aware design and animal-test boundary; this dossier does not assert a scientific dependency between it and the MIT studies.

Source-backed directed relationships and dated follow-up events are recorded on the linked discovery pages. They preserve the original catalog entries and never merge their counts.

What remains outside the evidence here?

These selected reports leave human efficacy and safety unresolved. Direct comparisons across candidates, independent replication and a comprehensive account of later clinical development were not assessed in this review. This is a coverage boundary, not a claim that such work does not exist.

Maintenance: this is a dated editorial snapshot, maintained by the catalog editors when new source-backed evidence is incorporated. No continuous monitoring or scheduled re-review is claimed. Corrections should update both the relevant entry and this synthesis.

Search antibiotic-related records across collections · Explore the biology catalog · Evidence and selection rules