AI found a new antibiotic: an MIT deep-learning model screened over 107 million molecules and discovered halicin

Verification status: IN CHECKING - not publish-ready, not pending, not verified. The headline outcome rests on a peer-reviewed Cell paper (Tier 1), restated by MIT’s own news release and independently corroborated by two outlets distinct from both the journal and MIT: the Royal Society of Chemistry’s Chemistry World and The Guardian (both Tier 2). The open item is that web.archive.org could not be reached this session to freshly snapshot the sources (all four are saved locally and the archive gap is flagged); read the weak-source line below. The terminal call is the checker’s.

The problem

New antibiotics are discovered slowly and expensively, while resistant bacteria emerge fast: the study opens on the premise that “due to the rapid emergence of antibiotic-resistant bacteria, there is a growing need to discover new antibiotics” (source). Traditional screening tests physical compounds one library at a time, which makes exploring very large chemical spaces prohibitive; as the MIT team put it, “the machine learning model can explore, in silico, large chemical spaces that can be prohibitively expensive for traditional experimental approaches” (source).

What was built

The team trained a deep neural network to predict which molecules have antibacterial activity, then pointed it at existing chemical libraries. “Once the model was trained, the researchers tested it on the Broad Institute’s Drug Repurposing Hub, a library of about 6,000 compounds. The model picked out one molecule that was predicted to have strong antibacterial activity and had a chemical structure different from any existing antibiotics” (source). The system is a prioritiser, not a replacement for the lab: it ranks candidates that humans then test at the bench and in animals. MIT reports the trained model “can screen more than a hundred million chemical compounds in a matter of days” (source).

The outcome

The model discovered a structurally unusual antibiotic. Per the Cell paper, “We performed predictions on multiple chemical libraries and discovered a molecule from the Drug Repurposing Hub-halicin-that is structurally divergent from conventional antibiotics and displays bactericidal activity against a wide phylogenetic spectrum of pathogens including Mycobacterium tuberculosis and carbapenem-resistant Enterobacteriaceae. Halicin also effectively treated Clostridioides difficile and pan-resistant Acinetobacter baumannii infections in murine models” (source).

The result was not a single lucky hit. The paper reports that “from a discrete set of 23 empirically tested predictions from >107 million molecules curated from the ZINC15 database, our model identified eight antibacterial compounds that are structurally distant from known antibiotics” (source). MIT reports that this screen was fast: “this screen, which took only three days, identified 23 candidates that were structurally dissimilar from existing antibiotics and predicted to be nontoxic to human cells” (source). The senior author was blunt about the significance: “our approach revealed this amazing molecule which is arguably one of the more powerful antibiotics that has been discovered” (source). Independent trade coverage restates the same figures: the Royal Society of Chemistry’s Chemistry World reported that of the ZINC15 database’s compounds “107 million were selected for screening” and that “23 compounds were chosen for further investigation, with eight displaying antibacterial activity against a range of pathogens” (source), and that “they discovered a compound, halicin, which has impressive antibiotic activity, despite having a chemical structure unlike conventional antibiotics” (source). A second independent outlet, The Guardian, reported the same screen: “they set the algorithm working on 107m of these. Three days later, the program returned a shortlist of 23 potential antibiotics, of which two appear to be particularly potent” (source), and independently restated halicin’s spectrum: “tests on bacteria collected from patients showed that halicin killed Mycobacterium tuberculosis, the bug that causes TB, and strains of Enterobacteriaceae that are resistant to carbapenems … Halicin also cleared C difficile and multidrug-resistant Acinetobacter baumannii infections in mice” (source).

Weakest load-bearing source, named. The screening scale, the three-day runtime and the named-researcher characterisations come from the MIT News release, which is an institutional press office, not an independent journal, so on its own it is a promotional secondary rather than a neutral audit (source). What keeps the file honest is that every load-bearing figure (>107 million molecules, 23 predictions, 8 compounds, halicin’s spectrum and its murine-model efficacy) is also stated in the peer-reviewed Cell paper itself, so the outcome does not stand or fall on the press release (source).

How this was verified. Method: the Cell 2020 figures (>107 million molecules from ZINC15; 23 tested predictions; 8 antibacterial compounds; halicin from the Drug Repurposing Hub; bactericidal spectrum including M. tuberculosis and carbapenem-resistant Enterobacteriaceae; murine-model efficacy against C. difficile and A. baumannii) were quote-matched verbatim against the Europe PMC core record for PMID 32084340 (DOI 10.1016/j.cell.2020.01.021), saved to sources/europepmc-32084340-core.json - Tier 1, first-party peer-reviewed. The screening-scale and speed figures (>100 million compounds in days; three-day / 23-candidate ZINC15 screen; ~6,000-compound Drug Repurposing Hub) and the Collins and Barzilay quotes were quote-matched verbatim against the MIT News release of 2020-02-20, saved to sources/mit-news-halicin-20200220.html - Tier 2, but this is the subject’s OWN institutional press office (MIT is the client here), so it is not an independent party. The genuinely independent Tier 2 corroboration comes from two outlets distinct from both the journal and MIT: the Royal Society of Chemistry’s Chemistry World (Jamie Durrani, 2020-02-21, sources/chemistryworld-halicin-20200221.html), which restates the 107-million screening scale and the 23-candidate / 8-antibacterial result, and The Guardian (2020-02-20, sources/guardian-halicin-20200220.html), which restates the 107-million / 23-candidate / three-day screen and halicin’s spectrum against M. tuberculosis and carbapenem-resistant Enterobacteriaceae plus its clearance of C. difficile and A. baumannii in mice; both quote-matched verbatim. Date verified: 2026-09-10. web.archive.org returned HTTP 429 headless this session, so no fresh snapshot could be taken; all four sources are saved locally and the archive gap is flagged. No confirmation was sought from the authors: asking the subject to confirm its own numbers is a testimonial, not an audit.

Sources

1. Cell (Cell Press) · A Deep Learning Approach to Antibiotic Discovery · Stokes JM, Yang K, Swanson K, Jin W, Cubillos-Ruiz A, Donghia NM, MacNair CR, French S, Carfrae LA, Bloom-Ackermann Z, Tran VM, Chiappino-Pepe A, Badran AH, Andrews IW, Chory EJ, Church GM, Brown ED, Jaakkola TS, Barzilay R, Collins JJ · 2020-02-20, 180(4):688-702.e13 (PMID 32084340, DOI 10.1016/j.cell.2020.01.021, PMCID PMC8349178) · https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=EXT_ID:32084340%20AND%20SRC:MED&resultType=core&format=json — Tier 1 (first-party peer-reviewed primary study; source of the >107 million / 23 / 8 figures, halicin’s spectrum, and the murine-model efficacy; abstract quote-matched verbatim; saved to sources/europepmc-32084340-core.json; web.archive.org 429 this session, archive gap flagged).
2. MIT News, Massachusetts Institute of Technology · Artificial intelligence yields new antibiotic · 2020-02-20 · https://news.mit.edu/2020/artificial-intelligence-identifies-new-antibiotic-0220 — Tier 2 (MIT’s own institutional press office; source of the >100-million-in-days screening scale, the three-day / 23-candidate ZINC15 screen, the ~6,000-compound Drug Repurposing Hub selection, and the Collins/Barzilay quotes; quote-matched verbatim; saved to sources/mit-news-halicin-20200220.html; web.archive.org 429 this session, archive gap flagged). Note: MIT is the client here, so this is the subject’s own release, not an independent party.
3. Chemistry World, Royal Society of Chemistry · AI tool screens 107 million molecules, discovers potent new antibiotics · Jamie Durrani · 2020-02-21 · https://www.chemistryworld.com/news/ai-tool-screens-107-million-molecules-discovers-potent-new-antibiotics/4011233.article — Tier 2 (genuinely independent trade press, distinct from both the journal and MIT; independently restates the 107-million screening scale, the 23-candidate / 8-antibacterial result, and halicin’s discovery and broad activity; quote-matched verbatim; saved to sources/chemistryworld-halicin-20200221.html; web.archive.org 429 this session, archive gap flagged).
4. The Guardian · Powerful antibiotic discovered using machine learning for first time · 2020-02-20 · https://www.theguardian.com/society/2020/feb/20/antibiotic-that-kills-drug-resistant-bacteria-discovered-through-ai — Tier 2 (second genuinely independent outlet, distinct from both the journal and MIT; independently restates the 107-million / 23-candidate / three-day screen and halicin’s spectrum against M. tuberculosis and carbapenem-resistant Enterobacteriaceae plus its clearance of C. difficile and A. baumannii in mice; quote-matched verbatim; saved to sources/guardian-halicin-20200220.html; web.archive.org 429 this session, archive gap flagged).

Related case files

• DeepMind’s AlphaFold, the CASP14 protein-structure breakthrough — the other landmark of deep learning applied to molecular science, on structure prediction rather than antibacterial screening.
• AlphaFold3, published in Nature without its code — deep learning extended from folding to how proteins bind other molecules, including drug-like ligands, the discovery-adjacent step; it also shows the reproducibility tension that halicin’s fully published Cell paper avoids.
• Boston Children’s used OpenAI’s o3 to re-analyse rare-disease cases — the same shape at a different target: a model surfacing a small set of novel, useful hits across an enormous candidate space for humans to validate, here new diagnoses across 376 cases rather than antibacterial candidates.

Path to green

Every headline figure is stated in the peer-reviewed Cell paper (Tier 1), restated by MIT’s own release and now independently corroborated by two genuinely independent Tier 2 outlets, Chemistry World (Royal Society of Chemistry) and The Guardian, which each restate the 107-million / 23-candidate screen and halicin’s spectrum and murine efficacy. That lifts corroboration on both critical claims to two_independent (score.mjs preview: c1 = c2 = 1.000, story confidence 1.000). The one open item is the archive gap: web.archive.org returned 429 headless this session, so the sources are saved locally and a fresh Wayback capture of all four should be taken once archive.org is reachable; the 2020 pages near-certainly already hold public snapshots. No confirmation was sought from the authors. The terminal call is the checker’s.