If A Computer Ever Finds A New Antibiotic… NAME IT AFTER HAL

London, 1928. Alexander Fleming comes back from a summer holiday to a messy lab, and on one forgotten dish a bit of mould has made a clear ring in the bacteria. That mould became penicillin. About ninety years later, a team at MIT found a new antibiotic too, but this time nobody waited for luck. A computer went looking. We climb into it one level at a time. Why we need new antibiotics: bacteria evolve, the few that survive a drug multiply, and some hospital germs like Acinetobacter baumannii now survive every antibiotic we know. How scientists normally search: one molecule at a time, near what already worked, like the man looking for his keys under the streetlight because that's where the light is. How the MIT team taught a computer what a germ killer looks like: about 2,500 molecules tested by hand against E. coli, then an AI that sees a molecule as dots and lines. What it found: halicin, an old molecule once studied for diabetes and then forgotten, that killed many dangerous bacteria, cleared a superbug infection in mice within about a day, and caused no resistance in E. coli over 30 days while ciprofloxacin climbed to about 200 times. And who really found it: the humans who built the data, picked the library and did the tests, and a model that made looking cheap enough to search in the dark. One honest note: as far as public records show, halicin is still not a medicine you can get. It is still being tested in animals, and a 2024 rat study found problems at a high dose. Sources Fleming, 1928: the messy lab, the contaminated plate, the clear zone around the mould: https://www.statnews.com/2017/03/01/p... Fleming's discovery, overview: https://pubmed.ncbi.nlm.nih.gov/15175... MIT News, Feb 20, 2020 (Collins and Barzilay, Stokes first author, about 2,500 training molecules on E. coli, about 6,000 molecules in the Drug Repurposing Hub, once studied for diabetes, named after HAL in 2001: A Space Odyssey, killed M. tuberculosis, cleared A. baumannii in mice within 24 hours, no resistance over 30 days vs ciprofloxacin within 1 to 3 days and about 200 times after 30 days, the electrochemical gradient, over 100 million ZINC15 molecules in three days): https://news.mit.edu/2020/artificial-... Stokes et al. (2020), "A Deep Learning Approach to Antibiotic Discovery", Cell (23 tested predictions from the 107 million ZINC15 molecules, 8 antibacterial): https://dspace.mit.edu/handle/1721.1/... Broad Institute release: https://www.broadinstitute.org/node/6... Halicin safety study in rats, 2024 (weight loss and slight kidney inflammation at a high dose): https://www.ncbi.nlm.nih.gov/pmc/arti... Halicin overview: https://en.wikipedia.org/wiki/Halicin The "map in the dark" is a real run of a tiny learner on a toy map we made up (2,500 toy training molecules, a 6,000-molecule toy library); its rank and scores are toy numbers, not the real study's. The resistance chart is our sketch of the reported numbers. The selection dish is a small simulation (400 cells, two drug rounds). The headline, the cupboard, the price tags, the pharmacy shelf and the cell's "electric skin" are simplified illustrations. Halicin's dots-and-lines drawing follows its real structure. The red camera eye is our own drawing, not the film. Painted scenes are illustrations; no real person is painted. Photos and images (Wikimedia Commons) Alexander Fleming in his lab: Imperial War Museums, public domain Fleming's penicillin mould sample, 1935: Science Museum London / Science and Society Picture Library, CC BY-SA 4.0 US wartime penicillin poster: NARA, public domain MIT Great Dome: Mys 721tx, CC BY-SA 3.0 Acinetobacter baumannii (electron microscope): CDC, public domain E. coli (electron microscope): Eric Erbe and Christopher Pooley, USDA ARS, public domain Mycobacterium tuberculosis (electron microscope): CDC / Janice Carr, public domain James J. Collins: GAEA AI, CC BY 3.0 Regina Barzilay: Christopher P. Michel, CC BY 2.0 Lab mouse: Rama, CC BY-SA 2.0 FR Structures: halicin (Innerstream, public domain), penicillin core (Yikrazuul, public domain), ciprofloxacin (AlyInWikiWonderland, CC0) Logo: MIT (Commons, public domain), used to name the institution. Icons: Material Design Icons (Apache 2.0). Chapters 0:00 A mess in London, 1928 0:25 He looked closer 0:35 Ninety years later, at MIT 0:52 How you search 1:10 Level 1: why a new antibiotic? 1:43 The empty cupboard 2:03 Level 2: how scientists search 2:27 The streetlight joke 2:49 Level 3: teaching a computer 3:07 2,500 molecules by hand 3:22 A molecule as dots and lines 3:43 Scoring the library 4:11 Level 4: did it work? 4:35 No resistance in 30 days 5:08 100 million molecules 5:25 Level 5: who found it? 5:50 The cost of looking 6:07 Where the light is 6:28 Making your own luck 6:55 A different angle