Reported Brilliant
A model trained on 2,500 molecules found an antibiotic that screening had walked past
It had been sitting in a library of known compounds, investigated as a diabetes drug and never pursued as an antibiotic.
A system that searches differently finds different things. The value is in what nobody told it to look for.
In February 2020 researchers at MIT trained a neural network on about 2,500 molecules to predict which would kill bacteria, then pointed it at the Broad Institute's library of known compounds.
It flagged one that had been investigated as a diabetes treatment and never taken further as an antibiotic. Structurally it looks nothing like existing antibiotics, which is why conventional screening had passed it over. Named halicin, it killed every species the team tested except Pseudomonas aeruginosa, including strains resistant to everything else available, and cleared infections in mice. E. coli developed no resistance to it over thirty days; resistance to ciprofloxacin appeared within one to three days. The team then screened more than 100 million molecules in three days.
Halicin is a candidate, not a medicine. Nobody has been treated with it. What the model did was look somewhere no one had thought to look.