The team succeeded by developing a system that can find molecular structures with desired traits (say, killing bacteria) more effectively than past systems. Unlike previous methods, the neural networks learn representations of molecules automatically, mapping them into continuous vectors that help predict their behavior. Once ready, the researchers trained their AI on 2,500 molecules that included both 1,700 established drugs and 800 natural products. When tasked with looking at a library of 6,000 compounds, the AI found that halicin would be highly effective.
Don't expect a prescription for halicin any time soon. MIT successfully used the medicine to eradicate A. baumanii (a common infection for US soldiers in Afghanistan and Iraq) in mice, but hasn't used it in human trials. This could be just the start of a much larger trend, mind you. The scientists have already used their model to screen over 100 million molecules in another database, finding 23 candidates. They also hope to design antibiotics from scratch and modify existing drugs to increase their effectiveness or reduce their unintended side effects. This is far from guaranteed to finish off "superbugs." If it takes out even some of them, though, it could save many lives.