GENEVA — Venomous Agent X is a deadly nerve agent, though you likely know it by another name: VX. It’s an amber, oil-like liquid that targets the body’s nervous system. A single drop on the skin can kill within minutes. In 2017, North Korea is believed to have used VX to assassinate Kim Jong Un’s estranged half-brother in a Malaysian airport. Kim Jong Nam suffered severe paralysis, dead in about 20 minutes from a weapon of mass destruction.
Sean Ekins and his team thought of the toxin for a possible experiment, one he needed to meet a last-minute deadline for a presentation at the Spiez Laboratory in Switzerland, at a conference examining how developments in science and technology might affect chemical and biological weapons regimes. Ekins is a scientist and CEO of Collaborations Pharmaceuticals, a lab that uses machine learning platforms to seek therapeutic treatments for rare and neglected diseases. He and his colleague Fabio Urbina wanted to test and see if they could flip their AI software, MegaSyn. Instead of steering the software away from toxicity, they wanted to see if they could guide the model toward it.
The scientists trained the software with some 2 million molecules from a public database, and then modeled for specific, toxic traits.
In just six hours, the AI generated some 40,000 molecules that met the scientists’ criteria, meaning that, based on their molecular structure, they all looked quite a lot like toxic chemical agents. The AI designed VX. It designed other known toxic agents. It even designed entirely new molecules that the scientists hadn’t programmed for, creating a sketch for potentially lethal and novel chemical compounds.
The experiment was computational — a digital recipe for molecules like VX, not a physical creation of it or any other substance. But Ekins and his team used open source, publicly available data. The AI they used was also largely open source as well; they just tweaked the models a little bit.
Read the full article here
