Drug discovery AI designs 40,000 potential chemical weapons in 6 hours
Of the areas where artificial intelligence is good, sifting through thousands of chemical compounds to identify drug candidates is easily one of the best. However, the researchers found it was also remarkably good at imagining possible chemical weapons – scary.
In a recent study published in the journal Nature Machine Intelligence, a team from pharmaceutical company Collaborations Pharmaceuticals, Inc. repurposed a drug discovery AI. He successfully identified 40,000 potential new chemical weapons in just 6 hours, some of them remarkably similar to the most powerful nerve agent ever created.
According to an interview with The Verge, researchers were shocked at how easy it was.
“For me, the concern was how easy it was to do. A lot of things we’ve used are free. You can go and download a toxicity dataset from anywhere. If you have someone who knows how to code in Python and has machine learning capabilities, then probably in a good weekend of work they could build something like this generative model driven by toxic datasets,” said Fabio Urbina, lead author of the article. , to the Verge.
“So that’s what really got us thinking about releasing this document; it was such a low barrier of entry for this type of misuse.
To redirect the AI to suggest something that is causing harm instead of healing, the researchers needed to direct it toward identifying toxicity.
By taking their AI MegaSyn, which generally rewards bioactivity (how the drug interacts with the target) and penalizes toxicity, they simply reversed the toxicity metrics but kept the bioactivity reward, now also rating the drugs more highly in depending on their toxicity.
In the 6 hours they ran the AI, it made some frightening developments. After targeting him for the generation of nerve agent-like compounds, he suggested VX, the most potent nerve agent ever created, used in the assassination of Kim Jong-un’s brother, Kim Jong-nam, as well as other agents used in chemical warfare.
Undeterred, he also designed agents that were supposed to be even more toxic than VX. The researchers said that while the predictions are unverified, and they “certainly don’t want to verify that” themselves, the predictive models created by MegaSyn so far are reliable. There will likely be false positives, and the compound would need to be synthesized to be tested, so it’s unclear how many of these compounds would actually be toxic.
The team believe this should be a telling moment for the use of AI in drug discovery, highlighting the ease of abuse of these algorithms.
“Without being overly alarmist, this should serve as a wake-up call to our colleagues in the ‘AI in drug discovery’ community,” the authors write.
“The reality is that this is not science fiction. We are just a very small company in a universe of several hundred companies using AI software for drug discovery and de novo design. How many of them have even considered reusing or abusing the possibilities?”