The story began in 2021, when Collaborations Pharmaceuticals, which uses computers to help its customers identify molecules that look like potential drugs, was invited to present a paper on how such drug-discovery technologies might be misused. The venue was a conference organised by the Spiez Laboratory, in Switzerland. This is a government-funded outfit that studies risks posed by nuclear, biological and chemical weapons. To prepare for the presentation some of Collaborations’ researchers carried out what they describe as a “thought exercise” that turned into a computational proof of concept for making biochemical weapons.
Their method was disturbingly simple. They took a piece of drug-discovery software, called MegaSyn (a piece of artificial intelligence, ai, which the company has developed for the purpose of putting virtual molecules together and then assessing their potential as medicines), and turned one of its functions upside down. Instead of penalising probable toxicity, as makes sense if a molecule is to be used medically, the modified version of MegaSyn prized it.
The result was terrifying. Trained on the chemical structures of a set of drug-like molecules (defined as substances easily synthesised and likely to be absorbed by the body) taken from a publicly available database, together with those molecules’ known toxicities, the modified software required a mere six hours to generate 40,000 virtual molecules that fell within the researchers’ predefined parameters for possible use as chemical weapons.
The list included many known nerve agents, notably vx, one of the most toxic. But the software also came up with not-yet-synthesised substances predicted to be deadlier still.
How to tweak drug-design software to create chemical weapons https://t.co/g3B5UcUBW1 #strategy #competitiveintelligence #marketing #pharma #productmarketing #biotech #pharmaceutical #healthcare pic.twitter.com/46VyMIRsKE
— Dr Timos Papagatsias (@_timos_) March 16, 2022