Pharmaceutical companies spend billions researching blockbuster drugs 鈥� and billions more on legal fees to protect their recipes for making them. But an artificial intelligence has shown it can find a legal way to copy such drugs in seconds.
Chemists create recipes for new compounds like drugs through a process called retrosynthesis. They start with the desired chemical structure and work backward, disconnecting chemical bonds until they are left with a list of simple, cheap starting materials. Making those bonds in the reverse order in the lab should then produce the compound.
Retrosynthesis is considered something of an art among chemists, because the need to choose which bonds to cut at each stage means there is no right answer, only more or less optimal solutions. But Bartosz Grzybowski at the Ulsan National Institute of Science and Technology (UNIST) in South Korea has been developing an AI called Chematica to master this art.
Advertisement
Last year Grzybowski finally proved that Chematica can devise better, shorter routes to complex compounds than chemists and that the routes really work in the lab.
Now he has started using the AI to find routes that are not just effective, but different from previous attempts. This matters because pharmaceutical companies hold patents that protect聽both the sale of the drug and many of the different routes to make it. In other words,聽competitors聽that are able to sell a drug after its patent expires often have to pay the original firm to use the still-patented routes.
Chemical pathways
That is, unless they can come up with routes that don鈥檛 infringe the patents 鈥� which is exactly what Chematica has done聽for an antibiotic, a bone cancer drug and a diabetes medicine.
The AI digested the hundreds of patents describing protected reactions for making the drugs, then developed synthetic routes that avoided them. The difficult part was programming Chematica so it would avoid the roadblocks, which Grzybowski had not done before.
鈥淧harmaceutical companies spend billions on making sure there are no loopholes and people think these patents are bulletproof,鈥� says Grzybowski. 鈥淏ut actually it seems quite possible to get around them.鈥�
In 2012, Grzybowski and his colleagues used Chematica to find ways of making several extremely toxic chemicals, including the nerve agent VX gas, including water, table salt and sulphuric acid. This earned him an invitation to the Pentagon, he says, where he pointed out that we ought to be better regulating chemicals that can be used to make chemical weapons.
This new work is a similar kind of wake-up call to pharmaceutical companies, he says: 鈥淚 think they should be a little more diligent.鈥�
Derek Lowe, a drug discovery chemist who works for a pharmaceutical firm based in Cambridge, Massachusetts, says the paper is interesting but not surprising. 鈥淭his is exactly the sort of thing I鈥檇 expect AI to be used for,鈥� he says.
He says that generic drug firms already work hard to find ways around patents. 鈥淭here have always been loopholes in patents; there are just so many ways of skinning these cats,鈥� he says. The AI is just a much better way of finding them.
Lowe does expect that pharmaceutical companies will increasingly adopt AIs like Chematica. In fact, Grzybowski鈥檚 software was bought by German pharma giant Merck last year, which sells access to it under the brand name . 鈥淎Is will not replace chemists,鈥� says Lowe. 鈥淏ut the chemists that use AIs will replace the chemists that don鈥檛.鈥�
Chem