Google DeepMind CEO Demis Hassabis suggested on Tuesday that clinical trials for the first drug designed with the help of artificial intelligence could begin this year.
“Several AI-designed drugs will be in clinical trials by the end of the year,” said Hassabis, who also runs DeepMind’s drug discovery spin-off Isomorphic Labs, during a panel discussion at the World Economic Forum in Davos. It’ll come in…that’s the plan.”
Isomorphic Labs has been trying to use machine learning to accelerate drug development since 2021. “Eventually, we will be able to imagine personalized medicine that is optimized, perhaps overnight, by AI systems to suit an individual’s metabolism,” he said.
AI hype is everywhere now, but Hassabis and his colleague John Jumper won a Nobel Prize for their work on AlphaFold, a deep learning system that can predict the structure of proteins.
Pharmaceutical companies are interested in AI because of the potential for significant time and cost savings. According to a recent article published in the Journal Nature Medicine, it takes 12 to 15 years and approximately $2.6 billion to successfully develop a new drug and get it approved for use.
Many drugs are not approved for use because clinical trials in humans ingesting the drug have a success rate of less than 10%. Anything that can reduce costs, accelerate development, and increase success rates has a significant impact on a pharmaceutical company’s bottom line.
Researchers believe there are many ways machine learning models can improve and speed up parts of the drug discovery process. Hassabis believes significant time and cost savings are possible.
This kind of optimism needs to be tempered, as high-quality training data is difficult to obtain due to privacy regulations, data sharing policies, and acquisition costs.
Hassabis believes these challenges are not insurmountable. “We can generate important data to fill in gaps that are not available in public data,” he says.
This can be done in collaboration with clinical research organizations or through the use of synthetic data, which AlphaFold2 uses extensively, he said. However, as previously discussed, synthetic data can be problematic.
“If you use synthetic data, you have to be very careful to make sure it actually represents the distribution correctly and that you’re not training your own errors in any way,” Hassabis said. say.
Hassabis doesn’t think AI will replace scientists anytime soon.
“True invention is not yet possible with AI. AI cannot come up with new hypotheses or new speculations. It could perhaps solve complex speculations, such as in mathematics. We are very close to a major breakthrough in this regard. “I think we’ll really see that this year, but it’s not the same as actually coming up with a theory or a hypothesis like a good human scientist does,” he said.
Hassabis is not alone in exploring the application of machine learning in drug discovery. Nvidia is also enthusiastic about AI-enhanced drug discovery, presumably because it gives it more reason to buy hardware.
Last fall, Nvidia open sourced the BioNeMo family of GPU-accelerated machine learning frameworks for drug development and molecular design. The company has also taken steps to repackage existing models such as DeepMind’s AlphaFold2 and MIT’s DiffDock 2.0 as microservices to make them easier to consume.
Nvidia is also partnering with major pharmaceutical companies, including Danish pharmaceutical giant Novo Nordisk, to bring new research systems online. One example of such efforts is Gefion, a Danish supercomputer that applies machine learning to biology and the development of new treatments. ®