By now, most tech investors, businesses and even consumers are aware of the promise of generative AI to transform fields from software programming and advertising to video game development and movie production.
But the use of large language and diffusion models to discover medicines is not widely discussed. While there still aren’t many biotech startups that are employing generative AI methods to create novel drug compounds, some early adopters have succeeded in raising substantive funding in recent months.
“We are in [the] very early days of how generative AI is going to be applied in life sciences, but I think the applications in drug discovery are potentially profound,” said Greg Yap, a partner at Menlo Ventures.
His firm is a backer of Genesis Therapeutics, a biotech startup that uses various AI methods, including generative AI, to design new drug candidates. Earlier this week, the Burlingame, Calif.-based company raised a $200 million Series B co-led by Andreessen Horowitz.
Genesis’ large deal size stands out at a time when raising new capital has become very difficult for tech and biotech startups. The median deal size for early-stage US-based life sciences companies that develop drugs has hovered around $20 million for the past three years, according to PitchBook data.
While the drugs designed by Genesis are not yet being tested in human trials, its computer-created compounds are showing promise. The company has been using generative AI in conjunction with predictive AI models since it spun out of Stanford University in 2019.
Other biotech companies that use generative AI include Recursion, which went public in 2021. In May, Recursion announced its plans to acquire Air Street Capital and Fifty Years-backed Valence Discovery, a startup with what Yap says are “strong AI capabilities.”
Shortly after that acquisition, Recursion received a $50 million investment from NVIDIA, whose chips are in high demand for generative AI applications.
Recursion plans to use its dataset in conjunction with Valence’s AI technologies and the giant chipmaker’s computing capabilities to build foundation models for biology and chemistry.
While the current versions of Genesis’ and Recursions’ models are based on the same principles as other foundational models, they incorporate something that most tech models don’t: the laws of physics. That’s because they aim to create molecular structures that can be made in a three-dimensional, physical world.
Although Genesis has been applying its models only to create simple molecules for medicines that can be taken in a pill form, Yap said that the future of generative AI in drug discovery is in making proteins because they can perform a wider variety of functions in the body and be injected or administered intravenously.
Designing proteins is much more complicated. “The physical structure of a small molecule is like a bicycle, but protein is like an airplane,” Yap said. Menlo Ventures has recently backed an undisclosed company applying generative AI to discover proteins.
The biggest roadblock to developing strong models is finding solid training data, according to Yap. “These models are being trained on every piece of public scientific data that describes molecular interaction, but some of the data is not very clean,” he said, adding that there are mistakes in publications.
But once the scientific community can find ways to improve the quality of training sets, generative AI for drug discovery is expected to take off.
“These models have now reached a stage of maturity that you can actually see that they will work. It’s just a question of when,” Yap said.
Related read: PitchBook’s Q1 2023 Digital Health Report
Correction: A previous version of this article misspelled the name of Valence Discovery and listed the wrong VC investors in the company. (Aug. 28, 2023)
Featured image by metamorworks/Getty Images
Learn more about our editorial standards.
