“We’re in a frothy time,” acknowledges Alexandra Snyder, the head of R&D at Generate:Biomedicines.
Generate is one of the AI-driven biotechs that make up a new generation of companies with huge funding, A-list science and tech stars and computing horsepower. On Tuesday, the $1 billion-plus launch of Xaira Therapeutics — the second-biggest initial biotech startup funding ever — took a field that’s already described by some as overhyped and whipped it into a frenzy.
Alexandra Snyder“Just as when you serve a frothy drink, you enjoy the froth and then it takes a little while for you to realize how much liquid is actually in the cup,” Snyder said in an interview this month, before Xaira’s debut. “We’re kind of in that stage right now — enjoying the froth.”
Over the next few years, the biotech industry will find out how valid those bets are.
The new leaders of the AI biotech field — Generate:Biomedicines, Isomorphic Labs and Xaira Therapeutics — are hoping they can do what an earlier crop of equally-hyped companies that launched a decade ago couldn’t.
BenevolentAI said that its technology would accelerate science faster than humanly possible. Exscientia aspired to be “the first company to automate drug design.” And Recursion Pharmaceuticals set the goal of creating 100 drug candidates in 10 years. All three have fallen short, and none has carried a drug to approval — the only benchmark that ultimately matters.
“I think there was an expectation that AI was going to revolutionize drug discovery immediately,” Nessan Bermingham, an operating partner at Khosla Ventures, said in an interview with Endpoints News. “The reality is, it’s not. It’s going to take some time.”
The newest trio of mega-startups believes the time is now. Xaira’s founding investors say its initial $1 billion could further expand. Generate has reeled in nearly $750 million from investors and partnerships. And while Isomorphic hasn’t disclosed funding, it is backed by the vast resources of Alphabet and has partnered with Eli Lilly and Novartis.
What’s different this time?
Much of the excitement this time around is based on the explosion of AI technologies, including the transformer model that helped power OpenAI’s ChatGPT, as well as Google DeepMind’s AlphaFold, a program invented by the team behind Isomorphic that predicts a protein’s three-dimensional structure. Scientists from Generate and Xaira used another method called a diffusion model to make their own programs, Chroma and RFdiffusion, for creating proteins from scratch. That process, called de novo protein design, is being used to create and refine new drugs.
David Baker“Ten years ago, there really wasn’t a demonstration that AI methods could, in fact, transform a key part of the drug discovery field,” said David Baker, a co-founder of Xaira who leads the Institute for Protein Design in Seattle. Several scientists who developed RFdiffusion and similar programs in his lab now work at Xaira.
The ability to generate massive amounts of biological data through genetic sequencing, microscopy images, protein structures and more has also exploded, along with the computing power to make sense of it.
Even so, many scientists believe that having enough data will continue to be the biggest bottleneck. After all, running lab experiments is a slower and more expensive process than scraping text and images from the internet.
Tellingly, none of the leading AI biotechs believe they can skip experimentation or replace wet labs. Generate has spent tens of millions building out its cryogenic electron microscope capabilities. Even Isomorphic, which is seeking to push the limits of a computer-first vision under CEO Demis Hassabis, has said it intends to build its own wet labs in the future.
A critical part of Xaira’s vision is producing massive amounts of experimental data to feed and improve its models.
“There will always be the design-test cycle,” said Marc Tessier-Lavigne, the former Stanford president who serves as Xaira’s CEO. “You’re not going to eliminate any of the steps. But can you make the steps both better and faster? Yes — but probably even more important than faster is increasing the chances of success. That is where AI is going to make a big difference.”
Broader technological advances are helping biotech too. Nvidia has raised the bar with its supercomputer builds that can process massive flows of data.
Alex Aravanis“I feel like we’ve turned the corner,” said Alex Aravanis, the CEO of Moonwalk Biosciences and former chief technology officer of Illumina, who oversaw the company’s AI models for predicting how never-before-seen genetic mutations contribute to disease. Aravanis, who is an advisor to Xaira’s founding incubator Foresite Labs, pointed to AlphaFold, Baker’s protein design, and Profluent Bio’s recent AI-designed gene editing tools as examples.
“We have had that moment where the models are good enough, the datasets are good enough,” he said. “They can do things that we could never do before.”
Numerous smaller startups, including Atomic AI, CHARM, Genesis, Iambic, Profluent Bio and Terray, have also launched in the past few years. They join a comparatively older group of privately held AI biotechs like Atomwise, Deep Genomics, Insilico Medicine, and insitro.
“Xaira is great. It makes us look very cheap,” said Alex Zhavoronkov, CEO of Insilico Medicine, an AI-focused biotech founded in 2014 that has put several drugs in the clinic and has raised over $400 million as it considers an IPO on the Hong Kong market.
‘On the precipice,’ but not over it
Nessan Bermingham“What’s very clear is that AI and machine learning are here to stay. It is on the precipice of having some really profound impact on drug discovery and development,” Bermingham said. “The problem we all have is we don’t know when that precipice is.”
Making medicines is nothing like upgrading a chatbot, where new versions can be quickly rolled out and tested. And the FDA’s requirements for preclinical safety tests in animals may act as a speed limit of sorts for AI drug designers.
Elizabeth Schwarzbach“IND-enabling studies are really hard to complete in less than 18 months. So we really have limited control on the timelines after we design a high-quality development candidate,” said Elizabeth Schwarzbach, chief business officer of BigHat Biosciences, a startup that has raised more than $100 million to develop antibodies with AI and announced a partnership with Johnson & Johnson on Wednesday.
As challenging as those preclinical tests can be for a small startup, it’s nothing compared to the lengthy and expensive process of testing drugs in people. Tessier-Lavigne said one of the company’s goals is to develop AI models for patient stratification and prediction of drug responses to increase the odds of trial success, but he didn’t share details on how it would work.
Additionally, not all in the field see some of the more ambitious AI ideas, like designing antibodies completely from scratch, as ready for prime time. Generate’s lead antibodies, for instance, weren’t designed with its de novo model called Chroma. And Snyder, the startup’s R&D head, said the field needs to improve on the data and modeling before de novo antibody generation can take off.
“On geological time, we’re pretty early,” Snyder said. “The people who made the first airplanes certainly couldn’t imagine a Concorde, even though they had the right idea.”
Many academics back that interpretation, describing a recent model called RFantibody from Baker’s lab as intriguing but early, producing antibodies that are still too weak to be good drugs. But one of those scientists, Nicholas Polizzi, a protein designer at Dana-Farber Cancer Institute, said that it’s hard to find the level of funding required to design, make and test de novo antibodies and improve the AI models in an academic lab.
“Try getting that from the NIH. It’s going to be tough,” Polizzi said. “So even if it’s premature, which it might be — probably is — it’s exciting to see that they’re well-funded. And the recipe that they need to achieve their goals is pretty clear. They need more data.”