Nabla Bio, a small startup founded by scientists from geneticist George Church’s lab at Harvard Medical School, has formed drug discovery collaborations with AstraZeneca, Bristol Myers Squibb and Takeda, the company told Endpoints News in an exclusive interview.
The startup is based on large language models developed in Church’s lab that predict protein structures. Nabla and its partners are building upon that AI research to design antibodies against proteins that have proven frustratingly difficult to drug with traditional methods.
Simon Chell“We can identify hits that you wouldn’t have seen previously. That is the biggest pull for these types of technologies,” said Simon Chell, VP of drug discovery at AstraZeneca. “That’s a much bigger leap in achievement than a shaving on a timeline.”
The companies are not disclosing specific targets or diseases that they’re working on or divulging the financial details of each partnership. Nabla said that the three deals in total are worth more than $550 million in upfront and milestone payments, plus royalties.
The Cambridge, MA-based startup also said it raised a $26 million Series A financing led by Radical Ventures. Existing investors chipped in, too, including Khosla Ventures and Zetta Venture Partners, which co-led Nabla’s $11 million seed round in 2021.
The sum pales in comparison to what other leading startups using AI to make antibodies have raised. That includes the $1 billion that Xaira Therapeutics raised at its launch last month and the nearly $750 million that Generate:Biomedicines has raised since 2020. But Nabla’s executives are unphased by the competition.
“Far more often than not, you don’t need billions of dollars to build transformative technology,” CEO Surge Biswas said during a visit to the startup’s small lab. “We raised what we needed, and we could have raised a lot more.”
Biswas wouldn’t say how much additional funding he turned away. But unlike Nabla’s competitors, which are trying to fill their own pipelines and run their own clinical trials, Nabla will be a partnership-focused company for now. “It seems almost selfish at this point to develop our own drugs,” he said.
‘A difference in kind’
The deals are part of a growing trend in which pharma companies partner with smaller biotechs developing their own AI models to improve antibody drug candidates or even create completely new ones from scratch — a process called de novo design. BigHat Biosciences has struck several partnerships, including with Merck and Johnson & Johnson. And Generate:Biomedicines has a collaboration with Amgen that could earn it up to $2.2 billion.
Biswas and colleagues from the Church Lab published their first major paper in 2019 examining a computer program that can predict the structures and properties of the biomolecules. A few years later, the team showed they could design new proteins better and faster than popular protein prediction software such as AlphaFold2, developed by researchers at Alphabet.
Dylan ReidDylan Reid, a partner at the AI-focused venture firm Zetta, which first invested in Nabla three years ago, thought it was a big shift from earlier AI work in biotech. “It felt very different than the kinds of things that you could do with traditional biostats,” he said. “The ability to generate sequences was a difference in kind versus something that was incremental.”
Biswas wouldn’t share specifics about how Nabla has built upon the approach since he left the Church Lab. But he emphasized the importance of improving the model with results from its lab experiments and not just using publicly available data.
Nabla is aiming its technology at a longstanding challenge for drugmakers. While scientists have mastered making antibodies against free-floating proteins, targeting the hundreds of proteins wedged in cell membranes is more difficult.
“They’re involved in every major physiological process,” Biswas said. “They are basically on/off switches for cells, and they have slightly different shapes when they are on and off, and a good drug needs to differentiate from that.”
These proteins include a laundry list of targets familiar to biologists including ion channels, transporters, and G protein-coupled receptors, or GPCRs. Proteins with very different functions often have very similar structures. In theory, antibodies can narrow in on those differences, yet like icebergs, only a small tip is exposed above the cell membrane, making it tough for antibodies to get a foothold.
Some biotechs are trying to solve these challenges with clever experimental approaches. Nabla’s solution is to design antibodies on a computer that bind to the right part of the right protein in the right conformation. And gauging the feasibility of that de novo design approach is a key focus of AstraZeneca’s interest in the startup.
“The generative outputs are fairly low affinity at the moment,” Chell said, echoing recent critiques of academic work in the space that found that de novo-generated antibodies were too weak to be used as drugs. But he still thinks it could be a good way to get leads on antibodies they otherwise might not find. “We’re definitely excited about the potential of it,” he added.