Protein-making AI models have fueled the launch of a $1 billion startup and a Nobel Prize win in 2024, but are they ready to actually make drugs?
Leaders of some of the largest startups working on that question say the technology is promising but not yet ready. But a tiny startup called Nabla Bio is far more bullish,
posting a preprint Thursday
highlighting its progress in creating antibodies from scratch that look therapeutically viable against some hard-to-drug targets in what it’s calling a first for the field.
This isn’t the holy grail of having fully optimized drug candidates at the push of a button. But Nabla’s research is a step forward. The biotech’s leaders say their antibodies showed strong binding affinities and, in general, looked and behaved more like drug candidates than previous
de novo
efforts.
Nabla said these are the first computer-generated antibodies against the family of proteins called GPCRs, which are notoriously hard to drug. The research includes eight examples of
de novo
antibodies against a range of targets, including proteins found in the membrane of cells like Claudin-4 and CXCR7.
“The antibodies are developable, and they could potentially be real drugs,” said Frances Anastassacos, co-founder and president of Nabla. “They will have to go through lead optimization, but unlocking target space is a very exciting space to be. That’s the real promise of
de novo
design.”
Other startups appear to be close behind, if not in lockstep, with Nabla, and that makes its first-ever claims hard to judge. For instance, a California startup called Abalone Bio told
Endpoints News
it has also generated
de novo
antibodies targeting GPCRs, the same hard-to-hit protein family cited in Nabla’s preprint.
Abalone has not published those results, but presented some of its findings at
an antibody engineering conference earlier this month
and also framed them as a first for the field. “To our knowledge, these are the first designed GPCR antibody sequences that are functionally active,” its research poster stated.
Protein design has arguably become the hottest space when applying AI to biology in 2024.
David Baker, a pioneer in the field, won a share of the Nobel Prize in Chemistry last month. He also co-founded Xaira Therapeutics, which
launched earlier this year with $1 billion
to bring AI into drug R&D, specifically designing antibodies. A year ago, Flagship Pioneering’s Generate:Biomedicines
published its own
de novo
model
, called Chroma, and has raised nearly $750 million to build an extensive pipeline of antibody drugs.
The leaders of those startups have also acknowledged the limitations of their recent work. Xaira CEO Marc Tessier-Lavigne recently told
Endpoints News
the AI models can generate initial hit candidates, but those still need to be improved by “old-school methods.” Generate R&D head Alexandra Snyder said earlier this year that designing from scratch isn’t yet ready for prime time, echoing what Baker has acknowledged in his own lab’s research.
Despite having a fraction of the resources of Xaira or Generate, Nabla hopes to lead the field. It
recently closed a $26 million Series A round
and employs 16 people in its 4,500-square-foot lab in Cambridge, MA. CEO Surge Biswas, who co-founded Nabla in 2020 based on his PhD work in Harvard University geneticist George Church’s lab, said technology, rather than capital, is the key limit to progress.
“Developing a system like we developed, the scarce commodity is not that you have $100 million to solve this problem,” Biswas said in an interview. “It’s not cheap, but you don’t need to raise giant amounts of capital to unlock some of these technologies.”
In an interview, Church called Nabla’s preprint one of “relatively few actual success stories” in applying
de novo
work for so-called undruggable targets.
“My hat is off to them for choosing targets well and not settling for a bunch of academic papers that settle five-decade-old challenges,” Church said.
Nabla’s preprint details an AI system called JAM, or Joint Atomic Modeling, in which scientists provide the sequence and structure of a desired drug target. The system then generates antibodies that may fit that target. The most promising creations showed strong binding affinities and looked developable in lab testing, according to the preprint.
There are still limitations to Nabla’s findings. They mainly made nanobodies rather than full-sized monoclonal antibodies that are commonplace in the drug industry. The paper included one example of a
de novo
monoclonal antibody, but that larger size also contributed to a lower binding success rate of 0.0017% for that project.
Nabla doesn’t plan to advance these antibodies itself, as the company is focused on partnering with pharma, with ongoing deals with AstraZeneca, Bristol Myers Squibb and Takeda.