An AI that designs antibodies atom by atom - and a wet lab that proves they work.
The inverted triangle in the logo is the “nabla” operator from vector calculus. A math symbol for the direction of steepest change, borrowed by a company that wants to change the steepest part of drug discovery.
In a lab off Memorial Drive in Cambridge, a computer proposes an antibody that has never existed. Days later, a robot makes 100,000 of its cousins and asks each one a single question: do you bind? Most of the answers are no. Enough of them are yes. That gap - between a molecule imagined and a molecule that works - is the whole business of Nabla Bio.
Nabla is small. Around twenty-two people. It has raised roughly $37 million, which in biotech is a rounding error. And yet three of the largest pharmaceutical companies on the planet - AstraZeneca, Bristol Myers Squibb, and Takeda - have signed deals to use its software. The October 2025 Takeda expansion alone carries more than a billion dollars in potential payments. For a company that fits in a single open-plan room, that is a strange kind of leverage.
The leverage has a name: JAM, short for Joint Atomic Modeling. It is the thing the pharma giants are actually buying. And to understand why they want it, you have to understand the problem it was built to embarrass.
“Nabla develops integrated AI and wet-lab technologies that enable atomically precise drug design and high-throughput measurement of drug function.”
Here is the dirty secret of finding a new biologic drug: it is mostly trial and error wearing a lab coat. You immunize animals, you screen enormous libraries, you fish for something that sticks to your target, and then you spend years fixing everything that is wrong with it. It is slow. It is costly. And against certain targets, it simply does not work.
The targets that break the old playbook are the interesting ones. Multipass membrane proteins - GPCRs, ion channels, transporters - sit woven through the cell wall like thread through fabric. They are central to disease and notoriously hard to drug with antibodies. The conventional toolkit tends to shrug at them. Which is a polite way of saying it fails.
Nabla's founders looked at this and asked an unfashionable question. What if you did not discover antibodies at all? What if you designed them, the way an engineer designs a bridge - on purpose, to spec, against any target you choose?
“Where drug discovery becomes design.”
Nabla was founded in 2020 by Surge Biswas and Frances Anastassacos. Biswas had recently finished his PhD in George Church's lab at Harvard, where he built one of the earliest protein language models - software that learns the grammar of proteins the way a large language model learns the grammar of English. Anastassacos came from biological engineering and a stint at Flagship Pioneering, the firm behind Moderna. They are also, as it happens, married. Make of that what you will; the company's risk profile was already high.
The bet was that the language-model idea, the one Biswas had sketched in academia, could be scaled into a true design engine. Not autocomplete for sentences. Autocomplete for molecules. Give the model a disease target, and let it fill in the antibody.
Khosla Ventures and Zetta Venture Partners found that plausible enough to put in $11 million of seed money in late 2021. The harder sell came later, when the bet had to produce something that could survive contact with a wet lab.
“Biswas created an early protein language model as a graduate student. That technology became central to Nabla's antibody design platform.”
JAM is a multimodal generative model trained on huge volumes of protein sequence and structure data, then sharpened with Nabla's own lab measurements - the human-relevant kind that public datasets do not contain. Hand it partial molecular context, a target or an epitope, and it autocompletes the rest. De novo antibodies. Multispecifics. Receptor decoys. Developability and affinity tuned in.
The cleverest part is almost philosophical. In 2025, Nabla showed that JAM gets better not by retraining but by being allowed to think for longer at the moment of design. They call it introspection: generate a batch of candidates, keep the best handful, use those to seed the next batch, repeat. It is best-of-N rejection sampling, and it means you can buy quality with compute. The same scaling-laws logic that made chatbots smarter, pointed at antibodies.
Then comes the part the software cannot fake. Nabla's wet lab can characterize the binding of around 100,000 antibodies in roughly three weeks. Design proposes; the lab disposes. The two halves feed each other, which is the entire point.
A generative model that takes a target and “autocompletes” the antibody - sequence and structure together.
Give the model more time to think at design time and it returns more, better binders. Compute buys quality.
~100,000 antibodies functionally screened in ~3 weeks. Every design gets graded by reality.
“JAM has shown double-digit success rates in de novo design - including picomolar binders to GPCRs in a true zero-shot setting.”
Skepticism is the correct response to an AI that promises designed drugs. So here is the evidence Nabla put on the table. It reported the first fully computationally designed antibody binders, of any kind, to the GPCR CXCR7 (ACKR3) and to Claudin-4. Not optimized versions of existing antibodies. Binders conjured from a target description and a model.
More striking: among the high-affinity CXCR7 designs, some did not just stick - they switched the receptor on. Antibody agonists. The first ever reported for CXCR7, and the first computationally designed antibody GPCR agonists of any kind. Designing a molecule that binds is hard. Designing one that binds and then does something useful is the part that makes pharma reach for its checkbook.
Pharma reached. Takeda came back for a second collaboration in October 2025, the kind of repeat business that says more than any press release. AstraZeneca and Bristol Myers Squibb are in. The deals are structured as upfront payments, milestones, and royalties - so most of the value is contingent, which is honest. The market is betting on the platform, not the press.
“The first antibody agonists reported for CXCR7 - and the first computationally designed antibody GPCR agonists of any kind.”
Strip away the model architecture and the deal terms and Nabla's mission is almost plain: make drug development a true design discipline, reduce the trial and error, get better medicines to patients faster. It sounds modest. It is not. Most of medicine still runs on discovery - on finding what nature or luck provides. Design means deciding what you want and building it.
The targets Nabla chose first are the ones that matter and resist: membrane proteins threaded through cell walls, implicated in disease, historically out of antibody reach. If you can design binders for those on purpose, the list of treatable diseases grows for reasons that have nothing to do with luck.
The skeptic's question is fair: plenty of AI biotechs have promised to redraw the map and delivered slideware. What separates Nabla, for now, is the loop. Software that designs, a lab that grades, and a model that improves when you spend more compute on it. That combination has a flywheel quality that pure-software and pure-wet-lab competitors lack.
It is also early. The big deal values are potential, not banked. The hard yards - molecules that survive the clinic and reach patients - are still ahead, and the clinic humbles everyone. Nabla has earned attention, not a victory lap.
So return to that lab off Memorial Drive. A computer proposes an antibody that has never existed. The difference now is who is watching. A year ago the proposal was a research curiosity. Today it is the thing AstraZeneca, Bristol Myers Squibb, and Takeda are paying to see filled in. The molecule still has to work. But the question has quietly changed - from whether you can design a drug at all, to which one you would like to design next.