Drug discovery · Generative protein design

Nabla Bio

A Cambridge company that writes antibodies on a computer and tests them in its own lab, aimed at the cell surface proteins conventional discovery keeps missing. Three large pharmaceutical partners, a billion dollars of contingent milestones, and not one molecule in a human being yet.

The switch that turned on

In a plate of Chinese hamster ovary cells rigged to light up when one receptor fires, two antibodies did the opposite of what they were built to do.

They were meant to block CXCR7, a receptor implicated in several cancers. Instead the light went up. The cells signalled harder, not less.

Neither came from a llama, a mouse or a phage library. Both had been written by a model, sight unseen, from the receptor sequence and a chosen patch of its surface. Nabla Bio reported them in May 2025 as the first antibody activators of CXCR7 ever described, and the first computationally designed antibody agonists of any G protein coupled receptor.

Then came the part that matters commercially. The company fed one validated activator back to the model as a prompt, generated 909 fresh designs, and found 704 bound the receptor and 348 switched it on. Of 43 measured carefully, 34 beat the original. One matched the potency of SDF1α, the natural ligand. One data point in, a lead series out, in about two weeks.

The two thirds nobody can drug

Almost every approved antibody drug hits a protein floating free in blood or sitting neatly on a cell surface. The proteins that weave in and out of the membrane several times are a different matter. They include the G protein coupled receptors, the ion channels and the transporters: roughly two thirds of all cell surface proteins, and fewer than ten percent of current biologics touch them.

The reasons are physical. Barely any of the protein sticks out above the membrane. The part that does changes shape depending on whether the receptor is on or off. Close relatives look nearly identical, so a drug that blocks one often blocks its neighbour. And the protein falls apart once pulled out of the membrane to serve as bait.

Nabla’s answer has two halves. The model designs the antibody. It also designs a soluble stand-in for the target, replacing the transmembrane section with a stable scaffold while preserving the outward-facing surface, which is what makes the receptor usable as a screening reagent at all.

Two graduate students on the same floor

Surge Biswas was in George Church’s lab at Harvard, applying language models to amino acid sequences. Frances Anastassacos was in William Shih’s lab, working on DNA origami. They worked about twenty feet apart, married, and founded the company in 2020. Church is the academic co-founder.

Biswas brought the science. The Church lab papers on low-N protein engineering and single-sequence structure prediction, in Nature Methods in 2021 and Nature Biotechnology in 2022, are the only Nabla-linked work indexed in PubMed. Anastassacos, now president, brought a Harvard PhD and a stint at Flagship Pioneering. They took an unannounced first cheque led by Fifty Years, then $11 million in December 2021 co-led by Khosla Ventures and Zetta Venture Partners. The team was seven people. By early 2025 it was about fifteen.

From one binder to a hit rate

In late 2024 Nabla reported a single VHH antibody that bound CXCR7. One, out of 20,000 designs tested. Its affinity was 21.8 nanomolar, it aggregated, and it did nothing functional. A 0.005 percent success rate.

In May 2025 the company applied what language model researchers call test-time scaling: instead of one pass, let the model propose, score and refine its own designs across six computational rounds before anything touches a bench. Against CXCR4, 7 of 35 designs came back sub-nanomolar and the best measured 370 picomolar, more than ten times tighter than ulocuplumab, a Bristol Myers Squibb antibody that has been through multiple Phase 1 and 2 trials. Against CXCR7 the best measured 2 nanomolar, roughly ten times better than a patented benchmark raised the old way by immunising a llama.

In November 2025 came JAM-2. Across 16 targets the company says the model had not seen, 45 designs per format produced binders for every one, at average success rates of 39 percent for VHH-Fc and 18 percent for full monoclonals. It also bound CXCR4 and CXCR7 directly on living cells, with no soluble stand-in, at 11.7 and 3.8 percent. In 2026 the same model was pointed at peptide-MHC targets, reporting a T cell engager that told a mutant KRAS peptide from the normal one by a single side chain and killed the presenting cells at 0.07 nanomolar.

What is proven, and what is still claimed

Read plainly: the binding data is real, measured in the company’s own lab, and better than the benchmarks it was run against. What is missing is independent replication, peer review of the two most impressive reports, an epitope binning experiment behind the claim about controlling where an antibody lands, and a molecule in a patient. The strongest external evidence is not a paper. It is that a Japanese pharmaceutical company wrote Nabla Bio into its SEC filings twice, then signed again.

What to watch

• First-in-human data. In October 2025 the company told Reuters it expected it within one to two years, placing it between late 2026 and late 2027.
• Peer review, or its absence, for JAM-2 and the 2026 peptide-MHC work. Both sit as PDFs on the company’s own servers.
• Whether AstraZeneca or Bristol Myers Squibb name Nabla Bio in their own disclosures, as Takeda now does.
• An epitope binning dataset, the experiment that would turn the epitope-control claim from a model output into a measurement.
• A new equity round. None has been announced since May 2024.

In their words

Related companies

• Abalone Bio: Also in Drug Discovery and Delivery. An East Bay biotech that screens antibodies for what they do to a receptor, not how tightly they stick to it
• Macromoltek: Also in Drug Discovery and Delivery. An Austin software company that has designed antibodies inside a computer since 2011, paid for by federal grants and customer fees rather than venture rounds
• Diffuse Bio: Also in Drug Discovery and Delivery. A Y Combinator company founded by one of the researchers who first pointed diffusion models at proteins
• Adaptyv Bio: Also in Drug Discovery and Delivery. A Swiss automated laboratory that turns AI designed protein sequences into measured binding data, and that has published the full results of three open design

Sources

1. Public recordTakeda Form 20-F, alliance table naming Nabla BioSEC EDGAR · Jun 17, 2026
2. Public recordTakeda 2025 and 2026 Annual Integrated ReportsSEC EDGAR exhibits
3. PreprintDe novo design of hundreds of functional GPCR-targeting antibodiesbioRxiv 2025.05.28.656709
4. PreprintDe novo design of epitope-specific antibodiesbioRxiv 2025.01.21.633066
5. Public recordPubMed records for Biswas SPubMed · Sep 2026
6. IndependentAI conjures up potential new antibody drugs in a matter of monthsScience · May 12, 2025
7. IndependentNabla Bio and Takeda expand AI drug design partnershipReuters · Oct 14, 2025
8. IndependentAI-based protein discovery start-up strikes a second deal with TakedaC&EN · Oct 14, 2025
9. IndependentNabla raised $26M, strikes three pharma dealsEndpoints News, via Zetta · May 2024
10. IndependentNabla Bio raises $11M for an antibody design platformTechCrunch · Dec 7, 2021
11. IndependentJAM-2: Performance Metrics and Validation GapsEngin Yapici, Medium · Dec 2, 2025
12. CompanySeries A and second Takeda collaboration announcementsBusiness Wire · 2024 and 2025
13. CompanyJAM-2 report, platform and news pagesNabla Bio · Nov 2025 to Sep 2026