Nabla Bio
Last reviewed
Jun 7, 2026
Sources
17 citations
Review status
Source-backed
Revision
v1 · 1,876 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Jun 7, 2026
Sources
17 citations
Review status
Source-backed
Revision
v1 · 1,876 words
Add missing citations, update stale details, or suggest a clearer explanation.
Nabla Bio is a Boston-based artificial intelligence biotechnology company that designs antibodies and other protein therapeutics computationally, an approach it calls generative protein design. The company spun out of the laboratory of geneticist George Church at Harvard University and the Wyss Institute for Biologically Inspired Engineering, and was founded in 2020 by Surge Biswas and Frances Anastassacos. Its central technology is a family of generative models named JAM, for Joint Atomic Modeling, which generates candidate antibody sequences directly from a target protein's sequence or structure rather than discovering them through animal immunization or large physical library screens. Nabla focuses on historically intractable drug targets, in particular multipass membrane proteins such as G protein-coupled receptors (GPCRs), ion channels, and transporters. It has signed collaborations with AstraZeneca, Bristol Myers Squibb, and Takeda, and in 2025 reported what it described as the first computationally designed antibodies against a GPCR.
Nabla Bio was incorporated in May 2020 and initially operated out of the Pagliuca Harvard Life Lab, a wet-lab incubator for Harvard-affiliated life science startups. The company was co-founded by Surge Biswas, who serves as chief executive officer, Frances Anastassacos, and George Church, who is listed as an academic co-founder. Biswas and Anastassacos, who are married, form the company's core leadership.
Biswas completed his PhD in Church's lab, where he and colleagues helped pioneer protein language modeling, a machine learning technique that treats amino acid sequences analogously to the way natural language processing models treat text. The approach trains models on large databases of natural protein sequences so they learn statistical patterns of what functional, well-behaved proteins look like. This line of work was closely related to the broader emergence of learned protein representations that also underpinned advances in protein structure prediction such as AlphaFold. Anastassacos earned a PhD in biological engineering and previously worked at the venture firm Flagship Pioneering, the investor behind Moderna. Church is a founding core faculty member of the Wyss Institute and leads its synthetic biology platform, which positions Nabla within the institute's tradition of translating academic research into companies.
From the outset the company paired computational design with high-throughput laboratory measurement, an integrated model intended to generate experimental data that feeds back into its models. Early descriptions emphasized the ability to express and biophysically characterize very large numbers of antibodies in parallel, on the order of tens of thousands to roughly a hundred thousand in a single experiment, and to link results back to the underlying DNA sequences.
Nabla's design engine is a generative system called JAM (Joint Atomic Modeling). JAM produces entirely new antibody sequences, including both single-domain antibodies (VHHs, the variable domains derived from heavy-chain-only antibodies) and full-length monoclonal antibodies, optimized jointly for binding affinity, manufacturability, and drug-like developability properties. The model draws on both public protein data and Nabla's internal biologics datasets.
A distinguishing feature of the more recent JAM-2 system is its use of test-time scaling, a technique borrowed from large language modeling in which a model expends additional computation at inference time to improve results. Nabla refers to its iterative inference procedure as introspection: the model generates many candidate designs, computationally scores and filters them, and refines toward the most promising candidates before any are made in the laboratory. In Nabla's reported benchmarks, allowing this iterative inference rather than a single design pass yielded substantially higher success rates and better affinities while preserving developability.
JAM-2 also allows a user to specify a particular epitope, the precise patch on the target antigen that the antibody should bind. According to Nabla, this epitope-level control makes it possible to design antibodies intended to act as agonists or antagonists, and the company reported success rates between roughly 30 percent and 70 percent at generating binders for user-defined patches on about half of the targets it tested. Nabla has stated a design-to-experiment feedback loop on the order of three to four weeks. Independent commentary has urged caution about interpreting such preprint metrics, and at least one published analysis flagged validation gaps, so the strongest performance claims should be read as company-reported and not yet peer reviewed.
Nabla's stated focus is de novo design of antibodies against targets that have resisted conventional antibody discovery, with an initial emphasis on multipass membrane proteins. These include GPCRs, which form the largest protein family in the human genome and account for roughly one-third of approved drug targets, as well as ion channels and transporters. Such proteins are difficult to raise antibodies against because they are embedded in the cell membrane, hard to purify in their native conformation, and present limited exposed surface.
In a preprint posted to bioRxiv in May 2025, titled "De novo design of hundreds of functional GPCR-targeting antibodies enabled by scaling test-time compute," Nabla scientists reported designing antibodies against the chemokine receptors CXCR4 and CXCR7, both GPCRs implicated in cancer. The company reported generating hundreds of functional single-domain antibodies, with top designs reaching picomolar to low-nanomolar affinities, high selectivity, and favorable early-stage developability, including designs that bound the receptors in their native cellular context. For CXCR7, Nabla reported generating more than 700 designs of which several hundred showed activator function, describing them as the first reported antibody activators for that receptor. The same body of work reported a roughly 22-fold improvement in success rate for a SARS-CoV-2 spike protein target when using multiple rounds of introspection versus a single design pass. Nabla has characterized these results as among the first demonstrations of computationally designed antibodies against GPCRs and other hard-to-drug membrane proteins. These are early-stage, preclinical findings; Nabla had no antibodies in human clinical trials as of early 2026, consistent with its status as a young, platform-focused company.
The work sits within the broader field of AI drug discovery and AI for science, where computational protein design has moved from optimizing existing molecules toward generating novel binders from scratch. Reporting in the trade and scientific press placed Nabla among a cohort of companies trying to bring de novo design into practical healthcare use, while also noting general industry skepticism about whether AI drug design claims translate into approved medicines.
Nabla's commercial model centers on collaborations with large pharmaceutical companies that apply its platform to their own targets. In May 2024 the company announced strategic collaborations with AstraZeneca, Bristol Myers Squibb, and Takeda, which it valued in aggregate at more than 550 million US dollars in upfront and milestone payments plus royalties. The bulk of that figure is success-based milestones contingent on programs advancing, rather than guaranteed cash, a standard structure for early platform deals.
In October 2025 Nabla signed a second, expanded collaboration with Takeda. The companies disclosed an upfront and research-funding component described as tens of millions of dollars, with total potential value, including success-based milestones, stated as more than 1 billion US dollars. The expanded deal targets areas including multispecific antibodies, receptor decoys, and other custom biologics for hard-to-treat diseases within Takeda's early discovery pipeline. As with the 2024 agreement, the headline billion-dollar figure reflects biobucks, the maximum payable only if multiple programs hit their milestones.
Nabla Bio has raised a comparatively modest amount of venture capital relative to the size of its pharma deals, reflecting a strategy of funding the company substantially through partnerships. It closed an 11 million US dollar seed round in December 2021, co-led by Khosla Ventures and Zetta Venture Partners, with participation from Fifty Years and Cantos Ventures. In May 2024 it raised a 26 million US dollar Series A led by Radical Ventures, alongside the three pharma collaborations announced the same day. Reported cumulative venture funding stood at roughly 37 million US dollars across these rounds.
A note on naming: Nabla Bio (protein design) is a distinct company from Nabla (an ambient AI clinical documentation assistant based in Paris and New York), and figures for the two are frequently conflated in aggregated databases.
| Date | Event | Counterparty / Investor | Disclosed value |
|---|---|---|---|
| December 2021 | Seed round | Khosla Ventures, Zetta Venture Partners (co-leads); Fifty Years, Cantos Ventures | 11 million USD |
| May 2024 | Series A | Radical Ventures (lead); existing investors | 26 million USD |
| May 2024 | Strategic collaborations (announced together) | AstraZeneca, Bristol Myers Squibb, Takeda | More than 550 million USD combined, upfront plus milestones, plus royalties |
| October 2025 | Expanded collaboration | Takeda | Tens of millions upfront; more than 1 billion USD total potential, including milestones |
Nabla Bio is frequently cited as an early example of generative AI applied to one of the harder problems in drug discovery: designing functional antibodies entirely in silico against targets, especially GPCRs and other membrane proteins, that have long frustrated conventional discovery. If its computational designs hold up through independent peer review and, ultimately, the clinic, the approach could shorten timelines and expand the universe of druggable targets, a claim Nabla's leadership has framed in terms of potentially increasing the number of addressable disease-relevant targets. The company's combination of academic pedigree from Church's lab, a generative modeling stack with test-time scaling, and a string of large pharma collaborations has made it a closely watched data point in the debate over how much of the promise of AI for biology will translate into approved medicines. At the same time, its results to date remain largely preprint-stage and preclinical, and observers have cautioned that demonstrated binders are not yet demonstrated drugs.