Nabla Bio
Nabla Bio designs antibodies on a computer for pharmaceutical partners, then makes and tests them in its own lab. It concentrates on the hardest class of target, proteins embedded in the cell membrane such as G protein coupled receptors, ion channels and transporters, which are difficult to present outside a cell and so resist conventional discovery.
The platform is JAM, for Joint Atomic Modeling. Given a target and a binding site, it completes the antibody: one module generates designs and a second scores how likely each one is to survive an experiment. The wet lab feeds the model, producing thousands of designs in mammalian cells, measuring them for binding, developability and function, and returning the results to training.
The JAM-2 technical report, released openly in 2025, gives results target by target across 16 targets unlike anything in the training data: binders to all 16, at average success rates of 39 percent for single domain designs and 18 percent for full length antibodies, from 45 designs per format. It sets those results against Chai Discovery's published numbers and names the target where Nabla produced no binder.
The business runs on discovery collaborations. AstraZeneca, Bristol Myers Squibb and Takeda signed in May 2024 for more than $550 million combined in upfront and milestone payments, and Takeda signed a second, larger agreement in October 2025 that pays double digit millions upfront against success based payments that may exceed $1 billion. In 2024 an AstraZeneca scientist said publicly that the designs were then fairly low affinity; the 2025 preprints report picomolar to low nanomolar affinity against CXCR4 and CXCR7.
Nabla is a Harvard spinout from George Church's laboratory, led by chief executive Surge Biswas and president Frances Anastassacos, launched in December 2021 and based in Cambridge, Massachusetts. It has raised about $37 million, including a $26 million Series A led by Radical Ventures in May 2024, and alongside the partner work it is building a portfolio of its own designed biologics to license out.
Capability Axes
An AI Health Index grade measures what a buyer can verify from public sources on the date shown. It is not a rating of how good the product is. A vendor can build an excellent system and grade low on an axis because it publishes nothing an outsider can check. How grades read
JAM is what Nabla delivers: a partner brings a target and an epitope, the model completes the antibody, and the design is the deliverable. The company describes its aim as making drug development a design discipline rather than a search.
Its experimental systems exist to teach JAM how to succeed in human biology, producing thousands of designs in mammalian systems and measuring binding, developability and function so the measurements sharpen the model. The published results are zero shot designs against targets with no close homology to anything in the training data, with no immunization or display campaign behind them.
One half of the system checks the other. The Generator builds complete protein complexes from partially specified ones and the Filter scores how likely a complex is to survive an experiment, with output fed back through rounds the company calls introspection, so the model screens its own designs before anything is made. Designs that get through are then made and measured in Nabla's own wet laboratory, which produces thousands of designs in mammalian systems and measures binding, developability and function in cellular and in vivo systems the company says reflect human biology. Bind rates are reported for each round, failures included.
The Filter score at which a design is dropped, and which decisions need a person, are not published.
JAM, for Joint Atomic Modeling, is a multimodal generative modeling system that the company says is trained on massive protein sequence and structure data and strengthened with its own human relevant measurements. It works as molecular autocomplete: given partial molecular context such as a target or an epitope, it completes the rest. Two modules carry names, a Generator that builds complete protein complexes from partially specified ones and a Filter that scores how likely a complex is to survive an experiment, with output fed back across rounds the company calls introspection. The successor, JAM-2, comes with an openly posted 46 page technical report tied to dated results.
No architecture, model class, parameter count, training procedure or training data size is given for either version.
JAM is built in house and trained on public protein sequence and structure data together with measurements Nabla generates in its own wet laboratory, and the company treats the data, the models and the integrated dry and wet laboratory systems as one engine it owns. Every design comes from that in house model rather than from an outside provider's.
The compute or hosting provider, open source components and third party licenses are not named.
A January 2025 preprint reported the first fully computationally designed antibodies against multipass membrane proteins, Claudin-4 and CXCR7, at double digit nanomolar affinity in single domain and paired formats. A May 2025 preprint applied test time scaling, letting the model compute longer at generation rather than retraining it, and reported hundreds of designs against CXCR4 and CXCR7 reaching picomolar to low nanomolar affinity, including what the company says are the first antibody activators reported for CXCR7. The JAM-2 technical report of 2025 adds binders to all 16 unseen targets from 45 designs per format, a second study of 10 more targets with 20 chosen epitopes each, designs that bind CXCR4 and CXCR7 on live cells, and developability profiling of hundreds of designs in which 57 percent passed all four core industry criteria at once. The 16 target run took a four person team under four weeks, from first design to purified affinity measurements. The founders' earlier protein engineering and structure prediction work appears in Nature Methods and Nature Biotechnology.
AstraZeneca, Bristol Myers Squibb and Takeda each began work on the platform in May 2024, and Takeda, a partner since 2022, returned in October 2025 for a second and broader agreement that reaches across its early stage development programs. The company reports low immunogenicity in primate studies for antibodies from the platform.
None of this work has been through peer review, the preprints are published under Nabla Bio's own name, and no molecule from the platform has entered a clinical study.
Every experimental result feeds back into JAM: designs are produced in mammalian systems, measured for binding, developability and function, and those measurements sharpen the next round of design. What moves through the loop is protein data and the company's own laboratory measurements rather than anything about patients, and Nabla runs antibody programs for AstraZeneca, Bristol Myers Squibb and Takeda at the same time on that one engine.
How a partner's targets, epitopes and assay results are retained, and whether they train models that serve another partner, is not set out.
The inputs to JAM are protein sequences, structures, epitope specifications and the company's own laboratory measurements, generated for preclinical discovery programs run with pharmaceutical partners. No patient data, clinical record or trial feed enters that work, and nothing Nabla offers is software a provider would run.
Its privacy notice covers the website alone, the contact details a visitor submits and the first party Google Analytics cookies the site sets, and makes no statement on protected health information or HIPAA.
Target and epitope information from AstraZeneca, Bristol Myers Squibb and Takeda sits inside Nabla's one environment, alongside what the company calls the largest developability dataset for computationally designed biologics. Its website privacy notice says precautions protect the information the site collects from loss, misuse and unauthorized access.
No SOC 2 report, ISO 27001 certificate or trust center is published.
Nabla designs biologics at the discovery stage. JAM is not sold as a product or presented as a medical device, and partners carry development and commercialization for what comes out of it. Alongside the partner programs the company is building a portfolio of its own designed biologics against high value disease targets, to license to partners with development expertise.
No investigational new drug application, trial start or development candidate has been announced.
The JAM-2 technical report, released openly as a 46 page document, reports performance target by target, epitope by epitope and format by format, with the denominators attached. Across 16 unseen targets chosen to have no close homology to any antibody antigen complex available at the model's training cutoff, JAM-2 produced binders to all of them, at average success rates of 39 percent for single domain designs and 18 percent for full length antibodies, from only 45 designs per format. Across a further 10 antigens with 20 user specified epitopes each, it produced binders to 30 to 70 percent of epitopes for half the targets. On the two G protein coupled receptors tested in native cellular context, success rates were 11.7 percent for CXCR4 and 3.8 percent for CXCR7, and developability is reported criterion by criterion across hundreds of designs.
The report benchmarks JAM-2 against a named competitor's published results on overlapping targets and records that it produced no binders to TNFL9, where the comparator reported a 100 percent success rate, before giving a methodological reason the comparison may be uneven. Measurement conditions come with the results: a goodness of fit threshold for reporting a dissociation constant, designs excluded where it was not met, avidity effects flagged, analyte concentrations below one micromolar, and the one target where those conditions did not hold.
The report is the company's own document and has not been peer reviewed.
JAM-2 results come with their denominators and their protocol: success rates per target and per format, live cell rates for CXCR4 and CXCR7, epitope coverage for the second target set and per criterion developability pass rates, together with the goodness of fit threshold a dissociation constant must meet to be reported, the designs excluded on that basis, the analyte concentrations used and the step taken to keep avidity out of the measurement.
No guarantee, indemnity or remedy for a design that fails is published, and the collaboration agreements are private.
Partners reach Nabla through a collaboration agreement rather than software. Nabla takes a target and an epitope from a pharmaceutical partner, designs and tests the antibody in its own laboratory, and hands back designs and the measurements behind them, with the data, the AI and the integrated dry and wet laboratory systems run as one engine. The whole loop sits in preclinical discovery, upstream of any patient record or clinical workflow.
No laboratory system connection, interoperability standard or application programming interface is offered to partners.
JAM runs inside Nabla rather than shipping to anyone. The model, the data and the wet laboratory form one engine the company owns and operates, and a partner buys its output rather than access to it. The team works fully in person at the Riverside Technology Center on Memorial Drive in Cambridge, Massachusetts.
Where compute sits and where partner target and design data rests are not set out.
Nabla earns from discovery collaborations rather than selling access, with a business development mailbox on its partnering page, and its second line of business is licensing out molecules it designed itself. The AstraZeneca, Bristol Myers Squibb and Takeda collaborations announced in May 2024 carry more than $550 million combined in upfront and milestone payments plus royalties. The second Takeda agreement, signed in October 2025, pays double digit millions in upfront and research cost payments against success based payments that may exceed $1 billion, and applies JAM across Takeda's early stage development programs.
No price, milestone schedule or royalty rate is published for any agreement.
Nabla designs antibodies against multipass membrane proteins such as G protein coupled receptors, ion channels and transporters, which are hard to present outside the cell membrane, as well as soluble targets, and offers generatively designed multispecific architectures as a capability. Its published work names CXCR4, CXCR7, Claudin-4, TrkA, TNFL9, IL-7R alpha and VEGFR2, plus benchmark sets of 16 and 10 targets, and the JAM-2 report gives per target numbers for both sets and for CXCR4 and CXCR7 bound on live cells. AstraZeneca, Bristol Myers Squibb and Takeda have contracted against the platform.
No disease area, indication or molecule of the company's own has been announced.
Alternatives to Nabla Bio
Vendors in the same categories as Nabla Bio whose graded profiles look most alike, with a line on what each product does.
Similar is not the same, so check that each one does the job you need. No vendor pays for placement.
Cellarity is a Flagship Pioneering company in Somerville, Massachusetts, founded in 2019, that designs drugs against the behavior of a whole cell rather than against a molecular target.
Clinical stage biotechnology company now trading as Verge Labs, applying machine learning to human tissue derived multiomic data in neurology.
Stanford spinout, now trading as Genesis Molecular AI, building GEMS, a small molecule discovery platform its own chemists and its partners' programs run on.
Enveda is a clinical stage biotechnology company in Boulder, Colorado that uses AI to discover medicines from the chemistry of living things.
PostEra is the medicinal chemistry counterpart to the biology platforms that dominate this category.
Protein design company in Emeryville, California, built on the ProGen family of protein language models, which grew out of research first showing that large language models can generate…
Ordered by how closely the graded profiles match across the fifteen axes, inside the same categories. See how each axis is graded in the methodology.
Compared With
Each comparison carries a written verdict, the buyer conditions that favor each vendor, and a graded side by side. Pairs that cross a category boundary are grouped separately, and their verdicts state where the boundary sits rather than manufacturing a head to head.
Pricing
Vendor-published figures are labeled as such. Figures labeled “Estimated” are derived from third-party sources and have not been confirmed by the vendor.
| Entry Price | Pricing Basis | BAA Tier | Implementation | Source |
|---|---|---|---|---|
|
Not published
|
Negotiated collaboration. Revenue comes from multi year biologics discovery agreements with pharmaceutical partners, paid through an upfront payment, milestone payments as programs progress, and royalties on any approved product. No software license, subscription or platform access is offered. | — | Nabla delivers designs rather than installing, licensing or configuring software, so there is no implementation event to charge for. It runs its own wet lab and carries the cost of testing designs before they reach a partner, and partners carry development and commercialization on their side. | Vendor Published |
The partnering page routes to a business development mailbox rather than to commercial terms. Deal announcements give more than $550 million in combined upfront and milestone payments plus royalties across the AstraZeneca, Bristol Myers Squibb and Takeda collaborations of May 2024, and double digit millions in upfront and research cost payments against success based payments that may exceed $1 billion on the second Takeda agreement of October 2025. The company describes itself as revenue generating and publishes no revenue figure or rate card.