Test-Time Scaling for De Novo Antibody Design

Read our technical report

Test-Time Scaling for De Novo Antibody Design

Read our technical report

Where Drug Discovery Becomes Design

Nabla Platform Hero

Technology

AI and human-relevant biology, built as one engine

Generative molecular design
At the core is JAM, our multimodal generative modeling system trained on massive protein sequence and structure data and strengthened with Nabla-generated, human-relevant measurements. JAM can take partial molecular context — such as a disease target or epitope — and “autocomplete” the rest, enabling de novo biologics design, epitope scaffolding/presentation, developability and affinity optimization, and rapid generation of developable multispecific architectures.

Human-relevant data generation at scale
Our experimental systems exist to teach JAM how to succeed de novo in human biology. We produce thousands of protein designs in mammalian systems and measure binding, a suite of developability properties, and function in cellular and in vivo contexts that closely reflect human biology. These tightly integrated data flows sharpen the model’s ability to design molecules with human-ready properties from the start — eliminating the need for iterative lab-in-the-loop screening.

Applications

Parallel de novo antibody design

High hit rates from JAM on soluble targets mean we can now design antibodies to many targets at once — running multiple parallel design campaigns in the time traditional discovery methods deliver one.

Unlike immunization or display-based approaches, JAM specifies the binding epitope up front, so each design set maps to a distinct site on the target. This enables systematic exploration of epitope biology, revealing functional differences across sites and guiding downstream therapeutic strategies with dramatically improved speed and precision.

(will be removed)

Generative design of multispecifics

JAM uses knowledge from both public protein data and Nabla’s large internal biologics datasets to learn what “well-behaved” proteins look like. Starting with validated binding heads, the model can generatively propose and explore multispecific architectures—linking multiple binding domains while preserving their individual target affinities and ensuring overall developability.

This approach tackles one of the hardest problems in next-generation biologics: most multispecifics fail because they misfold, aggregate, or lose binding. By designing in a high-likelihood space from the outset, JAM cuts out that attrition. The result is multispecific medicines that can engage multiple disease pathways at once, sharpen disease precision, reduce resistance, and enable combinations (e.g., simultaneous immune activation and checkpoint inhibition) that are out of reach with conventional antibodies.

(will be removed)

Drugging hard targets

Many high-value drug targets — such as GPCRs, ion channels, and transporters — are notoriously difficult to access. Their binding sites are small, barely available above the cell membrane, or nearly identical to closely related off-targets that differ by only a few amino acids. Conventional high-throughput screening or immunization struggles to find selective binders in this landscape.

Our platform overcomes these barriers with epitope-precise, de novo antibody design and high hit rates. JAM can design antibodies that bind exactly where needed, even on hard-to-reach epitopes, and Nabla’s human-relevant assays directly measure selectivity and function.


The result: therapeutic candidates for a vast expanse of challenging targets, expanding the space of diseases biologics can treat.

Will be removed

Publications