Generative de novo antibody design system jointly modeling sequence and all-atom structure, refined by scaling test-time compute.
No providers recorded yet. Browse all providers
G protein-coupled receptors are the target of more than a third of approved drugs, yet almost none of those drugs are antibodies. These multipass membrane proteins resist the purification and display steps conventional antibody discovery depends on, and immunization campaigns against them fail often enough that most GPCR programs default to small molecules or peptides. Computational design offers a way around the reagent problem, but generative protein design systems have historically produced GPCR binders at hit rates too low to be useful.
JAM (Joint Atomic Modeling) is the generative design system built by Nabla Bio to attack that problem. It treats a protein complex — the sequence and all-atom structure of one or several interacting chains — as its native input and output space, and generatively fills in whatever is left unspecified. Given a target's sequence, structure, and a chosen epitope, it writes both the sequence and the structure of an antibody against it. The same machinery designs soluble scaffolds that preserve a GPCR's extracellular surface while replacing its transmembrane region, giving the campaign its own screening reagent.
The system's headline result transfers the test-time scaling idea from language models to protein design. Rather than sampling designs in one forward pass, JAM runs an "introspection" loop: the Generator proposes many antibody-target complexes, a Filter module scores their experimental viability, and the top-scoring complexes are fed back for another round of resampling. Six rounds before any wet-lab work raised de novo binding success rates by more than an order of magnitude. JAM is proprietary; its architecture is deliberately undisclosed, and no weights, code, or public API are available.
The Generator and Filter were pretrained on public protein sequence and structure databases, including both experimentally determined and predicted structures, then post-trained on internal data from over 70 design-build-test rounds, each contributing 10⁴–10⁵ generated complexes paired with quantitative expression and binding measurements. Parameter counts and architectural details are not disclosed. Against CXCR4 and CXCR7, 20,000 three-round and roughly 90,000–107,000 six-round designs per target were screened on soluble proxies, then confirmed on cells. Six-round designs converted from proxy binder to native-receptor binder at 65% (CXCR7) and 55% (CXCR4), versus 3.1% for a single round — a 64-fold increase in the estimated lower-bound hit rate. Median on-cell K_D was 2 nM for CXCR4 and 34 nM for CXCR7, with a best design at 370 pM, more than tenfold tighter than the clinical-stage anti-CXCR4 antibody ulocuplumab. No design bound the alternate receptor despite the two sharing the ligand SDF1α. All designs differed by at least 18% from the nearest of over three billion database antibody sequences, and by more than 7 Å target-aligned RMSD from the nearest SAbDab complex.
The target application is therapeutic antibody discovery against protein classes where immunization and phage display underperform — GPCRs, ion channels, and other multipass membrane proteins. Of the CXCR7 designs profiled, 74% antagonized SDF1α-driven β-arrestin signaling, and two were agonists, the first antibody agonists reported for that receptor and the first computationally designed antibody agonists of any GPCR. Re-prompting with one validated agonist yielded over 300 further agonists, including one whose EC50 approaches the natural ligand's. Because the platform is closed, access is through partnership with Nabla Bio rather than direct use; the authors state that characterized design sequences will be released publicly.
JAM's contribution is evidence that inference-time compute is a genuine lever in biomolecular design, not only in language modeling: the same model, given more rounds to refine, moves from producing one usable GPCR binder in 20,000 to producing them routinely, with developability profiles competitive with clinical molecules. Two limits bound the result. The system is closed — no architecture, weights, or benchmarks are available, so the reported success rates cannot be reproduced outside Nabla Bio. And validation, thorough as it is at the assay level, remains in vitro and on-cell; no design has been tested in animals or the clinic. The preprint has not been peer reviewed.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.