Scaling test-time computation delivers the first computationally designed GPCR antibody agonists, overturning the assumption that de novo design cannot access receptor activation and establishing inference-time reasoning as a new axis for drug discovery.
Scaling test-time computation through iterative introspection enables de novo design of functional GPCR-targeting antibodies with >60-fold improved success rates, yielding hundreds of picomolar-affinity designs that include the first computationally designed antibody GPCR agonists and can be efficiently optimized into diverse therapeutic leads using single experimental examples to guide generation.
We present significant advances in de novo antibody design against G protein-coupled receptors (GPCRs) enabled by scaling the test-time compute used by our generative protein design system, JAM. We de novo design hundreds of VHH (single domain) antibodies against CXCR4 and CXCR7, with top designs showing picomolar to low-nanomolar affinities, high selectivity, and favorable early-stage developability profiles, matching or outperforming clinical-stage molecules in these dimensions. Further, high affinity designs potently modulate receptor function, with most acting as antagonists (inhibitors) and, strikingly, a subset functioning as agonists (activators) of CXCR7 — the first antibody agonists reported for this receptor, and the first computationally designed antibody GPCR agonists of any kind. Using a single experimentally validated agonist to further prompt JAM, we generate over 300 additional diverse agonists with superior properties, including a design with agonism EC50 rivaling that of CXCR7’s natural ligand, SDF1α. These results show that increasing the “reasoning” capacity of biomolecular generative models by scaling test-time computation will enable them to solve increasingly difficult problems in drug design.