Target-conditioned diffusion generates potent TNFR superfamily antagonists and agonists

Despite progress in the design of protein binding proteins, the shape matching of binder to target has not yet reached that of highly evolved native protein-protein complexes, and previous design efforts have failed for hard targets such as the TNF receptor (TNFR1) that have relatively flat and polar surfaces. We reasoned that free diffusion starting from random noise could enable generation of extensive shape-matching binders to challenging targets, and tested this approach on TNFR1 and related super family members. The diffused TNFR1 binders have nanomolar affinities that increase to single-digit picomolar upon refinement by partial diffusion, and their specificities can be completely switched by partial diffusion in the context of other family members. The designs function as antagonists as monomers, and as superagonists when presented trivalently for OX40 and at higher valency for 4-1BB. The ability to design high -affinity and specific antagonists and agonists for a difficult but pharmacologically important class of proteins entirely in silico, without any large-scale screening or experimental optimization, presages a new era in which binders are made by computation rather than much more laborious and less controllable random screening approaches.