Isaac Sappington, Martin Toul, David S. Lee, Stephanie A. Robinson, Inna Goreshnik, Clara McCurdy, Tung Ching Chan, Nic Buchholz, Buwei Huang, Dionne Vafeados, Mariana Garcia-Sanchez, Nicole Roullier, Matthias Glögl, Christopher J. Kim, Joseph L. Watson, Susana Vázquez Torres, Koen H. G. Verschueren, Kenneth Verstraete, Cynthia S. Hinck, Melisa Benard-Valle, Brian Coventry, Jeremiah Nelson Sims, Green Ahn, Xinru Wang, Andrew P. Hinck, Timothy P. Jenkins, Hannele Ruohola-Baker, Steven M. Banik, Savvas N. Savvides, David Baker
Designing proteins that bind with high affinity to hydrophilic protein target sites remains a challenging problem. Here we show that RFdiffusion can be conditioned to generate protein scaffolds that form geometrically matched extended β-sheets with target protein edge β-strands in which polar groups on the target are complemented with hydrogen bonding groups on the design. We use this approach to design binders against edge-strand target sites on KIT, PDGFRɑ, ALK-2, ALK-3, FCRL5, NRP1, and α-CTX, and obtain higher (pM to mid nM) affinities and success rates than unconditioned RFdiffusion. Despite sharing β-strand interactions, designs have high specificity, reflecting the precise customization of interacting β-strand geometry and additional designed binder-target interactions. A binder-KIT co-crystal structure is nearly identical to the design model, confirming the accuracy of the design approach. The ability to robustly generate binders to the hydrophilic interaction surfaces of exposed β-strands considerably increases the range of computational binder design. This study demonstrates the capability of deep learning protein design models in generating functionally validated β-strand pairing interfaces, expanding the structural diversity of de novo binding proteins and accessible target surfaces.