Expanding the Designable Landscape of Protein Interfaces through Computational Binder Design

Abstract

Protein binders can be used to recognize, inhibit, activate, localize, or deliver molecules in biological systems. Antibodies perform these tasks naturally, but designing new binders from scratch remains difficult, especially when target surfaces are polar, flexible, recessed, or conformationally heterogeneous. In this work, we set out to expand the range of protein interfaces that can be targeted by computational binder design. We developed and tested deep-learning-based design strategies for generating and optimizing protein binders against challenging target surfaces. First, RFdiffusion interface conditioning was used to design binders that form β-strand-pairing interactions with target edge strands, enabling the design of high-affinity and specific binders to polar surfaces that are poorly addressed by standard hotspot-based design. Second, AlphaFold2-based sequence optimization methods were developed to improve existing binder-target interfaces and to optimize binders across multiple related targets. These approaches produced experimentally validated binders and optimized variants across diverse systems, including receptors, toxins, fertilization proteins, immune targets, and viral antigens. Together, these studies show that computational binder design can be extended beyond simple hydrophobic and helical interfaces, but also reveal that target-state uncertainty remains a major limitation. Future progress will require better integration of generative design, differentiable optimization, target conformational modeling, and experimental screening.