Abstract

Generative AI-based protein design can create diverse structures for desired functions but offers human designers limited control over three-dimensional architectures. We developed ProteinSketch, a bimanual virtual reality platform in which immersive backbone sketches and volumetric envelopes are directly created and translated into constraints for diffusion-based protein generation. RFdiffusion-based real-time refinement enabled interactive exploration and construction of user-specified protein topologies. Volumetric conditioning of sketched envelopes enabled high-fidelity control of anisometric geometries, confirmed by cryo-electron microscopy. These user-defined volumetric constraints enabled functional-binder design and extension of pre-existing minibinders to surfaces otherwise difficult to access within complex molecular environments using conventional design approaches. This collaborative human-AI framework integrates human spatial reasoning with generative AI for spatially directed, shape-controlled design of protein structure and function.