What if designing a protein felt less like writing code and more like sculpting something with your hands? A new bioRxiv preprint introduces ProteinSketch, a VR environment where researchers can physically sketch protein backbones in 3D, shape the volume they want a protein to occupy, and let a diffusion model turn those gestures into molecular structures. Put on a headset, use both hands to draw a topology or define a curved 3D envelope, and RFdiffusion continuously refines that spatial idea into candidate proteins. This sounds almost like a futuristic interface demo, but the deeper problem is real: generative protein models are becoming extremely capable, while communicating high-level spatial intent to them remains surprisingly difficult. A sequence prompt is easy to type. “Build me a protein that bends around this molecular surface, enters this gap, and presents a binding interface here” is much harder. ProteinSketch makes geometry itself part of the conversation. The researchers used volumetric constraints to generate unusual anisometric protein shapes and confirmed shape fidelity by cryo-EM. They also used the same framework for functional binder design and to extend existing minibinders toward molecular surfaces that conventional design approaches struggled to reach. What I like about this direction is that it changes the interface to generative biology. We have spent enormous effort teaching models the geometry of proteins. The next challenge may be giving humans a more natural way to express the geometry they actually want. Maybe the future protein-design language will not be text at all. It may be shape, motion and space. Preprint : biorxiv.org/content/10.64898…