Sketch-Conditioned Local Protein Backbone Redesign

Local protein backbone redesign supports protein engineering when functional regions must be reshaped without disrupting the surrounding fold required for structural stability and molecular function. Existing protein generation methods primarily generate complete backbones from random priors or global structural conditions. They provide limited support for specifying a residue-level local topology through a sketch while keeping the surrounding protein context fixed. To address this gap, we formulate \textbf{sketch-guided} local protein backbone redesign under a fixed context. In this task, a user-provided 2D sketch serves as a noisy visual observation of local design intent and guides generation within a fixed structural context. We construct \textbf{Sketch2Backbone}, a large-scale dataset containing more than 150K paired local redesign examples derived from natural CATH domains. We further propose \textsc{VenusSketch}, a two-stage framework that first lifts the 2D sketch into a coarse 3D C$\alpha$ trace. A context-anchored conditional flow then refines the redesigned region, while a clamping projection preserves the surrounding structure exactly at every refinement step. Experiments on held-out CATH domains evaluate paired-reference recovery under the controlled noisy-sketch protocol and show competitive post-refolding structural self-consistency.