Daisuke Ino / 稲生 大輔

Programming protein shape as an explicit design layer via CAD blueprint-guided diffusion 1. The study treats global protein shape (curvature, twist, chirality, assembly envelope) as a first-class, explicitly programmable design variable by integrating CAD blueprints directly into diffusion-based backbone generation. 2. Method: a CAD-derived “shape potential” is applied during RFdiffusion sampling on a Cα point cloud, combining (i) a signed-distance-function boundary penalty to keep residues inside the target volume and (ii) a global shape-matching term (point-cloud similarity via Chamfer distance) to enforce agreement with the blueprint continuously during denoising. 3. This enables multiple design modes with one framework: (a) de novo generation from noise into a specified 3D envelope, (b) topology-preserving reshaping of an existing fold into alternative global shape states, and (c) assembly-scale programming by tuning local shape states of shared building blocks to propagate into global architecture. 4. De novo envelope design was tested on five CAD targets (bracket, cone, sphere, ellipsoid, cube). Backbones were generated under shape constraints, sequences were designed with ProteinMPNN, and structures were checked by AlphaFold2; sharper corners/turns (cube/bracket) were harder and longer chains reduced designability. 5. Experimental validation for single-chain shapes: 15 designs selected; 11 expressed solubly and 10 were monodisperse by SEC with CD/thermal melts supporting stable folds. Two crystal structures confirmed envelope-level fidelity: Cone SC1 matched closely (Cα RMSD ~1.3 Å), while Bracket SC4 preserved the bracket contour despite one local helix-to-β rearrangement. 6. Topology-preserving reshaping: the authors targeted a rare state in nature—left-handed β-sheet twist within an α/β architecture. Starting from right-handed backbones produced by fold-conditioned diffusion, CAD blueprints encoding left-handed twist redirected diffusion trajectories to the rare twist while keeping secondary-structure order/connectivity largely intact. 7. Cryo-EM validation of the reshaped design yielded ~4.0 Å (monomer) and ~4.1 Å (dimer) reconstructions consistent with the intended curved α/β architecture and the designed twist state (with some local N-terminal flexibility), supporting the claim that “shape state” can be edited separately from fold topology. 8. Assembly programming: a shared ~340-residue helical-bundle scaffold was reshaped into distinct local curvature/twist states, then docked and interface-designed to form higher-order architectures. By varying local shape states, the same scaffold produced closed rings with different bending (270°, 180°, 90°) and open helical filaments with tunable pitch/diameter, plus programmable inversion of superhelical handedness. 9. Shape-to-function link: low-Reynolds-number hydrodynamic simulations showed symmetric ring controls generate negligible net pumping, while rotating helical propellers generate axial flow whose direction flips with handedness and whose magnitude is tunable by architecture (strongest in the 90° variants). Pumping scaled with an effective helical slope metric P/(πD), aligning with expectations from helical pumping theory. 10. Prototype ATP-dependent “molecular swimmer”: a designed helical propeller was coupled to a PS3 F1-ATPase rotor via a designed bridge (rotor-binding motifs from BindCraft + motif-fusion via RFdiffusion). nsTEM supported the intended Rotor–Bridge–Propeller organization, and single-particle tracking showed an ATP-dependent increase in apparent diffusion only when all three components were present (suggesting mechanical coupling, while not yet proving directional swimming). 💻Code: github.com/yilunqi/Shape-con… 📜Paper: biorxiv.org/content/10.64898… #ProteinDesign #DiffusionModels #GenerativeAI #StructuralBiology #CryoEM #XrayCrystallography #CAD #Nanotechnology #Biophysics #SyntheticBiology