Extended Data Fig. 5: RIF docking grid-based search algorithm, β-barrel scaffold construction and post-design ligand-docking simulations.

a, Illustration of grid-based hierarchical search strategy in RIF docking. After generating an ensemble of interactions for the target ligand (Fig. 3), each one of the selected scaffold is docked into the fixed RIF using the grid-based hierarchical searching algorithm. This search procedure starts from coarse sampling grids to fine sampling grids in 3D space. An example 2D grid scheme is shown in the upper row, from the lowest resolution (coarse sampling, left) to the highest resolution (fine sampling, right). At each searching stage, the backbone is assigned to different grids on the basis of its relative position and the resulting docking configurations are scored. The top-scored backbone positions (highlighted by cyan circles in the 2D scheme) are shown as 3D structures in the lower row for each searching resolution and are continued for the next grid search and scoring. The 3D structure example shown here was streptavidin structure (PDB ID: 1STP) with grid searching resolutions of 8.0 Å, 4.0 Å, 2.0 Å, and 1.0 Å. b–d, β-barrel scaffold construction for small-molecule binding. Three geometric constraints (b) were used to describe each backbone hydrogen bond and drive the backbone assembly during Rosetta low-resolution centroid modelling. Backbones generated with all three constraints had a very narrow Φ/Ψ distribution as a result of strong constraints (c, Ramachandran plot in upper left, set 1, density coloured in blue); by omitting N–H–O angle constraint, backbone torsion diversity slightly improved (c, upper right, set 2). These two raw backbone sets yielded few non-redundant RIF docking solutions (d, blue bars). After two rounds of sequence design calculation using Rosetta full-atom force field (Supplementary Methods), regularized backbones (peptide bonds with proper dihedral geometry) and broadened Φ/Ψ distribution (c, Ramachandran plot in the lower row, density coloured in orange) yielded more unique RIF docking solutions (d, orange bars). e, Computed metrics for 42 designs ordered and tested. Results from ab initio folding simulation were scaled to 0.0 to 1.0, in which 1.0 represents a funnel-shaped folding landscape⁵⁵. f, Alternative ligand-binding conformations revealed by post-design ligand-docking simulations. The lowest-energy docking conformation using the design model (by simply taking out the ligand from the pocket) was similar to the designed DFHBI-binding mode (top left, grey; designed binding mode was circled in grey in the energy landscape in the lower row). Docking simulations using an apo-protein model refined by molecular dynamics simulations revealed an alternative equal-energy docking conformation (top right, green) that is indicated by a green circle in the docking energy landscapes (bottom). Both binding modes rely on three hydrogen-bonding residues from RIF docking (top).