Bulged torso restraints improve native-like HCDR3 sampling and recovery.
Using Rosetta LoopModel, 1,000 models of the benchmark antibody 4G5Z (circles) were generated with or without bulged restraints and these models were then scored with or without bulged restraints (panel A, modeled and scored without restraints; panel B, modeled without but scored with restraints; panel C, modeled with but scored without restraints; panel D, modeled and scored with restraints). The native crystal structure 4G5Z was also minimized using Rosetta FastRelax, generating 20 structures (black x’s). The total HCDR3 score (in Rosetta Energy Units, or REU) is shown versus the Cα root mean square deviation of the HCDR3 loop, normalized to that of a protein loop containing 16 residues (RMSD16, in Å) to the native crystal structure. Models with scores ranked in the top 10% and RMSD16 ≤ 2 Å have been colored blue, while models with scores ranked below the top 10% and RMSD16 > 2 Å have been colored red. Improved native-like HCDR3 sampling is observed as a greater density of low RMSD16 models (blue circles) in comparison to Panel A, while improved model recovery is defined as a greater correlation between RMSD16 and score (colored vs. gray circles) in comparison to Panel A, as seen in panels C and D.