Fig 4. Impacts of solvent exposure and learned backbone error on side-chain prediction accuracy.
(A) Comparison of repacked side-chain RMSD for PEARS, SCWRL, Rosetta, and DeepSCAb with increasing relative side-chain solvent accessible surface area (SC SASA). (B) Comparison of error in DeepSCAb-predicted backbones versus error in side-chain dihedral prediction. Backbone error is measured as the Cβ deviation between the predicted structure and the native when the framework residues are aligned. Side-chain dihedral error is measured as a cosine distance between the predicted and native dihedrals.